Method, system and apparatus for reasoning using probability information

Abstract

Input data in the reasoning process are generally highly correlated and mutually dependent. These correlations provide intrinsic redundancy in the input data, which can be used to refine the inference results, thereby improving inference performance. Probability information is used to refine results from the inference process. The probability information indicates, for each of a plurality of potential results available from the reasoning process, a probability of obtaining a respective potential result and another potential result from the plurality of potential results. By refining the reasoning result in this way, the reasoning performance can be improved.

Description

Technical Field

[0001] This application relates to reasoning, and in particular to reasoning using probabilistic information. Background Art

[0002] Future wireless communication systems, such as sixth-generation or “6G” wireless communication, are expected to trend toward increasingly diverse application scenarios, including the use of artificial intelligence (AI) (e.g., machine-learning (ML)) and perception to serve a large number of devices.

[0003] A common application of machine learning is to perform inference to extract insights from data. In the context of wireless communication networks, machine learning processes such as deep neural networks (DNNs) can be trained to perform inference using data from devices in the network. For example, a machine learning process may be deployed in a data center that is far away from the devices providing the data, which means that large amounts of data may need to be transferred from the devices to the machine learning process over the network. Since wireless connections may not provide sufficient bandwidth and stability to transfer data to the machine learning process, this data transfer is only feasible when the device is connected to the network through a wired or fiber connection that can provide broadband and stable connections. Summary of the invention

[0004] In next-generation networks, inference can be performed jointly by the network and the devices in the network, rather than just in a centralized data center. Inference jobs can be distributed to multiple devices so that each device performs one or more tasks as part of a distributed machine learning process. This can alleviate the computational load on each device compared to a situation where one device performs the entire inference job, while also reducing the amount of data that each device may need to communicate as part of the machine learning process (e.g., reducing traffic load). Because the computational and traffic loads on each device are reduced, inference can be performed using low-complexity devices such as IoT devices. This means that inference can be performed using low-cost hardware that can even be battery-powered.

[0005] However, challenges may still arise because low-cost devices may not be as reliable in both communication and computation. In real-time systems, which require not only inference accuracy but also low latency, the low capabilities of a single device can lead to poor inference performance. Many devices may be required to work together to improve inference performance (e.g., improve accuracy and latency) and achieve the required quality of service.

[0006] According to aspects of the present invention, the inference results from the distributed inference process can be refined by exploiting the correlations present in the data performing the inference. The input data of the inference process are usually highly correlated and interdependent. These correlations provide inherent redundancy in the input data, which can be used to refine the inference results, thereby improving the inference performance. For example, this can improve the accuracy of the inference results. In addition, the use of these correlations can adapt the inference to specific applications, environments, and changes in the environment.

[0007] In a first aspect, a method is provided. The method involves: receiving an input of a component reasoning process from a network device, the component reasoning process constituting a part of a distributed reasoning process representing a machine learning process. The method also involves: executing the component reasoning process on the input to obtain a first reasoning result. The method also involves: sending a second reasoning result based on the first reasoning result and probability information to the network device. The probability information indicates, for each potential result of a plurality of potential results obtainable from the component reasoning process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results.

[0008] The machine learning process may include a classification process. The plurality of potential results may include a plurality of categories, so that the probability information indicates, for each of the plurality of categories obtainable from the component reasoning process, a probability of obtaining the corresponding category and another category from the plurality of categories. The first reasoning result includes a corresponding first confidence c for each category i in the plurality of categories. 1 (i) The second inference result includes a corresponding second confidence c for each category in the plurality of categories. 2 (i).

[0009] The probability information may include: for each of the multiple categories, assuming that other potential results j have been obtained from the component reasoning process, obtaining the corresponding conditional probability P(i, j|j) of the corresponding category i and the other category j from the multiple potential results. The method also involves: determining the corresponding second confidence c of the corresponding category i in the multiple categories according to the following equation 2 (i)

[0010]

[0011] The machine learning process may include a regression process. The plurality of potential outcomes may include a plurality of values ​​of a parameter. The first inference result may include a first probability distribution P of the parameter associated with the first data i in the input 1 (x i). The second inference result may include a second probability distribution P of the parameter associated with the first data i 2 (x I ).

[0012] The probability information may include obtaining a first value x of the parameter i and obtain the second value of the parameter x j The joint probability P(x i ,x j ). The method further comprises determining the second probability distribution according to the following equation:

[0013] P 2 (x i )=∫P 1 (x j )P(x i ,x j )dx j .

[0014] The probability information may include, assuming that the second value x of the parameter has been obtained j , get the first value x of the parameter i The conditional probability P(x i |x j ). The method further comprises determining the second probability distribution according to the following equation:

[0015] P 2 (x i )=∫P 1 (x j )P(x i |x j )dx j .

[0016] The method may also include receiving an indication of the probability information from the network device.

[0017] The method may further involve: updating the probability information based on the first reasoning result or the second reasoning result.

[0018] The method may also involve one or more of: indicating the updated probability information to the network device; indicating the updated probability information to an apparatus for performing inference as part of the distributed inference process.

[0019] Also provided is an apparatus (e.g., an entity) for performing the above method. In yet another aspect, a memory (e.g., a non-transitory processor-readable medium) is provided. The memory includes instructions (e.g., processor-executable instructions) that, when executed by a processor of the apparatus, cause the apparatus to perform the above method.

[0020] In a second aspect, a method performed by a network device is provided. The method includes: sending a corresponding first input of a plurality of first inputs to each of a plurality of processing devices. Each corresponding first input is used in a component reasoning process as part of a distributed reasoning process representing a machine learning process. The method also includes: receiving, from the plurality of processing devices, a first reasoning result obtained based on probability information and the plurality of first inputs, wherein the probability information indicates, for each potential result of a plurality of potential results obtainable from the component reasoning process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results.

[0021] The plurality of first inputs may include a plurality of second inputs and at least one redundant input. The method may further include: encoding the plurality of second inputs to generate the at least one redundant input; and decoding the first inference result to obtain a second inference result.

[0022] The method may further involve receiving, from at least one of the plurality of processing devices, a respective update to the probability information, the respective update being based on the first inference results obtained from the component inference processes.

[0023] The method may further involve indicating the probability information to the plurality of processing devices by indicating the same probability information to each of the plurality of processing devices.

[0024] The method may further involve indicating the probability information to the plurality of processing devices by indicating to at least one of the plurality of processing devices first probability information specific to the respective processing device.

[0025] The plurality of potential results obtainable from the component reasoning process may be a first plurality of potential results. The method may further involve: obtaining second probability information indicating, for each potential result in a second plurality of potential results obtainable from the distributed reasoning process, a probability of obtaining a corresponding potential result and another potential result from the second plurality of potential results. The method may further involve: for each of the at least one of the plurality of processing devices, selecting the first probability information from the second probability information based on the first plurality of potential results.

[0026] A network device for performing the above method is also provided. In yet another aspect, a memory (e.g., a non-transitory processor-readable medium) is provided. The memory includes instructions (e.g., processor-executable instructions) that, when executed by a processor of a network device, cause the network device to perform the above method.

[0027] In a third aspect, a method performed by a network device is provided. The method involves: sending a corresponding first input of a plurality of first inputs to each of a plurality of processing devices. The corresponding first input is used in a component reasoning process as part of a distributed reasoning process representing a machine learning process. The method also includes: receiving a first reasoning result based on the plurality of first inputs from the plurality of processing devices. The method also includes: determining a second reasoning result based on the first reasoning result and probability information, wherein the probability information indicates, for each potential result of a plurality of potential results that can be obtained from the component reasoning process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results.

[0028] The plurality of first inputs may include a plurality of second inputs and at least one redundant input. The method may further include: encoding the plurality of second inputs to generate the at least one redundant input; and decoding the first inference result to obtain a second inference result.

[0029] In a fourth aspect, a method is provided. The method includes: performing an inference process on input data to obtain a first inference result. The method also includes: determining a second inference result based on the first inference result and probability information. The probability information indicates, for each potential result among a plurality of potential results obtainable from the inference process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results.

[0030] In yet another aspect, a system is provided. The system includes a first device for obtaining a first inference result as part of a distributed inference process representing a machine learning process. The system also includes a second device in communication with the first device. The second device is used to obtain a second inference result as part of the distributed inference process, the second inference result being based on the first inference result and probability information. The probability information indicates, for each potential result of a plurality of potential results obtainable from the inference process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more fully understand the present embodiment and its advantages, the following description is made by way of example with reference to the accompanying drawings, in which:

[0032] Figure 1 is a schematic diagram of a communication system in which embodiments of the present invention may occur;

[0033] Figure 2 is another schematic diagram of a communication system in which embodiments of the present invention may occur;

[0034] Figure 3 is a block diagram illustrating units or modules in a device in which embodiments of the present invention may be implemented;

[0035] Figure 4 is a block diagram illustrating units or modules in a device in which embodiments of the present invention may be implemented;

[0036] Figure 5 is a block diagram of an exemplary system for implementing a coding reasoning process provided by an embodiment of the present invention;

[0037] Figure 6 and Figure 7 showing the number of times each object in the plurality of objects appears in the same image as another object in the two sets of images;

[0038] Figure 8 shows the number of times objects of different categories appear in the same image of the image collection;

[0039] Fig. 9 and Fig.10 Correlation plots of image classes in the two datasets are shown;

[0040] Fig.11 is a diagram of a method provided by an embodiment of the present invention;

[0041] Figures 12 to 15 A flow chart of a method provided by an embodiment of the present invention is shown;

[0042] Fig.16 and Fig.17 The detection and classification of objects in an image provided by an embodiment of the present invention are shown;

[0043] Fig.18 The object detection rate of the distributed reasoning process performed according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0044] The operation and structure of the current exemplary embodiment are discussed in detail below. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in any of a variety of specific contexts. The specific embodiments discussed are merely illustrative of the specific structure of the present invention and the specific way to use the present invention, and do not limit the scope of the present invention.

[0045] refer to Figure 1, as an illustrative example but not limiting, a simplified schematic diagram of a communication system is provided. The communication system 100 includes a radio access network 120. The radio access network 120 may be a next generation (e.g., sixth generation (6G) or higher) radio access network, or a traditional (e.g., 5G, 4G, 3G, or 2G) radio access network. One or more communication electronic devices (electronic devices, ED) 110a to 110j (collectively referred to as 110) may be interconnected with each other, or connected to one or more network nodes (170a, 170b, collectively referred to as 170) in the radio access network 120. The core network 130 may be part of the communication system and may be dependent on or independent of the radio access technology used in the communication system 100. In addition, the communication system 100 includes a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160.

[0046] Figure 2 An exemplary communication system 100 is shown. Typically, the communication system 100 enables multiple wireless or wired elements to transmit data and other content. The purpose of the communication system 100 may be to provide content such as voice, data, video and / or text through broadcast, multicast and unicast, etc. The communication system 100 may operate by sharing resources (e.g., carrier spectrum bandwidth) between its constituent elements. The communication system 100 may include a terrestrial communication system and / or a non-terrestrial communication system. The communication system 100 may provide a wide range of communication services and applications (e.g., earth monitoring, remote sensing, passive sensing and positioning, navigation and tracking, autonomous distribution and mobility, etc.). The communication system 100 may provide a high degree of availability and robustness through the joint operation of a terrestrial communication system and a non-terrestrial communication system. For example, integrating a non-terrestrial communication system (or its components) into a terrestrial communication system may realize a heterogeneous network including multiple layers. Compared with traditional communication networks, heterogeneous networks may obtain better overall performance through efficient multi-link joint operation, more flexible function sharing, and faster physical layer link switching between terrestrial networks and non-terrestrial networks.

[0047] The terrestrial communication system and the non-terrestrial communication system can be considered as subsystems of the communication system. In the example shown, the communication system 100 includes electronic devices (ED) 110a to 110d (generally referred to as ED 110), radio access networks (RAN) 120a and 120b, non-terrestrial communication networks 120c, core networks 130, public switched telephone networks (PSTN) 140, Internet 150 and other networks 160. RAN 120a and 120b include corresponding base stations (BS) 170a and 170b, which can be generally referred to as terrestrial transmit and receive points (T-TRP) 170a and 170b. Non-terrestrial communication network 120c includes access node 120c, which can be generally referred to as non-terrestrial transmit and receive point (NT-TRP) 172.

[0048] Alternatively or additionally, any ED 110 may be used to connect, access or communicate with any other T-TRP 170a and 170b and NT-TRP 172, the Internet 150, the core network 130, the PSTN 140, other networks 160, or any combination thereof. In some examples, ED 110a may perform uplink and / or downlink transmissions with T-TRP 170a via interface 190a. In some examples, ED 110a, 110b, and 110d may also communicate directly with each other via one or more sidelink air interfaces 190b. In some examples, ED 110d may perform uplink and / or downlink transmissions with NT-TRP 172 via interface 190c.

[0049] The air interfaces 190a and 190b may use similar communication technologies, such as any suitable wireless access technology. For example, the communication system 100 may implement one or more channel access methods in the air interfaces 190a and 190b, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal FDMA (OFDMA), or single-carrier FDMA (SC-FDMA). The air interfaces 190a and 190b may utilize other high-dimensional signal spaces, which may involve a combination of orthogonal dimensions and / or non-orthogonal dimensions.

[0050] The air interface 190c may enable communication between the ED 110d and one or more NT-TRPs 172 via a wireless link or simply a link. For some examples, the link is a dedicated connection for unicast transmission, a connection for broadcast transmission, or a connection between a group of EDs and one or more NT-TRPs for multicast transmission.

[0051] The RANs 120a and 120b communicate with the core network 130 to provide various services, such as voice, data, and other services to the EDs 110a, 110b, and 110c. The RANs 120a and 120b and / or the core network 130 may communicate directly or indirectly with one or more other RANs (not shown), which may or may not be directly served by the core network 130 and may or may not employ the same radio access technology as the RANs 120a, 120b, or both. The core network 130 may also serve as a gateway access between (i) the RANs 120a and 120b and / or the EDs 110a, 110b, and 110c and (ii) other networks, such as the PSTN 140, the Internet 150, and other networks 160. In addition, some or all of the EDs 110a, 110b, and 110c may include functionality to communicate with different wireless networks over different wireless links using different wireless technologies and / or protocols. Instead of (or in addition to) wireless communications, the EDs 110a, 110b, and 110c may also communicate with a service provider or switch (not shown) and with the Internet 150 over a wired communication channel. The PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). The Internet 150 may include a network of computers and / or subnets (intranets) and incorporate protocols such as the Internet Protocol (IP), the Transmission Control Protocol (TCP), and the User Datagram Protocol (UDP). The EDs 110a, 110b, and 110c may be multimode devices capable of operating in accordance with a variety of wireless access technologies and include a plurality of transceivers required to support these technologies.

[0052] Figure 3Another example of ED 110 and base stations 170a, 170b and / or 170c is shown. ED 110 is used to connect people, objects, machines, etc. ED 110 can be widely used in various scenarios, such as cellular communication, device-to-device (D2D), vehicle to everything (V2X), peer-to-peer (P2P), machine-to-machine (M2M), machine-type communication (MTC), Internet of Things (IOT), virtual reality (VR), augmented reality (AR), industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, smart city, drone, robot, remote sensing, passive perception, positioning, navigation and tracking, autonomous distribution and mobility, etc.

[0053] Each ED 110 represents any end-user device suitable for wireless operation, and may include (or may be referred to as): user equipment / device (UE), wireless transmit / receive unit (WTRU), mobile station, fixed or mobile user unit, cellular phone, station (STA), machine type communication (MTC) equipment, personal digital assistant (PDA), smart phone, laptop, computer, tablet, wireless sensor, consumer electronic device, smart book, vehicle, car, truck, bus, train or IoT device, industrial equipment or devices in the aforementioned equipment (e.g., communication module, modem or chip), etc. The next generation ED 110 may be referred to using other terms. Base stations 170a and 170b are T-TRPs and will be referred to as T-TRP 170 hereinafter. Figure 3 170 and / or NT-TRP 172. Each ED 110 connected to T-TRP 170 and / or NT-TRP 172 may be dynamically or semi-statically opened (i.e., established, activated, or enabled), closed (i.e., released, deactivated, or disabled), and / or configured in response to one or more of connection availability and connection necessity.

[0054] ED 110 includes a transmitter 201 and a receiver 203 coupled to one or more antennas 204. Only one antenna 204 is shown in this figure. One, some or all of the antennas may also be panels. For example, transmitter 201 and receiver 203 may be integrated into a transceiver. The transceiver is used to modulate data or other content so that it can be sent by at least one antenna 204 or a network interface controller (NIC). The transceiver is also used to demodulate data or other content received by at least one antenna 204. Each transceiver includes any suitable structure for generating signals for wireless or wired transmission, and / or for processing signals received wirelessly or wired. Each antenna 204 includes any suitable structure for sending and / or receiving wireless or wired signals.

[0055] ED 110 includes at least one memory 208. Memory 208 stores instructions and data used, generated, or collected by ED 110. For example, memory 208 can store software instructions or modules that are used to implement some or all of the functions and / or embodiments described herein and are executed by one or more processing units 210. Each memory 208 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices. Any suitable type of memory can be used, such as random access memory (RAM), read only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, processor cache, etc.

[0056] ED 110 may also include one or more input / output devices (not shown) or interfaces (e.g., connected to Figure 1 The input / output devices support interaction with users or other devices in the network. Each input / output device includes any suitable structure for providing information to or receiving information from a user, such as a speaker, microphone, keypad, keyboard, display or touch screen, including network interface communications.

[0057] ED 110 also includes a processor 210 for performing operations, including operations related to preparing transmissions for uplink transmission to NT-TRP 172 and / or T-TRP 170, operations related to processing downlink transmissions received from NT-TRP 172 and / or T-TRP 170, and operations related to processing sidelink transmissions to and from another ED 110. Processing operations related to preparing transmissions for uplink transmission may include operations such as encoding, modulation, transmit beamforming, and generating symbols for transmission. Processing operations related to processing downlink transmissions may include operations such as receive beamforming, demodulation, and decoding received symbols. According to an embodiment, the downlink transmission may be received by receiver 203, possibly using receive beamforming, and processor 210 may extract signaling from the downlink transmission (e.g., by detecting and / or decoding the signaling). An example of signaling may be a reference signal sent by NT-TRP 172 and / or T-TRP 170. In some embodiments, the processor 276 implements transmit beamforming and / or receive beamforming based on an indication of a beam direction (e.g., beam angle information (BAI)) received from the T-TRP 170. In some embodiments, the processor 210 may perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as operations related to detecting a synchronization sequence, decoding, and obtaining system information. In some embodiments, the processor 210 may perform channel estimation, for example, using a reference signal received from the NT-TRP 172 and / or the T-TRP 170.

[0058] Although not shown, the processor 210 may form part of the transmitter 201 and / or the receiver 203. Although not shown, the memory 208 may constitute part of the processor 210.

[0059] Each of the processor 210 and the processing components of the transmitter 201 and the receiver 203 may be implemented by the same or different one or more processors that are used to execute instructions stored in the memory (e.g., the memory 208). Alternatively, the processor 210 and some or all of the processing components of the transmitter 201 and the receiver 203 may be implemented using dedicated circuits, such as a programmed field-programmable gate array (FPGA), a graphical processing unit (GPU), or an application-specific integrated circuit (ASIC).

[0060] The T-TRP 170 may be known by other names in some implementations, such as a base station, a base transceiver station (BTS), a wireless base station, a network node, a network device, a network-side device, a transmission / reception node, a NodeB, an evolved NodeB (eNodeB or eNB), a home eNodeB, a next generation NodeB (gNB), a transmission point (TP), a site controller, an access point (AP) or a wireless router, a relay station, a remote radio head, a ground node, a ground network device or a ground base station, a base band unit (BBU), a remote radio unit (RRU), an active antenna unit (AAU), a remote radio head (RRH), a central unit (CU), a distributed unit (DU), a positioning node, etc. The T-TRP 170 may be a macro BS, a micro BS, a relay node, a host node, etc., or a combination thereof. T-TRP 170 may refer to the aforementioned device, or to a means in the aforementioned device (eg, a communication module, a modem, or a chip).

[0061] In some embodiments, various parts of T-TRP 170 can be distributed. For example, some modules of T-TRP 170 can be located away from the device that houses the antenna of T-TRP 170, and can be coupled to the device that houses the antenna through a communication link (not shown) sometimes called a fronthaul (e.g., a common public radio interface (CPRI)). Therefore, in some embodiments, the term T-TRP 170 can also refer to the modules on the network side that perform processing operations such as determining the location of ED110, resource allocation (scheduling), message generation, and encoding / decoding, which are not necessarily part of the device that houses the antenna of T-TRP 170. These modules can also be coupled to other T-TRPs. In some embodiments, T-TRP 170 can actually be multiple T-TRPs that operate together to serve ED 110 through coordinated multi-point transmission, etc.

[0062] The T-TRP 170 includes at least one transmitter 252 and at least one receiver 254 coupled to one or more antennas 256. Only one antenna 256 is shown in this figure. One, some or all of the antennas may also be panels. The transmitter 252 and the receiver 254 may be integrated into a transceiver. The T-TRP 170 also includes a processor 260 for performing operations, including operations related to preparing transmissions for downlink transmission to the ED 110, processing uplink transmissions received from the ED 110, preparing transmissions for backhaul transmission to the NT-TRP 172, and processing transmissions received from the NT-TRP 172 via backhaul. Processing operations related to preparing transmissions for downlink or backhaul transmission may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to processing transmissions received in the uplink or via backhaul may include operations such as receive beamforming, demodulation, and decoding received symbols. The processor 260 may also perform operations related to network access (e.g., initial access) and / or downlink synchronization, such as generating the content of a synchronization signal block (SSB), generating system information, etc. In some embodiments, the processor 260 also generates a beam direction indication, such as a BAI, that can be scheduled for transmission by the scheduler 253. The processor 260 performs other network-side processing operations described herein, such as determining the location of the ED 110, determining the location of the deployed NT-TRP 172, etc. In some embodiments, the processor 260 may generate signaling, such as one or more parameters for configuring the ED 110 and / or one or more parameters of the NT-TRP 172. Any signaling generated by the processor 260 is sent by the transmitter 252. It should be noted that the "signaling" used herein may also be referred to as control signaling. Dynamic signaling may be sent in a control channel, such as a physical downlink control channel (PDCCH), and static or semi-static higher layer signaling may be included in a message sent in a data channel (such as a physical downlink shared channel (PDSCH)).

[0063] The scheduler 253 may be coupled to the processor 260. The scheduler 253 may be included within the T-TRP 170 or operate separately from the T-TRP 170, and the scheduler 253 may schedule uplink transmissions, downlink transmissions, and / or backhaul transmissions, including issuing scheduling grants and / or configuring scheduling-free ("configuration grants") resources. The T-TRP 170 also includes a memory 258 for storing information and data. The memory 258 stores instructions and data used, generated, or collected by the T-TRP 170. For example, the memory 258 may store software instructions or modules for implementing some or all of the functionality and / or embodiments described herein and executed by the processor 260.

[0064] Although not shown, the processor 260 may constitute a part of the transmitter 252 and / or the receiver 254. In addition, although not shown, the processor 260 may implement the scheduler 253. Although not shown, the memory 258 may constitute a part of the processor 260.

[0065] The processor 260, the scheduler 253, and the processing components of the transmitter 252 and the receiver 254 may each be implemented by the same or different one or more processors that execute instructions stored in a memory (e.g., the memory 258). Alternatively, some or all of the processing components of the processor 260, the scheduler 253, and the transmitter 252 and the receiver 254 may be implemented using dedicated circuits, such as an FPGA, a GPU, or an ASIC.

[0066] Although the NT-TRP 172 is shown as a drone only as an example, the NT-TRP 172 can be implemented in any suitable non-ground form. In addition, the NT-TRP 172 may be known by other names in some implementations, such as a non-ground node, a non-ground network device, or a non-ground base station. The NT-TRP 172 includes a transmitter 272 and a receiver 274 coupled to one or more antennas 280. Only one antenna 280 is shown in this figure. One, part, or all of the antennas may also be panels. The transmitter 272 and the receiver 274 may be integrated into a transceiver. The NT-TRP 172 also includes a processor 276 for performing operations, including operations related to: preparing transmissions for downlink transmission to the ED 110, processing uplink transmissions received from the ED 110, preparing transmissions for backhaul transmission to the T-TRP 170, and processing transmissions received from the T-TRP 170 via backhaul. Processing operations associated with preparing transmissions for downlink or backhaul transmission may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations associated with processing transmissions received in the uplink or via backhaul may include operations such as receive beamforming, demodulation, and decoding received symbols. In some embodiments, processor 276 implements transmit beamforming and / or receive beamforming based on beam direction information (e.g., BAI) received from T-TRP 170. In some embodiments, processor 276 may generate signaling, such as one or more parameters for configuring ED 110. In some embodiments, NT-TRP 172 implements physical layer processing, but does not implement higher-level functions such as functions at the medium access control (MAC) or radio link control (RLC) layer. Since this is just an example, more generally, NT-TRP 172 may implement higher-level functions in addition to physical layer processing.

[0067] The NT-TRP 172 also includes a memory 278 for storing information and data. Although not shown, the processor 276 may constitute a part of the transmitter 272 and / or the receiver 274. Although not shown, the memory 278 may constitute a part of the processor 276.

[0068] Processor 276 and the processing components of transmitter 272 and receiver 274 can each be implemented by the same or different one or more processors that execute instructions stored in a memory (e.g., memory 278). Alternatively, processor 276 and some or all of the processing components of transmitter 272 and receiver 274 can be implemented using dedicated circuits, such as a programmed FPGA, GPU, or ASIC. In some embodiments, NT-TRP 172 can actually be multiple NT-TRPs that operate together to serve ED 110 through coordinated multi-point transmission, etc.

[0069] The T-TRP 170, NT-TRP 172, and / or ED 110 may include other components, but these components are omitted for clarity.

[0070] One or more steps of the exemplary methods provided herein may be performed based on Figure 4 The corresponding unit or module is executed. Figure 4 Units or modules in a device, such as in ED 110, in T-TRP 170, or in NT-TRP 172, are shown. For example, a signal may be sent by a sending unit or a sending module. For example, a signal may be sent by a sending unit or a sending module. A signal may be received by a receiving unit or a receiving module. A signal may be processed by a processing unit or a processing module. Other steps may be performed by an artificial intelligence (AI) or machine learning (ML) module. The corresponding units or modules may be implemented using hardware, one or more components or devices that execute software, or a combination thereof. For example, one or more of the units or modules may be an integrated circuit, such as a programmed FPGA, GPU, or ASIC. It should be understood that if these modules are implemented, for example, using software executed by a processor, these modules may be retrieved by the processor in whole or in part as needed, retrieved individually or collectively for processing, retrieved in one or more instances, and these modules themselves may include instructions for further deployment and instantiation.

[0071] Additional details regarding ED 110, T-TRP 170, and NT-TRP 172 are known to those skilled in the art. Therefore, these details are omitted here.

[0072] Performing Inference

[0073] In cases where a centralized machine learning process (e.g., a centralized DNN) performs inference using data from a set of distributed devices, the reliability of the machine learning process may depend on the quality, reliability, and latency of the transmission between the machine learning process and the devices.

[0074] This situation can be mitigated by performing inference locally relative to the data source. Thus, for example, a machine learning process can be deployed closer to the device. However, machine learning processes can be computationally intensive. For example, a DNN may have as many as 10 to 100 billion neurons. Therefore, performing inference using a machine learning process on a single client device can be challenging.

[0075] In contrast, by distributing the machine learning process across multiple devices, low-cost and low-power devices can be used to implement the machine learning process. Distributed reasoning may be particularly advantageous in the context of wireless communication networks because input data for the reasoning process is typically collected by devices in the access network, such as electronic communication devices and TRPs. By distributing the machine learning processing across multiple devices, the machine learning process can be implemented in or near the access network, thereby reducing the risk of loss or delay of input data for the machine learning process.

[0076] However, when the machine learning process is distributed across multiple devices, there is a risk that a device will not return its results due to errors in computation or transmission, etc. This risk can be mitigated by introducing redundancy so that lost results can be recovered.

[0077] Encoded reasoning is one way to introduce this redundancy. In encoded reasoning, the inputs to the distributed learning process are encoded to produce redundant inputs. The inputs and redundant inputs are processed by component reasoning processes (which can be the same or different) to produce inference results and redundant results. The redundant results can be used to recover lost inference results and / or refine the inference results.

[0078] Figure 5 An exemplary system 500 for implementing coding reasoning is shown. The system includes a first reasoning unit 502 and a second reasoning unit 504. The first reasoning unit 502 is used to perform a first input X 1 Perform the component reasoning process as part of the distributed reasoning process to obtain the first result Y 1 =f(X 1 The second reasoning unit 504 is used to analyze the second input X 2 Perform a component reasoning process (eg, the same component reasoning process) to obtain a second result Y 2 =f(X 2 ), as part of the distributed inference process.

[0079] The system 500 further includes a redundant reasoning unit 506 for receiving a redundant input X3=h(X1, X2) obtained by encoding the first input and the second input. The redundant reasoning unit 506 is used to perform a component reasoning process on the redundant input so as to output a third result Y3=f(h(X1, X2)), which can be used to restore one of the first result or the second result.

[0080] The reasoning units 502 to 506 may be implemented in any processing device. In some examples, the reasoning units 502 to 506 may be implemented at a corresponding electronic device (e.g., a terminal device, a user device, or an IoT device). The electronic device may be any suitable electronic device, such as a fixed camera or a mobile phone, a vehicle-mounted sensor, etc.

[0081] The system 500 may further include an encoding unit and a decoding unit (not shown). The encoding unit processes the input X 1 , X 2 Encode (eg, process) to generate redundant input h(X 1 ,X 2 ). The decoding unit uses the first inference result and the second inference result Y 1 , Y 2 At least one of the redundant inference results Y 3 Decode to recover lost results and / or refine the inference result Y 1 , Y 2 In some examples, the encoding unit and the decoding unit may be a single unit (e.g., a combined encoder-decoder). The encoding unit and / or the decoding unit may be implemented in a network device, such as a TRP, a base station, or an access point, or other device (e.g., an electronic device). The encoding unit and / or the decoding unit may be implemented at one of the inference units 502 to 506.

[0082] In some examples, the redundant input is based on a linear combination of the inputs. The redundant reasoning unit 506 can be trained to provide redundant results that are linear combinations of the reasoning results. This approach can implement distributed encoding reasoning, but it increases the complexity at the redundant reasoning unit 506. In other implementations of encoding reasoning, when redundant inputs are generated based on inputs, AI / ML invariant transformations can be enforced. For example, the redundant input X 3A cascade of inputs may be included. Generating redundant inputs in this manner avoids the need for additional training of the redundant reasoning unit 506, which means that the same component reasoning process can be used at the first reasoning unit and the second reasoning unit 502 to 504 and the redundant reasoning unit 506 to process the input and the redundant input. This means that reasoning tasks can be deployed without time-consuming training, which means that the deployment of reasoning tasks does not need to be done on a case-by-case basis. As reasoning applications become more specialized and diverse in the 6G era, this is expected to save service providers a lot of time and resources.

[0083] The coding reasoning takes inspiration from error correction coding, also referred to herein as channel coding. This can be illustrated by considering channel coding, which encodes an input binary sequence by adding some redundant bits. The redundant bits can be calculated from the input binary sequence in a predefined method and placed together with the input binary sequence, which can generate some correlations between the redundant bits and the input binary sequence within the codeword (the output of the channel encoder). When the codeword is decoded by the channel decoder, the channel decoder uses these correlations to recover the most likely binary sequence. This produces a coding gain, which can be attributed to the correlation between the original input sequence and the redundant bits.

[0084] In contrast to error-correcting coding, which typically processes binary inputs, encoded reasoning can use a wider range of data as input, such as images, audio, video, or point clouds. The output of encoded reasoning can be reasoning results, such as categories (also referred to as labels in this article) and / or quantities.

[0085] The inputs to the encoding inference process may not follow the same classical statistical assumptions as error correction coding. In particular, the inputs to the encoding inference process may not be independent and memoryless. Instead, in many applications, the inputs to the encoding inference process may be highly correlated and dependent.

[0086] Taking image classification as an example, some objects may be more likely to appear in the same image than others. For example, a zebra is more likely to appear in an image of a giraffe than in an image of a whale. This is an example of spatial correlation within a dataset.

[0087] In other examples, the inference results may be time-dependent. Events in the real world are often causally related, in that one event causes another. Therefore, real-world data often includes temporal dependencies. An example of this is audio and video clips, where events and / or objects that appear in adjacent time frames may be dependent on each other.

[0088] Aspects of the present invention use the correlations expected to be present in the input data to improve reasoning performance. These correlations can be quantified in probability information, which indicates the probability of obtaining one reasoning result and another reasoning result. The probability can be a joint probability or a conditional probability. In an embodiment of the present invention, the reasoning results obtained from the reasoning process (e.g., a machine learning process or a distributed reasoning process representing the machine learning process) can be refined using probability information, which indicates the probability of obtaining both a specific reasoning result and another reasoning result.

[0089] Reasoning performance can be improved by using correlations expected to exist in the input data to refine the inference results. In certain embodiments, probabilistic information can be used to refine the inference results from the encoded reasoning process, resulting in a process called correlated encoded reasoning. This supports the use of two types of redundancy, namely redundancy inherent in the input data and redundancy generated by encoded reasoning, to jointly improve the performance of the reasoning algorithm.

[0090] Quantifying correlations in probabilistic information

[0091] You can refer to Figures 6 to 10 Describe in more detail the nature of probability information and methods of obtaining it, Figures 6 to 10 Illustrated are examples of correlations that may occur in the input data used for the inference process. Figures 6 to 10 Relates to the detection and recognition of objects in images, also known as image classification. Image classification is used herein as an example of a reasoning process to which aspects of the present invention may be applied. In general, aspects of the present invention may be applied to any suitable reasoning process.

[0092] Figure 6 is a table obtained using images from the COCO training dataset (COCO-train2017 dataset; Lin, Tsung-Yi et al., “Microsoft Coco: Common Objects in Context”, European Conference on Computer Vision, Springer, Cham, 2014). Figure 7 A similar table is for the COCO validation dataset (COCO-val2017 dataset; Lin, Tsung-Yi et al., "Microsoft Coco: Common Objects in Context", European Conference on Computer Vision, Springer, Cham, 2014). Objects in images in the COCO dataset are labeled as one of 80 categories. These 80 categories are grouped into 12 superclasses. These 12 superclasses are grouped along Figure 6 and Figure 7 The first row and column of the table shown are listed.

[0093] Figure 6 and Figure 7The tables in show, for each image in the corresponding COCO dataset, the number of times an object in a particular superclass appears in the same image as an object in another superclass. The cells in the table are colored to reflect the strength of the correlation: the darker the cell associated with a particular superclass pair, the more images there are associated with both superclasses in the pair. Each of these tables forms a correlation graph, allowing visualization of statistical similarities between different datasets. For any dataset with associated categories, a similar table can be generated by, for each datum (e.g., each image) in the dataset, incrementing a counter associated with a set (e.g., a pair) of categories each time the set (e.g., a pair) of categories is present in the same datum.

[0094] Figure 6 and Figure 7 The table in shows that some superclasses of objects are more likely to appear in the same image than other objects. For example, the training dataset ( Figure 6 ) has 4264 images with both outdoor objects and vehicles, while the training dataset ( Figure 6 ) have both appliances and vehicles. This suggests that vehicles are more likely to appear in the same image with outdoor objects than in images that include appliances.

[0095] For example, the validation dataset ( Figure 7 ) have both food objects and kitchen objects, while only 291 images have both food objects and moving objects. This suggests that food objects and kitchen objects are more likely to appear in the same image than food objects and moving objects.

[0096] Both tables also show that there are a large number of images that include both people and objects from another superclass. This is illustrated by the darker shading on the row labeled people, indicating that the superclass "people" is highly correlated with all other superclasses. This reflects that people (humans) are central to the world observed by humans.

[0097] Similar correlations were found in the 80 categories in the COCO dataset. Figures 8 to 10 In the description. Figure 8 The co-occurrence of several pairs of categories of the COCO validation dataset is shown. Figure 8 As shown in the figure, cars and traffic lights appear together in more than 2,000 images, while giraffes and stop signs appear together in a small number of images. This means that the probability of a car and a traffic light appearing in the same image is much higher than the probability of a giraffe and a stop sign appearing in the same image.

[0098] Fig. 9 and Fig.10 It is shown that for the COCO training dataset ( Fig. 9) and COCO validation dataset ( Fig.10 ) for all 80 categories in the dataset, and a plot of the number of times an object in the corresponding category appears in the same image as an object in another corresponding category. Each row and column is associated with a corresponding category, so that each cell is associated with a pair of categories. The darker the cell for a particular pair of categories, the greater the number of images that include objects from both categories. Although the training and validation datasets are independent, the same correlation exists in both images. Therefore, both datasets show that objects of the same category are likely to appear in the same image.

[0099] According to an embodiment of the present invention, these correlations can be quantified in probability information. The probability information indicates, for each of the multiple categories into which objects in an image in the COCO dataset can be classified, the probability that an object in the corresponding category and an object in another category in the multiple categories appear in the same image.

[0100] In some examples, the probability information may include a set of conditional probabilities {P(i|j)}, i, j = 1, ... N, where each corresponding conditional probability P(i|j) is the probability that an object in category i appears in a particular image given that an object in category j appears in the image. The set of conditional probabilities may be determined by counting the number of times that objects in each category pair appear in an image of a particular COCO dataset to obtain a corresponding co-occurrence count A(i, j) for each pair. For example, a co-occurrence count may also be referred to as a coincidence count.

[0101] The conditional probability of an object in category i existing, assuming that an object in category j exists, is obtained by normalizing the co-occurrence counts of the corresponding pairs relative to the total number of occurrences of objects in category i. For example, the corresponding conditional probability of an object in category i existing, assuming that an object in category j exists, P(i|j), can be determined according to the following equation:

[0102]

[0103] In other examples, the probability information may include a joint probability set {P(i,j)}, i=1,...N and j=1,...N, and for each class pair i, j, the joint probability set includes the probability P(i,j) that objects in the two classes appear in a particular image. The joint probability set may be determined based on the co-occurrence count A(i,j) described above. In particular, the joint probability of a particular class pair may be obtained by normalizing the co-occurrence counts of the corresponding pair relative to the total number of occurrences of all classes. For example, the joint probability P(i,j) of a pair of classes may be determined according to the following equation:

[0104]

[0105] Although the above examples involve classification of objects in images (also known as image classification or recognition), and in particular use of the COCO dataset, it will be appreciated that the above methods can be generalized to other datasets and / or other reasoning problems. Therefore, different datasets can be used to generate probability information in the same or similar manner. In general, equation (1) can be used to determine the conditional probability of a pair of categories based on the co-occurrence counts A(i, j) of categories i and j of any suitable dataset. Equation (2) can be used to determine the joint probability of a pair of categories based on the co-occurrence counts A(i, j) of categories i and j of any suitable dataset.

[0106] The co-occurrence counts may be related to the class pairs that exist in the same observation (e.g., the same image). This can be described as the co-occurrence counts are defined for two or more semantics (e.g., classes, objects) within one observation. Therefore, the above joint probability can be expressed as P(a i ,b i ), where a and b are semantics and i is an observation. Assuming that semantic b exists in observation i, the conditional probability that semantic a exists in observation i can be expressed as P(a i |b i ). Assuming that semantic a exists in observation i, the conditional probability that semantic b exists in observation i can be expressed as P(b i |a i ).

[0107] Alternatively, the co-occurrence counts may relate to pairs of categories that are present in different observations (e.g., in adjacent or neighboring observations). For example, the co-occurrence counts may reflect the frequency with which a first category is present in a frame of a media clip (e.g., a video) and a second category is present in a subsequent frame of the media clip. This can be described as a co-occurrence count being defined for two or more semantics (e.g., categories, objects) across multiple observations. Thus, the above joint probability may be expressed as P(a i ,b j ), where a and b are the first semantics and the second semantics, and i and j are the first observation and the second observation. Assuming that semantic b exists in observation j, the conditional probability that semantic a exists in observation i can be expressed as P(a i |b j ). Assuming that semantic a exists in observation i, the conditional probability that semantic b exists in observation j can be expressed as P(b j |a i ).

[0108] Thus, the above-described method for determining probability information may be used to determine probability information for refining inference results obtained from any classification process. In this context, a classification process may be any process that attempts to classify, label, or categorize data. Thus, categories referred to herein may include categories, categorizations, category labels, or any other suitable way of categorizing or classifying information.

[0109] It should be understood that the classification process is an example of an inference process to which the method of the present invention can be applied.

[0110] In some examples, the probability information can be used to refine the inference results obtained from the regression process. Regression processes are often used to extract information from the values ​​of multiple variables. Thus, regression can be used to obtain inference results about quantities (e.g., temperature or brightness) across multiple observations (e.g., time and / or location). For example, a regression process can be used to identify patterns or trends in a data set, including temperature measurements at multiple time points.

[0111] Therefore, in some examples, the probability information may include the joint probability P(x i ,x j ). The joint probability indicates that the first parameter has value x at observation i i and has value x at observation j j probability.

[0112] The first parameter x may also be referred to as a variable (e.g., a continuous or discrete variable) or a parameter. Each observation in observations i, j may be a specific instance (e.g., a value) of the second parameter. For example, the value x of the first parameter x is i Can be used with a specific value y i is associated with (e.g., measured at) a second parameter y. Thus, for example, the value x i Can be related to a specific time t i In another example, the value x i Can be used with specific locations i Associated.

[0113] Get the first value of parameter x i and get the second value of parameter x j The joint probability P(x i ,x j ) can be determined according to the following equation:

[0114]

[0115] Where a and b can be real values. In practice, a and b can be quantized into discrete values.

[0116] Additionally or alternatively, the probability information may include the conditional probability P(x i |x j ), which indicates the conditional probability that the first parameter has the value x at observation j j , the first parameter has a first value x at observation i i The conditional probability can be determined according to the following equation:

[0117]

[0118] Where a and b can be real values. In practice, a and b can be quantized into discrete values.

[0119] In general, the above method can be used to obtain probability information of any reasoning process (such as a machine learning process or a component reasoning process of a distributed process), wherein the distributed reasoning process represents a machine learning process. The reasoning process may involve a classification process and / or a regression process.

[0120] In particular, the above method can be applied to obtain probability information, which indicates, for each potential result in a plurality of potential results that can be obtained from the reasoning process, the probability of obtaining the corresponding potential result and another potential result from the plurality of potential results. For example, a potential result may include a confidence associated with a particular category (e.g., for a classification process). Alternatively, a potential result may include multiple values ​​of a parameter (e.g., for a regression process). The probability information may include a joint probability or a conditional probability. The probability information may also be referred to as correlation information.

[0121] In some examples, the probability information can be stored in a table. The table can be called a relevance table. The table can be square, where each row and each column represents a corresponding semantics or observation. For example, the entries in the table can be joint probabilities or conditional probabilities. Each row in the table can be normalized so that the sum of each row can be equal to 1. Each column in the table can be normalized so that the sum of each column can be equal to 1.

[0122] In some examples, it may be advantageous for the probability information to include joint probabilities rather than conditional probabilities because joint probabilities are symmetric. i ,x j )=P(x j ,x i ). This means that, when stored in a table, the joint probabilities will be symmetric along the diagonal. Thus, in certain embodiments, the probability information may comprise an upper triangular matrix or a lower triangular matrix of the joint probabilities. This reduces the storage required to store the probability information and / or the transmission resources required to transmit the probability information by half.

[0123] Methods for refining an inference result using probabilistic information are now described. In this context, refining an inference result refers to determining a second inference result based on a first inference result. Although the second inference result may be more accurate and / or more precise than the first inference result (e.g., such that the use of probabilistic information improves the result), it should be understood that this is not a requirement, and the use of the term "refined" only means that the second inference result is different from the first inference result produced using probabilistic information.

[0124] Refining the inference results

[0125] As briefly described above, probability information indicating the probability of obtaining a particular potential outcome from the reasoning process and obtaining another potential outcome from the reasoning process can be used to refine the output of the reasoning process.

[0126] For example, an image classification process may return respective confidences indicating the likelihood of food, kitchen objects, and outdoor objects being present in the image. The confidence associated with the presence of kitchen objects in the image may be raised based on a high confidence associated with the presence of food in the image and a high likelihood that food and kitchen objects co-occur. Conversely, the confidence associated with outdoor object detection may be lowered based on a low likelihood that food and outdoor objects co-occur.

[0127] Fig.11 An example of determining a refined inference result based on probability information (labeled as “Statistical information from training images”) and a preliminary inference result (labeled as “Inference output: category confidence”) is shown.

[0128] In this example, preliminary inference results are obtained by inputting the image into the YOLO algorithm (Farhadi, Ali, and Joseph Redmon. "Yolov3: Incremental Improvements", Computer Vision and Pattern Recognition, Berlin / Heidelberg, Germany: Springer, 2018), which is an example of an image classification process. However, in general, any suitable inference process can be used, such as any suitable image classification process.

[0129] The YOLO algorithm provides preliminary reasoning results based on the input image, which include the coordinates of the bounding box in the image, the objectness score, and the initial category confidence associated with each of the N categories. The coordinates of the bounding box include the x coordinate t of the bounding box x and the y coordinate t y , and the width t of the bounding box w and height t h Target score p o Indicates the confidence of the object detected in the image. The initial category confidence c associated with the corresponding category i 1(i) indicates the likelihood that the object in the bounding box is in category i. Category confidence can also be called category score. Initial category confidence can also be called initial marginal probability (e.g., for a particular category).

[0130] Based on the conditional probability of a specific category i and the initial confidence c 1 (i), the refined confidence c that the object in the image is in category i 2 (i) can be determined according to the following equation:

[0131]

[0132]

[0133] where and are the sums of each of the other N–1 categories in the set of N categories, and P(i|j) is the conditional probability that an object in category i exists given that an object in category j exists. The refined confidence c 2 It can be normalized according to the following equation:

[0134] Therefore, the refined confidence of a specific class can be normalized based on the sum of the refined confidences of all classes.

[0135] In the example above, conditional probabilities are used to refine the confidence. The refined confidence c that the detected object is in a particular category i 2 (i) Based on the initial confidence value c 1 (i) (e.g., the confidence value provided by the inference process) and the joint probability, determined according to the following equation:

[0136]

[0137] where P(i,j) is the joint probability that an object from category i and category j exists in the image. For more information on characterizing marginalized probabilities, such as the refined confidence c 2 (i), see "Factor Graphs and Sum-Product Algorithms", Frank R. Kschischang, Brendan J. Frey and HA. Loeliger, IEEE Transactions on Information Theory 47, No. 2 (2001): pp. 498-519. In Kschischang, a factor graph is defined between random variables (variable nodes) and their relations (check nodes). According to the present invention, this can be adjusted by representing objects or events as variable nodes and their joint probabilities as check nodes.

[0138] The refined confidence c 2(i) (i = 1, ... N) can be normalized using the same method described above in the method using conditional probabilities. In fact, P(i, j) is the joint probability of co-occurrence of category / object i and category / object j given the statistics from the training data before performing the inference process. In contrast, c 1 (i) is the probability of class i given an independent observation after performing the inference process.

[0139] A refined confidence c can be determined for each of the N categories that can be identified by the YOLO algorithm 2 The refined confidence c 2 It can be used to classify objects in an image. 2 When the refined confidence of (i) satisfies (e.g., is greater than or equal to) a threshold, it can be determined that an object in the specific category exists. For example, when the following relationship is satisfied, it can be determined that the specific category c 2 The object in (i) exists:

[0140] c 2 (i)>θ,

[0141] Here, θ is the threshold value. This step can be called thresholding.

[0142] This example using the YOLO algorithm illustrates how probabilistic information can be used to refine the results of an image classification process. In general, the above approach can be applied to any classification process. Thus, in the context of a classification problem, probabilistic information can be used to refine the results of an image classification process based on a first or initial confidence c associated with class i. 1 (i) to obtain the second or refined confidence c associated with category i 2 (i).

[0143] As mentioned above, the methods described in this article can be applied to other inference processes, such as regression processes.

[0144] The regression process can be used to return the initial probability distribution P of parameter x at observation i 1 (x i ). This initial probability distribution can be called the initial marginal probability distribution. The initial probability distribution can indicate that the parameter x takes a specific value x from the potential range of values ​​at observation i. i Thus, for example, a parameter may be a temperature with potential values ​​in the range of 0 to 100 degrees Celsius. The initial probability distribution P 1 (x i ) can indicate the temperature at a certain time t i With value x i possibility.

[0145] The initial or first probability distribution may be refined based on the probability information to obtain a second probability distribution. In some examples, the probability information may include obtaining a first value x of a parameter i and get the second value of parameter x j The joint probability P(x) of i ,x j ), and a second or refined probability distribution can be determined based on the joint probability. For example, the second probability distribution can be determined according to the following equation:

[0146] P 2 (x i )=∫P 1 (x j )P(x i ,x j )dx j . #(8)

[0147] In other words, the second probability distribution can be obtained by using the first probability distribution P provided by the regression process. 1 (x j ) for x j P(x i ,x j ) is integrated to determine. The second probability distribution can be normalized so that the total probability sum is 1. The second probability distribution can be normalized according to the following equation:

[0148]

[0149] In some embodiments, the second probability distribution P returned by equation (9) 2 may have already been normalized, so no further normalization is needed.

[0150] In other examples, the probability information may include, assuming that the parameter has a second value x j (e.g., as determined above), obtaining a first value x of the parameter i The conditional probability P(x i |x j ), and a second or refined probability distribution can be determined based on the conditional probability. For example, the second probability distribution can be determined according to the following equation:

[0151] P 2 (x i )=∫P 1 (x j )P(x i |x i )dx j . #(10)

[0152] Therefore, the second probability distribution can be obtained by using the first probability distribution P provided by the regression process 1 (x j ) for x j P 1 (x j )P(x i |x j ) is integrated to determine. The second probability distribution can be normalized so that the total probability sum is 1. The second probability distribution can be normalized according to the following equation:

[0153]

[0154] In some embodiments, the second probability distribution P returned by equation (11) 2 may have already been normalized, so no further normalization is needed.

[0155] In summary, the regression process provides the marginal probability P(x i ) can be combined with the local observations to obtain P(x i ) is the refined probability distribution of .

[0156] As described above, some inference processes involve both regression and classification. Thus, in some examples, the methods described above with respect to regression and classification can be combined and applied to the inference results provided by the inference process. For example, the refinement techniques described above with respect to the regression process can be used to refine the coordinates of the bounding box provided by the image classification process, and the refinement techniques described above with respect to the classification process can be used to refine the classification of the object detected within the bounding box.

[0157] Update probability information

[0158] As described above, the probability information can be used to refine one or more inference results from the inference process. In some examples, updates to the probability information can be determined based on the inference results.

[0159] For example, assuming that category j exists, the updated conditional probability P′(i|j) that a particular category i exists can be based on the conditional probabilities P(i|j) of the existence of categories i and j provided by the inference process and the initial confidence c 1 (i) and c) 1 (j) is determined. The updated conditional probability can be determined according to the following equation:

[0160]

[0161] The denominator is used for normalization. In some embodiments, normalization may not be necessary, and the denominator may be omitted (eg, may be equal to 1).

[0162] In another example, the joint probability P(i, j) of the existence of categories i and j provided by the inference process and the initial confidence c 1 (i) and c) 1 (j) can determine the updated joint probability P′(i, j) that a specific class i exists and a class j exists. The updated joint probability can be determined according to the following equation:

[0163]

[0164] The denominator is used for normalization. In some embodiments, normalization may not be necessary, and the denominator may be omitted (eg, may be equal to 1).

[0165] In another example, suppose the parameter has a second value x at observation j j , get the first value x of the parameter at observation i i The updated conditional probability P′(x i |x j ) can be based on the initial probability distribution P of the parameter x at observation i provided by the inference process 1 (x i ) and the initial probability distribution P of parameter x at observation j 1 (x j ). The updated conditional probability can be determined according to the following equation:

[0166]

[0167] The denominator is used for normalization. In some embodiments, normalization may not be necessary, and the denominator may be omitted (eg, may be equal to 1).

[0168] In another example, get the first value x of the parameter at observation i i and obtain the second value x of the parameter at observation j j The updated joint probability P′(x i ,x j ) can be based on the initial probability distribution P of the parameter x at observation i provided by the inference process 1 (x i ) and the initial probability distribution P of parameter x at observation j 1 (x j ). The updated joint probability can be determined according to the following equation:

[0169]

[0170] The denominator is used for normalization. In some embodiments, normalization may not be necessary, and the denominator may be omitted (eg, may be equal to 1).

[0171] Typically, the probability information may be updated based on an initial or first inference result provided by the inference process. The initial inference result may include one or more confidence levels and / or one or more probability distributions, such as the examples given above. In some examples, the probability information may be updated based on a refined inference result that is determined based on the initial inference result provided by the inference process and the probability information itself.

[0172] In some examples, the inference results can be obtained by multiple devices (also referred to as processing devices). This will be described in more detail in the following distributed reasoning and related coded reasoning. In these examples, the same probability information can be used to refine the inference results from multiple devices. For example, the probability information can be updated based on the inference results from multiple devices. Alternatively, the probability information can be specific to a specific device or a specific device group. Therefore, the probability information can be updated based on the inference results from a specific device or device group. In some examples, the same probability information can be initially used for multiple devices, and the probability information specific to each device can be determined based on the same probability information by iteratively updating the probability information based on the inference results provided by the specific device. That is, the default probability information can be initially used for all devices and then refined according to the actual inference results provided by the device.

[0173] Refine inference results from a single device

[0174] The methods described herein can be applied to an inference process (e.g., a non-distributed inference process) performed by a single device. Thus, for example, a processing device can perform an inference process on inference data to obtain an inference result. The inference process may involve a classification and / or regression process. The inference result can be refined based on the probability information as described above. The refinement can be performed by the same processing device that performs the inference. Alternatively, the refinement can be performed elsewhere. For example, an electronic device connected to a cell served by a network device can perform inference to obtain an inference result, and the electronic device can send the inference result to the network device for refinement using the methods described herein.

[0175] Distributed Correlation Reasoning

[0176] In other examples, the methods described herein can be applied to distributed reasoning, wherein component reasoning processes are performed at multiple processing devices to obtain corresponding initial reasoning results at each processing device. The distributed reasoning process can represent a machine learning process, such as a neural network (e.g., a deep neural network (DNN)), a k-nearest neighbor process, a linear regression process, a logistic regression process, a support vector machine, or any other suitable machine learning process. The initial or first reasoning result can be refined based on probabilistic information using any of the above methods to obtain a refined or second reasoning result. These refinement techniques may be particularly advantageous for distributed reasoning processes because they can be implemented without incurring a significant processing burden. For example, these refinement techniques can improve reasoning performance at less computational cost than using more training data or using a more complex reasoning process (e.g., using different or more involved machine learning processes or algorithms) to train the reasoning process.

[0177] The refinement may be performed at the processing device or elsewhere. For example, the processing device may send the initial reasoning result to the network device, and the network device may determine the refined reasoning result based on the probability information and the initial reasoning result. By refining the reasoning result elsewhere (e.g., not at the processing device), the reasoning performance may be improved without further processing at the processing device. This may save processing resources at the processing device, which may be particularly advantageous for low-complexity and / or low-power processing devices.

[0178] In some examples, the inference results may be further processed after refinement. This may be particularly suitable for a classification process, in which the initial inference results may include corresponding initial confidences for one or more categories, and the refined inference results may include refined or second confidences for one or more categories. According to aspects of the present invention, the refined confidences may be compared to thresholds to confirm the detection of one or more categories. This may be referred to as thresholding and may be performed as described above. Thresholding may be performed by a device that performs a component inference process and / or a device that refines the inference results. In some examples, thresholding may be performed elsewhere. For example, a network device may obtain a refined inference result and send the refined inference result to another device (e.g., a device in a core network) to perform thresholding.

[0179] Related Coding Reasoning

[0180] In some embodiments, the methods described herein can be applied to encoded reasoning. Reasoning in which the input is encoded to provide redundancy and probability information is used to refine the reasoning result can be referred to as relevant encoded reasoning.

[0181] In the above about Figure 5 In the system 500 described above, for example, based on the first reasoning result Y provided by the first reasoning unit 502, 1 The probability information may indicate, for each of the plurality of potential results that may be obtained from the component reasoning process performed by the first reasoning unit 502, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results. The probability information may be obtained using any of the methods described above in the section “Quantifying Correlation in Probability Information”. The probability information may be obtained based on the first reasoning result Y using any of the methods described above in the section “Refining Reasoning Results”. 1 and probability information to determine the refined first inference result.

[0182] Similarly, based on the second inference result Y provided by the second inference unit 504, 2 The probability information used to refine the first reasoning result and the second reasoning result may be the same or different. For example, the component reasoning processes at the first reasoning unit 504 and the second reasoning unit 506 may be able to provide different reasoning results (e.g., may be classified according to different categories), and the probability information may be specific to the corresponding component reasoning process.

[0183] The refinement may be performed at the corresponding reasoning unit 504, 506 or elsewhere. For example, another device ( Figure 5 (not shown) can receive the first reasoning result, the second reasoning result and the redundant reasoning result Y from the first reasoning unit, the second reasoning unit and the redundant reasoning unit 502 to 506 1 , Y 2 and Y 3 , and refine the first inference result and the second inference result as described above. The other device may be any one of the encoding unit, decoding unit or encoder-decoder described above.

[0184] In some examples, the method described in this article can also be used to refine the redundant reasoning result Y based on probability information. 3 This applies when the input X 1 , X 2 Taken from a similar scenario, because this might mean that input X 1 、x 2 Any inherent correlation in is preserved in the redundant input h(X 1 ,X 2 ). In other examples, the redundant inference results may not be refined based on the probability information. 1 ,X 2 ) for input X 1 , X2 are combined, so X 1 , X 2 Any inherent correlations in may not be preserved in redundant inputs, which can lead to unexpected results when refining redundant inference results.

[0185] The refined inference results may be decoded (e.g., at the decoding unit or encoder-decoder described above) to determine missing inference results or further refine the inference results. A method for decoding inference results from a related encoded inference process will be described in more detail below in conjunction with method 1300.

[0186] An example method to implement relevant encoding inference is as follows:

[0187] Step 1: Perform independent inference

[0188] Get inference results (probabilities) for each individual category / object / event.

[0189] Step 2: Refine reasoning through relevance knowledge (only for system input)

[0190] >Update the inference result (probability) according to the marginal probability expression

[0191] Normalized inference results (probability)

[0192] Step 3: Refine reasoning using redundant input

[0193] >Message passing between the reasoning units 502, 504 and the redundant reasoning unit 506 based on “set operations”.

[0194] Normalized inference results (probability)

[0195] Step 4: Get the final inference result

[0196] Perform thresholding or make decisions based on the refined inference results

[0197] Step 1 may be performed according to the description of performing the component reasoning process provided herein. Step 2 may be performed according to the method described in the "Refine Reasoning Results" section. Part 3 may be performed according to the decoding described in method 1300 below. Step 4 may involve performing thresholding as described herein. Steps 1 to 4 may be performed one or more times (e.g., may be performed iteratively).

[0198] In some examples, the message passing between the reasoning units 502, 504 and the redundant reasoning unit 506 based on "set operation" in step 3 can be omitted. Without step 3, the exemplary method becomes only correlation reasoning and can be implemented in a single device (e.g., without the support of a network).

[0199] Generalize to two or more possible outcomes

[0200] In the above description, the probability information is described as indicating that, for each potential result in a plurality of potential results that can be obtained from the component reasoning process, the probability of obtaining the corresponding potential result and another potential result from the plurality of potential results. In general, the probability information may relate to the probability of obtaining the corresponding potential result and one or more other potential results from the plurality of potential results. Therefore, the above equations (1) to (15) may be applicable and / or generalized to embodiments in which the probability information relates to the probability of obtaining the corresponding potential result and two or more other potential results. For example, in the classification process, the corresponding initial confidence c of the returned category i=1,2,…,N 1 In the embodiment of (i), the specific category c 2 The refined confidence of (i) can be determined according to the following equation:

[0201]

[0202] Among them, P(i,j,k) is the joint probability that categories i, j and k exist, P(k,m) is the joint probability that k and m exist, and P(k,q) is the joint probability that categories k and q exist.

[0203] Exemplary Methods

[0204] Fig.12 1 shows a method 1200 provided by an embodiment of the present invention. In the above description, the method 1200 is described as being performed by a processing device. However, in general, the method 1200 can be performed by any suitable device, and in some examples, can be performed by more than one device. In a specific example, the method 1200 can be performed by an inference unit, such as any one of the first inference unit 502 and the second inference unit 504 described above. In some examples, the processing device can be an electronic device, such as the one described above with respect to Figures 1 to 4 Any of the electronic devices 110 described herein. In a specific example, the processing device may be a sensing device. Thus, for example, the method 1200 may be used to implement joint or collaborative sensing.

[0205] Method 1200 may begin at step 1202: a processing device receives input for a component reasoning process. In some examples, the processing device receives input from a network device. The network device may be a TRP, such as the one described above with respect to Figures 1 to 4 Any TRP 170 described. For example, the network device may be a base station and the processing device may be connected to a cell served by the base station.

[0206] In other examples, the input may be obtained in other ways. For example, the processing device may include a sensing device, and the input may include sensory data obtained (eg, measured, sensed, and / or calculated) by the processing device.

[0207] The input may include any data on which reasoning may be performed. Thus, for example, the input may include one or more of the following: image data, audio data, video data, measurement data, network data for a communication network (e.g., indicating traffic, usage, performance, or any other network parameter), user data, or any suitable data.

[0208] The component reasoning process constitutes part of a distributed reasoning process representing a machine learning process. The component reasoning process can be any suitable process (e.g., an algorithm) including one or more tasks to be performed as part of a distributed reasoning process. The component reasoning process and / or the distributed reasoning process can include any suitable machine learning process, such as a neural network (e.g., a deep neural network (DNN)), a k-nearest neighbor process, a linear regression process, a logistic regression process, a support vector machine, or any other suitable machine learning process. For example, the component reasoning process and / or the distributed reasoning process can include a regression process, a classification process (e.g., a classifier), or a combination of a regression process and a classification process. It will be appreciated by those skilled in the art that the choice of a machine learning process is typically specific to the reasoning task. For example, the reasoning task can include image classification, and the component process can include a neural network trained to classify images, such as a deep neural network.

[0209] In the context of the present invention, a distributed reasoning process may be any reasoning process that includes tasks that may be performed by multiple devices. In some examples, the distributed reasoning process may be performed by multiple processing devices, where each processing device performs a component reasoning process. Each processing device may perform the same component reasoning process. Alternatively, different processing devices may perform different component reasoning processes.

[0210] In some examples, the distributed reasoning process may include an encoded reasoning process. This will be discussed below in conjunction with Fig.13 Described in more detail, but in some examples will be understood to apply to method 1200.

[0211] Continuing the discussion of method 1200, in step 1204, the processing device performs a component reasoning process on the input to obtain a first reasoning result. It should be understood that different reasoning processes provide different results, and therefore, the form of the first reasoning result may depend on the component reasoning process, the distributed reasoning process, and / or the machine learning process represented by the distributed reasoning process.

[0212] In some examples, the machine learning process may include a classification process, and the first reasoning result may include one or more categories, and a corresponding confidence for each category. Confidence may also be referred to as a confidence score, a confidence index, a confidence level, a category score, a trust score, or any other suitable term. The confidence indicates the likelihood that the assignment of the category based on the input is correct. For example, an image classification process may provide one or more categories for objects detected in an image, and for each category, an associated confidence indicating the likelihood of the object being in the corresponding category. In some examples, the confidence may take a value in the range of 0 to 1, with larger values ​​indicating that the category is more likely to be correct.

[0213] In some examples, the machine learning process represented by the distributed learning process may include a regression process, and the first inference result may include a first probability distribution P of a parameter associated with the first data i. 1 (x i ). Thus, for example, step 1204 may involve performing a component inference process on the input image to obtain a corresponding first probability distribution of coordinates of a bounding box in the image. For example, a bounding box may indicate the presence of an object in the image. As described above, a bounding box may have coordinates (t x , t y , t w , t h ). Therefore, the first probability distribution may include the corresponding distribution The first probability distribution may take any suitable form.In some examples, for each potential value of a plurality of potential values ​​of the parameter, the first probability distribution may include a corresponding probability.

[0214] In some examples, the machine learning process may include classification and regression processes. Therefore, the first result may include a combination of the first results described above with respect to the classification and regression processes.

[0215] The method 1200 may also involve obtaining probability information indicating, for each potential result in a plurality of potential results obtainable from the component reasoning process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results.

[0216] The form of the multiple potential outcomes may depend on the component reasoning process. For example, for a classification process, the potential outcomes may include multiple categories. For example, for a regression process, the potential outcomes may include a range of values ​​or a set of values.

[0217] The probability information may be the probability information described above in the section "Quantifying Correlations in Probability Information". In some examples, the processing device may determine the probability information using any of the methods described in that section. In other examples, the processing device may receive an indication of the probability information. The processing device may receive the indication of the probability information from a network device or from another device. Alternatively, the processing device may be configured with the indication of the probability information.

[0218] The indication of the probability information may include the probability information itself. For example, the processing device may receive any of the joint and / or conditional probabilities described above. Alternatively, the indication may take another form. In a specific example, the indication may include an identifier used with a lookup table available at the processing device (e.g., stored at the processing device). For example, the lookup table may also be referred to as a correlation table or a confidence table. The processing device may look up the identifier in the lookup table to determine the probability information. For example, a table such as Table 1 may be configured for the processing device.

[0219] ID Probability meaning 00 0 zero 01 <0.5 Low 10 <1 medium 11 1 high

[0220] Table 1

[0221] Table 1 is an example of a lookup table that can be used in some embodiments of the present invention. The table has three columns. The first column includes an identifier (ID) that the processing device can receive (e.g., from a network device). The second column includes a probability or probability range that the processing device can determine by looking up the associated identifier in the lookup table. The third column (which can be omitted) includes the meaning of the associated probability. Each probability in the lookup table can be the probability of obtaining a potential result and another potential result from the component reasoning process. Thus, for example, the identifier (ID) 00 can be associated with the conditional probability of detecting an object in the category "giraffe" in the image when an object in the category "stop sign" is also detected in the image.

[0222] Table 2 shows another example of a lookup table that may be used.

[0223] ID Joint / Conditional Probability meaning 00 0 independent 01 <0.5 Weak correlation 10 <1 Moderate correlation 11 1 Strong correlation

[0224] Table 2

[0225] In some examples, the processing device may store the received indication so that the indication can be retrieved from the memory at the processing device to determine the probability information using a lookup table (e.g., as needed). Table 3 shows an example of a table that stores corresponding identifiers for each pair of categories in a set of categories 1, 2, and 3, which can be used with a lookup table (e.g., Table 1 or Table 2) to determine the probability information for the corresponding pair. For example, each processing device may store a table of identifiers for each combination of possible results from a component reasoning process. For example, a network device may store a corresponding table. The table stored at the network device may include an identifier for each combination of possible results from a distributed reasoning process (e.g., possible results from all component reasoning processes that constitute part of a distributed reasoning process). The network device may store a larger table than any individual processing device.

[0226] Correlation ID Category 1 Category 2 Category 3 Category 1 11 10 01 Category 2 10 11 00 Category 3 01 00 11

[0227] Table 3

[0228] Although the above description relates to the processing device receiving one identifier, it should be understood that in general the processing device may receive one or more identifiers. In some examples, the processing device may receive an identifier for each combination of potential results that can be obtained from the component reasoning process. For example, the processing device may receive a table such as Table 3, which includes corresponding identifiers for each pair of categories that can be obtained from the component classification process. In general, the probability information can be quantized and / or encoded in the indication in any suitable manner.

[0229] The processing device may also determine a second reasoning result based on the first reasoning result and the probability information. Therefore, the processing device may refine the first reasoning result based on the probability information. For example, any method described above in the "Refine Reasoning Result" section may be used to perform this step.

[0230] In step 1206 , the processing device sends the second inference result to the network device.

[0231] The method 1200 may further involve: updating the probability information based on the first reasoning result and / or the second reasoning result.

[0232] The processing device may indicate the updated probability information to the network device. The processing device may send the probability information itself to the network device. Alternatively, the processing device may indicate an update to the probability information using other means. For example, the processing device may send an updated identifier for each combination of potential outcomes. The network device may determine the probability of a combination of potential outcomes based on the updated identifier (e.g., using a lookup table, such as any of the above-described lookup tables).

[0233] Additionally or alternatively, the processing device may indicate the updated probability information to another device (e.g., another processing device). In some examples, the processing device may send the updated probability information to another processing device (e.g., to another device for performing component reasoning as part of a distributed learning process). The processing device may indicate the updated probability information to another device using any of the methods described above with respect to indicating the updated probability information to the network device.

[0234] In some examples, a processing device may receive an indication of an update to the probability information. The processing device may receive the indication from a network device or another device (e.g., another processing device (e.g., another device for performing a component reasoning process as part of a distributed reasoning process)). In some examples, each processing device may exchange its (updated) probability information with at least one other processing device. The updated probability information at a particular processing device may be referred to as local relevance information. Thus, for example, each processing device may exchange its local relevance information with its neighboring processing devices.

[0235] The processing device may update the probability information based on the received indication of the update to the probability information.The processing device may update the probability information using any of the methods described above in the section "Updating the probability information".

[0236] Also provided is an apparatus for performing the method 1200. In yet another aspect, a memory (e.g., a non-transitory processor-readable medium) is provided. The memory includes instructions (e.g., processor-executable instructions) that, when executed by a processor of the apparatus, cause the apparatus to perform the method 1200. In yet another aspect, provided is an apparatus including a processor and a memory. The memory includes instructions (e.g., processor-executable instructions) that, when executed by the processor, cause the apparatus to perform the method 1200.

[0237] Fig.13 FIG. 1 is a flowchart of a method 1300 performed by a network device according to an embodiment of the present invention. The network device may be a TRP, such as Figures 1 to 4 Any TRP 170 described. The network device can be the network device mentioned in the description of the method 1200 above.

[0238] In step 1302, the network device sends a corresponding first input of the plurality of first inputs to each of the plurality of processing devices. Some or all of the plurality of processing devices may be used to perform the above method 1200. Thus, for example, some or all of the processing devices may be the processing devices described above with respect to method 1200. The processing device may be an electronic device. In some examples, the processing device may be connected to a cell served by the network device.

[0239] Each processing device in the plurality of processing devices is used to perform a component reasoning process as part of a distributed reasoning process representing a machine learning process. The component reasoning process, the distributed reasoning process, and / or the machine learning process may be as described above with respect to method 1200.

[0240] Each first input is used in a component inference process at a corresponding processing device.

[0241] The method may also involve: the network device indicating the probability information to the plurality of processing devices. The probability information is defined as described above in the description of the method 1200. The processing device may receive the indication of the probability information according to receiving the indication of the probability information in the above method 1200. Thus, for example, the network device may send the probability information itself or an indication (e.g., an identifier) ​​that enables the processing device to determine the probability information to the network device.

[0242] In some examples, the network device may indicate the same probability information to all processing devices. For example, the network device may broadcast the same probability information. For example, the same probability information may indicate the probability of obtaining a corresponding potential result and another potential result from a plurality of potential results, and the potential results may be obtained from a distributed reasoning process. Therefore, the plurality of potential results may be results obtainable by any component reasoning process. The same probability information may also be referred to as global probability information or global correlation information. Therefore, the network device may perform a broadcast to distribute the global correlation information to the reasoning unit (e.g., to the processing device).

[0243] In other examples, the probability information indicated by the network device may be specific to a specific processing device. The probability information may be specific to the component reasoning process that the specific processing device is used to perform. For different processing devices, the multiple potential results that can be obtained from the component reasoning process may be different. For example, one processing device may be able to classify input data using a smaller number of categories than another processing device. Therefore, the network device may only provide probability information related to potential results to the specific processing device, and the potential results may be obtained from the component reasoning process to be performed by the specific processing device. Sending only probability information related to a specific processing device can save transmission resources of the network device and the processing device, while also saving memory at the processing device.

[0244] For example, a network device may store probability information including a 100×100 table including, for each of 100 categories, the corresponding probability of each category co-occurring with another category. However, a first processing device referred to as user equipment A (UEA) may be used to classify data according to only 50 categories. A second processing device referred to as user equipment B (UEB) may be used to classify data according to 5 categories. Therefore, the network device may extract a 50×50 table from the complete 100×100 table and send the 50×50 table to UEA as probability information. The network device may extract a 5×5 table from the complete 100×100 table and send the 5×5 table to UEB as probability information. This may reduce the communication overhead between the network device and UEA and UEB, and save memory of UEA and UEB.

[0245] Thus, in some examples, the plurality of potential results obtainable from the component reasoning process may be a first plurality of potential results, and the network device may obtain second probability information indicating, for each potential result in the second plurality of potential results obtainable from the distributed reasoning process, the probability of obtaining the corresponding potential result and another potential result from the second plurality of potential results. The first plurality of potential results may be a subset of the second plurality of potential results. The network device may select the first probability information from the second probability information based on the first plurality of potential results for one or more (each) of at least one of the plurality of processing devices. Thus, the network device may send a subset of the probability information to the corresponding processing device for one or more of the processing devices in the processing devices based on the plurality of potential results obtainable by the component reasoning process to be performed by the corresponding processing device.

[0246] In some examples, the probability information may be specific to a particular location (e.g., a particular area). For example, a processing device may be provided with specific probability information based on its location. This may be particularly suitable for situations where the input to the component reasoning process at a particular processing device is specific to the location of the processing device. By using location-specific probability information, reasoning can be customized according to the location of the processing device, which can further improve performance. For example, an electronic device entering an area (e.g., a cell) of a network may be provided with probability information based on the area (e.g., specific to the cell). In a sense, the probability information may reflect historical knowledge (or prior knowledge) about where it was created.

[0247] In some examples, the processing device may not receive the probability information from the network device. Instead, the processing device may obtain the probability information by other means. For example, the probability information may be configured for the processing device or received from another device.

[0248] In step 1304, the network device receives a first reasoning result from a plurality of processing devices. The first reasoning result is obtained at the processing device based on the probability information and the plurality of first inputs. Therefore, before the first reasoning result received in step 1304 is sent to the network device, the first reasoning result is refined at the processing device based on the probability information. For example, the first reasoning result may be obtained using any of the methods described above in the section “Refining the Reasoning Result”. Step 1304 may correspond to step 1206 described above.

[0249] The network device may also receive updates to the probability information from at least one processing device. In some examples, the network device may receive corresponding updates from all processing devices. The update may be based on the first reasoning result obtained from the component reasoning process. The update may be determined using any of the methods described above in the "Updating Probability Information" section. At least one processing device may send an update to the network device based on the indication of the updated probability information of the network device described above in method 1200.

[0250] The network device may update its probability information (e.g., global probability information) based on updates received from at least one processing device. In the simplest example, the network device may replace its probability information with updates received from a processing device. In another example, the network device may average the probability information received from all processing devices in at least one processing device. The network device may weight the average based on, for example, the size of the input processed by the respective processing device and / or the confidence (e.g., trust) of the respective processing device.

[0251] As described above, the probability information may be specific to a particular processing device and / or a particular location. Therefore, a network device may update the probability information of the device based only on updates received from a particular processing device. In another example, a network device may update the probability information of a location based on updates received from one or more processing devices in or associated with a location (e.g., a particular area). This may enable a network device or wireless network to adaptively match a target application based on a particular geographic location and / or a particular scenario.

[0252] In the above description of method 1300, the network device sends multiple first inputs to multiple processing devices as part of a distributed reasoning process. In some embodiments, the multiple first inputs include multiple second inputs and at least one redundant input. The at least one redundant input is redundant because it includes data that is also included in the multiple second inputs. Thus, for example, the at least one redundant input can be used to recover lost reasoning results from the distributed reasoning process and / or refine the reasoning results from the distributed reasoning process.

[0253] Thus, the distributed reasoning process may include an encoding reasoning process. The network device may process multiple second inputs to generate at least one redundant input before sending the first input. This processing may be referred to as encoding because it provides redundancy in a manner similar to encoding theory.

[0254] In some examples, multiple second inputs are processed so that each redundant input in at least one redundant input includes a cascade of data from at least two second inputs in the multiple second inputs. In this context, cascading can refer to combining data from at least two of the multiple inputs without mixing data from different inputs. Therefore, for example, data from at least two of the multiple inputs can be combined into a common data set without superimposing (e.g., adding) data from different inputs. For example, data from at least two of the multiple inputs can be placed side by side in the same data set. For example, data from one input can be attached to another input. In another example, data from, for example, three or more data sets can be tiled. Tiling may be particularly suitable for data with two or more dimensions. By generating redundant inputs in this way, the same component reasoning process as the component reasoning process used at the processing device performing reasoning on the second input can be used at the processing device performing reasoning on the redundant input.

[0255] Method 1300 may also involve: after the network device receives the first reasoning result from the multiple processing results, decoding the first reasoning result to obtain the second reasoning result. The first reasoning result includes a redundant result based on at least one redundant input and other first results based on multiple second inputs. At least two first reasoning results may be decoded, wherein at least two first reasoning results include redundant results. In other words, at least one redundant result of the one or more redundant results and zero or more other first results may be decoded. Therefore, for example, decoding may be performed based on two or more redundant results. Alternatively, at least one redundant result of the redundant results or one or more other first results may be decoded.

[0256] For example, the second inference result may include an estimate of a missing result from an instance of the same component inference process (e.g., a result that should have been returned by the processing device but was not returned). Even if there is no data loss in the distributed inference process, using the process decoding results and redundant results is still advantageous because it can provide a more accurate and / or more insightful second inference result.

[0257] In an example where the at least one redundant input includes a concatenation of data from at least two of the plurality of second inputs, the network device may decode the first inference result by performing one or more linear operations and / or one or more set operations. Performing the one or more set operations may be particularly suitable for an example where the machine learning process includes a classification process such that the first inference result includes a plurality of categories and one or more redundant categories.

[0258] There are multiple ways to decode the first inference result using linear operations. For example, this may depend on the first inference result, the distributed inference process, and / or the inference being sought. In this context, a linear operation is any operation that preserves vector addition and scalar multiplication operations. Thus, for all x and y, and all constants a, the one or more linear operations may include any operation f(.) that satisfies the following equation:

[0259] f(x+y)=f(x)+f(y);

[0260] f(ax)=af(x).

[0261] Performing one or more linear operations may be particularly suitable for examples of machine learning processes including regression processes.

[0262] There are multiple ways to use set operations to decode the first inference result. Performing one or more set operations may be particularly suitable for examples where the machine learning process includes a classification process, so that the first inference result includes multiple categories and one or more redundant categories. In some examples, a confidence propagation process (e.g., an algorithm) can be used to decode the first inference result. This can be illustrated by the following example: the redundant categories constitute a set R, the categories in the first inference result constitute a set S, and N is a neighbor set. By performing the following steps one or more times, category i in the first inference result can be decoded to obtain category j of the second inference result:

[0263] {class j→i}={class i' ,i'∈N(j)∩R}-union({class i”}for all i”∈N(j)∪Sand i”≠i)

[0264] {class i→j}={class i}∪union({class j'→i}for all j'∈N(i)and j'≠j)

[0265] {class i}={class i}∪union({class j'→i}for all j'∈N(i))

[0266] Among them, ∪ represents the union of two categories, "union" represents the union of more than two categories, "–" represents set difference, and ∩ represents set intersection. The neighbor set N of a particular category j may include each of the labels used to infer category j. This particular belief propagation process may reduce the complexity of decoding because the category in the second inference result may be determined without performing an exhaustive search. The belief propagation process may be particularly suitable for the case where sparse codes are used for encoding because the belief propagation process converges faster for sparse codes.

[0267] Although this example of the belief propagation algorithm uses union, intersection, and difference set operations, any suitable set operation may generally be used to decode the first inference result. Thus, for example, the one or more set operations may include one or more of: union, intersection, complement, and difference.

[0268] A network device for performing the method 1300 is also provided. In yet another aspect, a memory (e.g., a non-transitory processor-readable medium) is provided. The memory includes instructions (e.g., processor-executable instructions) that, when executed by a processor of the network device, cause the network device to perform the method 1300. In yet another aspect, a network device is provided, including a processor and a memory. The memory includes instructions (e.g., processor-executable instructions) that, when executed by the processor, cause the network device to perform the method 1300.

[0269] In method 1300, before receiving the first inference result to be received in step 1304 at the network device, the first inference result is refined at the processing device based on the probability information. In other embodiments, the network device may use the probability information to refine the inference results provided by the component inference process. Fig.14 Describe, Fig.14 FIG. 14 is a flowchart of a method 1400 performed by a network device according to an embodiment of the present invention. The network device may be a TRP, such as Figures 1 to 4 Any TRP 170 described.

[0270] Method 1400 may be substantially the same as method 1300, except that the network device, rather than the processing device, refines the inference result based on the probability information.

[0271] Method 1400 involves, at step 1402, a network device sending a plurality of first inputs to a plurality of processing devices. Each respective first input of the plurality of first inputs is used for a component reasoning process at a respective processing device as part of a distributed reasoning process representing a machine learning process. Step 1402 may be performed in accordance with step 1302.

[0272] In step 1404, the network device receives first reasoning results based on multiple first inputs from multiple processing devices. Step 1404 can be performed according to step 1304, except that the first reasoning result in step 1404 is not based on probability information. Therefore, in method 1400, the processing device may not be able to obtain the probability information described above in method 1300.

[0273] In step 1406, the network device determines a second reasoning result based on the first reasoning result and the probability information. The probability information is defined as described in the description of methods 1200 and 1300 above.

[0274] The network device 1406 may also update the probability information based on the first reasoning result and / or the second reasoning result. The network device 1406 may update the probability information based on the reasoning results from all processing devices. Alternatively, the network device 1406 may maintain different probability information for each processing device or group of processing devices. Thus, for example, the network device 1406 may update the probability information of the group based on the reasoning results of the processing devices in the group of one or more processing devices. The network device may update the probability information using any of the methods described above in the "Updating Probability Information" section.

[0275] In the above description of method 1400, the network device sends multiple first inputs to multiple processing devices as part of a distributed reasoning process. In some embodiments, the multiple first inputs include multiple second inputs and at least one redundant input. Therefore, the distributed reasoning process may include an encoded reasoning process. This can be implemented in the same manner as the encoded reasoning process described above with respect to method 1300.

[0276] Therefore, the network device may encode the plurality of second inputs to generate at least one redundant input. The network device may also decode the second inference result to obtain a third inference result. Therefore, the network device may decode the inference result after refining the inference result based on the probability information. Therefore, the network device may use redundancy to recover lost results and / or further refine the results after refining the inference result using the probability information.

[0277] A network device for performing the method 1400 is also provided. In yet another aspect, a memory (e.g., a non-transitory processor-readable medium) is provided. The memory includes instructions (e.g., processor-executable instructions) that, when executed by a processor of the network device, cause the network device to perform the method 1400. In yet another aspect, a network device is provided, including a processor and a memory. The memory includes instructions (e.g., processor-executable instructions) that, when executed by the processor, cause the network device to perform the method 1400.

[0278] Although many of the examples described herein are provided in the context of distributed reasoning, it should be appreciated that many of the techniques described herein are also applicable in the context of reasoning performed by a single device.

[0279] Fig.15 The flowchart of the method 1500 provided by the embodiment of the present invention is shown. The method 1500 can be implemented by any suitable device, such as an electronic device or a network device.

[0280] Method 1500 may involve obtaining input data. For example, the input data may be obtained by receiving the input data from another device. In some examples, the input data may be collected from multiple devices.

[0281] Method 1500 involves, in step 1502, performing an inference process on input data to obtain a first inference result. The inference process may involve a machine learning process, such as any suitable machine learning process, such as a neural network (e.g., a deep neural network (DNN)), a k-nearest neighbor process, a linear regression process, a logistic regression process, a support vector machine, or any other suitable machine learning process. For example, the inference process may include a regression process, a classification process (e.g., a classifier), or a combination of a regression process and a classification process. It will be appreciated by those skilled in the art that the selection of a machine learning process is typically specific to the inference task. For example, the inference task may include image classification, and the component process may include a neural network trained to classify images, such as a deep neural network.

[0282] In step 1504, method 1500 involves: determining a second reasoning result based on the first reasoning result and probability information, wherein the probability information indicates, for each potential result in a plurality of potential results that can be obtained from the reasoning process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results. Step 1504 can be performed according to determining the second reasoning result as described above in method 1200 or according to the above step 1406, wherein the reasoning process replaces the component reasoning process mentioned above.

[0283] The method 1500 may also involve sending the second inference result to another device. The other device may be the same device that provided the input data, or a different device. In some examples, the method 1500 may involve using the second inference result.

[0284] Also provided is an apparatus for performing the method 1500. In yet another aspect, a memory (e.g., a non-transitory processor-readable medium) is provided. The memory includes instructions (e.g., processor-executable instructions) that, when executed by a processor of the apparatus, cause the apparatus to perform the method 1500. In yet another aspect, provided is an apparatus including a processor and a memory. The memory includes instructions (e.g., processor-executable instructions) that, when executed by the processor, cause the apparatus to perform the method 1500.

[0285] result

[0286] Fig.16 and Fig.17 Simplified line graphs based on photos on which distributed inference representing an image classification process is performed are shown. In each figure, four copies of the corresponding image are shown. Boxes are overlaid to show the objects identified in each image. In both figures, the rightmost image shows the objects manually identified (e.g., by a person) in these images. These images represent the "ground truth," i.e., the information that the image classification process attempts to obtain.

[0287] exist Fig.16 , the ground truth image shows a first bird 1602, a first giraffe 1604, a second giraffe 1606, a second bird 1608, a third bird 1610, and a fourth bird 1612. Fig.17 In the ,background image shows three boats in the background, and a crowded scene of umbrellas, people, and deck chairs in the foreground.,In the ground truth image, all boats, umbrellas, people, and deck chairs are detected and,identified.

[0288] exist Fig.16 and Fig.17, three additional images are labeled from left to right “Detection”, “Related Detection”, and “Related Decoding”. These images show the objects detected in the images when using the three different methods. In all three methods, objects are detected and classified using the YOLOv3 model (Farhadi, Ali, and Joseph Redmon, “Yolov3: Incremental Improvements”, Computer Vision and Pattern Recognition, Berlin / Heidelberg, Germany: Springer, 2018), which is trained using the COCO-train2017 dataset (Lin, Tsung-Yi et al., “Microsoft coco: Common Objects in Context”, European Conference on Computer Vision, Springer, Cham, 2014). For the related detection image, the confidence associated with the initial classification provided by the YOLOv3 model is refined based on the probabilistic information using the methods described in this article. For the related decoded image, the probabilistic information and redundancy provided by the encoded inference are implemented using the methods described in this article to classify the objects in the image.

[0289] like Fig.16 As shown, all objects 1602 to 1610 except the fourth bird 1612 are detected in the detection image (e.g., have boxes around them), indicating that the YOLOv3 model is able to correctly detect and identify objects 1602 to 1610. These objects are also detected in the relevant detection and relevant decoded images, indicating that the use of probability information does not degrade classification performance. In addition, the fourth bird 1612 is detected and identified as a bird (e.g., has a box around it) in the relevant detection image and the relevant decoded image. This shows that using probability information to refine the inference results according to the method described in this article can improve the performance of the image classification process, especially the ability to correctly detect and identify small objects in the image. Although Fig.16 This advantage is illustrated, but the improvement also exists when using the method described herein to classify objects in other images.

[0290] like Fig.17 As shown, fewer objects are detected in the detection image compared to the associated detection and the associated decoded image. This suggests that using probabilistic information to refine the inference results according to the method described in this paper can improve the performance of the image classification process by correctly detecting and identifying more objects, especially more objects in crowded images. Fig.17 In the , more objects are detected and identified in the associated decoded image than in the associated detection image. This suggests that refining the inference results using the probabilistic information and redundancy provided by the encoded inference according to the method described in this paper can further improve the performance of the image classification process. The performance gains obtained by using both probabilistic information (also called relevant inference) and encoded inference simultaneously suggest that these techniques can enhance each other.

[0291] By reference Fig.18 The improved performance provided by the methods disclosed herein can be further illustrated by Fig.18 The object detection rate of the distributed reasoning process performed according to an embodiment of the present invention is shown.

[0292] exist Fig.18 In each example shown, a YOLOv3 model (Farhadi, Ali, and Joseph Redmon, "Yolov3: Incremental Improvements", Computer Vision and Pattern Recognition, Berlin / Heidelberg, Germany: Springer, 2018) was trained using the COCO-train2017 dataset (Lin, Tsung-Yi et al., "Microsoft coco: Common objects in context", European Conference on Computer Vision, Springer, Cham, 2014). Inference was performed on images from the COCO-val2017 dataset (Lin, Tsung-Yi et al., "Microsoft coco: Common objects in context", European Conference on Computer Vision, Springer, Cham, 2014) to detect 36,781 labeled objects in 5,000 images. Each input image is fed into the corresponding instance of the trained YOLOv3 model, which outputs bounding box estimates and class predictions for one or more objects detected in the image.

[0293] This process is repeated with deletion probabilities ranging from 0 to 0.8, where the deletion probability indicates the likelihood that each instance of the YOLOv3 model returns its corresponding output. So, for example, a deletion probability of 0 indicates that all outputs of the YOLOv3 model are returned.

[0294] The lower dashed line with circle marks (labeled "Inference") shows the detection rate of the inference process performed without using probabilistic information or redundancy (e.g., without using encoding). The dashed line with triangle marks (labeled "Correlated Inference") shows the detection rate of the distributed inference process according to an embodiment of the present invention, in which probabilistic information is used to refine the inference results. The solid line with star marks (labeled "(7,4) Hamming Coding Inference") shows the detection rate of object detection performed by the distributed inference process, in which the input image is encoded according to the (7,4) Hamming code with the following parity check matrix:

[0295]

[0296] Therefore, 7 inputs including 4 input images and 3 redundant input images are generated using the (7,4) Hamming code. The generator matrix of the Hamming code can be expressed as:

[0297]

[0298] The solid line with square markers (labeled "(24,12) 2nd-order coding inference") shows an implementation using a (24,12) code with 2nd order. For the (24,12) code, the images are grouped into batches of 12 images and 12 redundant images are generated for each batch so that 24 images are input to an instance of YOLOv3 for each batch. Since the (24,12) code is 2nd order, each redundant image contains data from two images. The parity check matrix of the (24,12) 2nd-order code can be expressed as:

[0299]

[0300] The solid line with triangle marks (labeled as “(24,12) 2nd order correlation coding inference”) shows the detection rate of the distributed inference process according to an embodiment of the present invention, in which probability information is used to refine the inference result and the input is encoded according to the above-mentioned (24,12) code.

[0301] like Fig.18 As shown, using probabilistic information to refine individual inferences (e.g., without any redundancy) results in significantly improved performance for all deletion rates, with the strongest improvement shown at low deletion rates. Performance improves further when probabilistic information is used in conjunction with the redundancy provided by encoded inference. This suggests that correlated inference and encoded inference can reinforce each other. In particular, correlated encoded inference can improve detection rates by more than 20% for at least some deletion rates.

[0302] Although many of the examples provided above are described in the context of image classification, it should be understood that the present invention is not limited to this, and aspects of the present invention can be implemented in a wide range of applications, such as network reasoning, environmental detection, and / or autonomous driving. Aspects of the present invention can be implemented in a wide range of system architectures. The embodiments described herein can be implemented in various communication networks, such as 5G, 6G, and Wi-Fi. In some cases, a network is not required. In some examples, aspects of the present invention can be implemented in next-generation mobile and wireless network services, cloud and edge computing services, and / or perception services. For example, aspects of the present invention can be implemented to achieve joint perception or detection in a wireless network.

[0303] It should be understood that one or more steps of the embodiment method provided herein can be performed by corresponding units or modules. For example, a signal can be sent by a sending unit or a sending module. A signal can be received by a receiving unit or a receiving module. A signal can be processed by a processing unit or a processing module. The corresponding unit / module can be hardware, software, or a combination thereof. For example, one or more units / modules can be integrated circuits, such as field programmable gate arrays (FPGA) or application-specific integrated circuits (ASIC). It should be understood that if these modules are software, these modules can be retrieved in whole or in part by a processor as needed, retrieved individually or collectively for processing, retrieved in one or more instances as needed, and these modules themselves can include instructions for further deployment and instantiation.

[0304] Although combinations of features are shown in the illustrated embodiments, it is not necessary to combine all of the features to achieve the advantages of the various embodiments of the present invention. In other words, a system or method designed according to an embodiment of the present invention does not necessarily include all of the features shown in any of the drawings or all of the parts schematically shown in the drawings. In addition, selected features of an exemplary embodiment may be combined with selected features of other exemplary embodiments.

[0305] Although the present invention has been described with reference to illustrative embodiments, this specification is not to be interpreted in a limiting sense. According to the above guidance, many modifications and variations of the present invention may be made. Therefore, it should be understood that within the scope of the appended claims, the present invention may be implemented in a manner different from that specifically described herein.

Claims

1. A method, characterized in that include: receiving inputs of component inference processes from a network device, the component inference processes forming part of a distributed inference process representing a machine learning process; executing the component reasoning process on the input to obtain a first reasoning result; Sending a second inference result based on the first inference result and probability information to the network device, wherein the probability information indicates, for each potential result of a plurality of potential results obtainable from the component inference process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

2. The method according to claim 1, characterized in that: The machine learning process includes a classification process, The plurality of potential results include a plurality of categories, such that the probability information indicates, for each category in the plurality of categories obtainable from the component reasoning process, a probability of obtaining the corresponding category and another category from the plurality of categories, The first inference result includes a corresponding first confidence c1(i) for each category i in the plurality of categories, and the second inference result includes a corresponding second confidence c2(i) for each category in the plurality of categories.

3. The method according to claim 2, characterized in that The probability information includes: for each category in the plurality of categories, assuming that other categories j have been obtained from the component reasoning process, obtaining the corresponding conditional probability P(i, j|j) of the corresponding category i and the other category j from the plurality of potential results, and the method further includes determining the corresponding second confidence c2(i) of the corresponding category i in the plurality of categories according to the following equation:

4. The method according to claim 1, characterized in that: The machine learning process includes a regression process, The plurality of potential outcomes includes a plurality of values ​​for a parameter, The first inference result includes a first probability distribution P1(x i ), the second inference result includes a second probability distribution P2(x I ).

5. The method according to claim 4, characterized in that The probability information includes obtaining the first value x of the parameter i and obtain the second value of the parameter x j The joint probability P(x i ,x j ), and the method further comprises determining the second probability distribution according to the following equation: P2(x i )=∫P1(x j )P(x i ,x j )dx j 。 6. The method according to claim 4, characterized in that The probability information includes assuming that the second value x of the parameter has been obtained j , get the first value x of the parameter i The conditional probability P(x i |x j ), and the method further comprises determining the second probability distribution according to the following equation: P2(x i )=∫P1(x j )P(x i |x j )dx j 。 7. The method according to any one of claims 1 to 6, characterized in that Also includes: An indication of the probability information is received from the network device.

8. The method according to any one of claims 1 to 7, characterized in that Also includes: The probability information is updated based on the first inference result or the second inference result.

9. The method according to claim 8, characterized in that Also includes one or more of the following: indicating updated probability information to the network device; The updated probability information is indicated to a device configured to perform inference as part of the distributed inference process.

10. A method performed by a network device, characterized in that: The method comprises: sending a respective first input of a plurality of first inputs to each processing device of a plurality of processing devices, each respective first input for a component reasoning process as part of a distributed reasoning process representing a machine learning process; A first inference result obtained based on probability information and the multiple first inputs is received from the multiple processing devices, wherein the probability information indicates, for each potential result among a plurality of potential results obtainable from the component inference process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

11. The method according to claim 10, characterized in that The plurality of first inputs include a plurality of second inputs and at least one redundant input, the method further comprising: encoding the plurality of second inputs to generate the at least one redundant input; The first inference result is decoded to obtain a second inference result.

12. The method according to claim 10 or 11, characterized in that: Also includes: A respective update to the probability information is received from at least one processing device of the plurality of processing devices, the respective update being based on the first inference results obtained from the component inference processes.

13. The method according to any one of claims 10 to 12, characterized in that Also included is indicating the probability information to the plurality of processing devices by indicating the same probability information to each of the plurality of processing devices.

14. The method according to any one of claims 10 to 12, characterized in that Also included is indicating the probability information to at least one of the plurality of processing devices by indicating first probability information specific to the respective processing device to the plurality of processing devices.

15. The method according to claim 14, characterized in that The plurality of potential results obtainable from the component reasoning process is a first plurality of potential results, the method further comprising: Obtaining second probability information, the second probability information indicating, for each potential result in a second plurality of potential results obtainable from the distributed reasoning process, a probability of obtaining the corresponding potential result and another potential result from the second plurality of potential results; For each processing device of the at least one processing device of the plurality of processing devices, the first probability information is selected from the second probability information based on the first plurality of potential outcomes.

16. A method performed by a network device, characterized in that: The method comprises: sending a respective first input of a plurality of first inputs to each processing device of a plurality of processing devices, the respective first inputs being used for a component reasoning process as part of a distributed reasoning process representing a machine learning process; receiving, from the plurality of processing devices, first inference results based on the plurality of first inputs; A second reasoning result is determined based on the first reasoning result and probability information, wherein the probability information indicates, for each potential result among a plurality of potential results obtainable from the component reasoning process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

17. The method according to claim 16, characterized in that The plurality of first inputs include a plurality of second inputs and at least one redundant input, the method further comprising: encoding the plurality of second inputs to generate the at least one redundant input; The second inference result is decoded to obtain a third inference result.

18. A method, characterized in that include: Performing a reasoning process on the input data to obtain a first reasoning result; A second reasoning result is determined based on the first reasoning result and probability information, wherein the probability information indicates, for each potential result among a plurality of potential results obtainable from the reasoning process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

19. A device, characterized in that: include: processor; a memory storing instructions that, when executed by the processor, cause the apparatus to: receiving inputs of component inference processes from a network device, the component inference processes forming part of a distributed inference process representing a machine learning process; executing the component reasoning process on the input to obtain a first reasoning result; Sending a second inference result based on the first inference result and probability information to the network device, wherein the probability information indicates, for each potential result of a plurality of potential results obtainable from the component inference process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

20. The device according to claim 19, characterized in that: The machine learning process includes a classification process, The plurality of potential results include a plurality of categories, such that the probability information indicates, for each category in the plurality of categories obtainable from the component reasoning process, a probability of obtaining the corresponding category and another category from the plurality of categories, The first inference result includes a corresponding first confidence c1(i) for each category i in the plurality of categories, and the second inference result includes a corresponding second confidence c2(i) for each category in the plurality of categories.

21. The device according to claim 20, characterized in that The probability information includes: for each category in the plurality of categories, assuming that other categories j have been obtained from the component reasoning process, obtaining the corresponding conditional probability P(i, j|j) of the corresponding category i and the other category j from the plurality of potential results, and, when the instructions are executed by the processor, further causing the device to determine the corresponding second confidence c2(i) of the corresponding category i in the plurality of categories according to the following equation:

22. The device according to claim 19, characterized in that: The machine learning process includes a regression process, The plurality of potential outcomes includes a plurality of values ​​for a parameter, The first inference result includes a first probability distribution P1(x i ), the second inference result includes a second probability distribution P2(x i ).

23. The device according to claim 22, characterized in that The probability information includes obtaining the first value x of the parameter i and obtain the second value of the parameter x j The joint probability P(x i ,x j ), and when the instructions are executed by the processor, the device further determines the second probability distribution according to the following equation: P2(x i )=∫P1(x j )P(x i ,x j )dx j 。 24. The device according to claim 22, characterized in that The probability information includes: assuming that the second value x of the parameter has been obtained j , get the first value x of the parameter i The conditional probability P(x i |x j ), and when the instructions are executed by the processor, the device further determines the second probability distribution according to the following equation: P2(x i )=∫P1(x j )P(x i |x j )dx j 。 25. The device according to any one of claims 19 to 24, characterized in that When the instructions are executed by the processor, the device is also caused to: An indication of the probability information is received from the network device.

26. The device according to any one of claims 19 to 25, characterized in that When the instructions are executed by the processor, the device is also caused to: The probability information is updated based on the first inference result or the second inference result.

27. The device according to claim 26, characterized in that When the instructions are executed by the processor, the device is further caused to perform one or more of the following operations: indicating updated probability information to the network device; The updated probability information is indicated to a device configured to perform inference as part of the distributed inference process.

28. A network device, characterized in that: include: processor; a memory storing instructions that, when executed by the processor, cause the network device to: sending a respective first input of a plurality of first inputs to each processing device of a plurality of processing devices, each respective first input for a component reasoning process as part of a distributed reasoning process representing a machine learning process; A first inference result obtained based on probability information and the multiple first inputs is received from the multiple processing devices, wherein the probability information indicates, for each potential result among a plurality of potential results obtainable from the component inference process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

29. The network device according to claim 28, characterized in that: The plurality of first inputs include a plurality of second inputs and at least one redundant input, and when the instructions are executed by the processor, the network device is further caused to: encoding the plurality of second inputs to generate the at least one redundant input; The first inference result is decoded to obtain a second inference result.

30. The network device according to claim 28 or 29, characterized in that: When the instructions are executed by the processor, the network device is further caused to: A respective update to the probability information is received from at least one processing device of the plurality of processing devices, the respective update being based on the first inference results obtained from the component inference processes.

31. The network device according to any one of claims 28 to 30, characterized in that: When the instructions are executed by the processor, the network device further causes the network device to indicate the probability information to the plurality of processing devices by indicating the same probability information to each of the plurality of processing devices.

32. The network device according to any one of claims 28 to 30, characterized in that: When the instructions are executed by the processor, the instructions further cause the network device to indicate the probability information to at least one of the plurality of processing devices by indicating first probability information specific to the corresponding processing device to the plurality of processing devices.

33. The network device according to claim 32, characterized in that: The plurality of potential results obtainable from the component reasoning process is a first plurality of potential results, and when the instructions are executed by the processor, further cause the network device to: Obtaining second probability information, the second probability information indicating, for each potential result in a second plurality of potential results obtainable from the distributed reasoning process, a probability of obtaining the corresponding potential result and another potential result from the second plurality of potential results; For each processing device of the at least one processing device of the plurality of processing devices, the first probability information is selected from the second probability information based on the first plurality of potential outcomes.

34. A network device, characterized in that: include: processor; a memory storing instructions that, when executed by the network device, cause the apparatus to: sending a respective first input of a plurality of first inputs to each processing device of a plurality of processing devices, the respective first inputs being used for a component reasoning process as part of a distributed reasoning process representing a machine learning process; receiving, from the plurality of processing devices, first inference results based on the plurality of first inputs; A second reasoning result is determined based on the first reasoning result and probability information, wherein the probability information indicates, for each potential result among a plurality of potential results obtainable from the component reasoning process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

35. The network device according to claim 34, characterized in that: The plurality of first inputs include a plurality of second inputs and at least one redundant input, and when the instructions are executed by the processor, the network device is further caused to: encoding the plurality of second inputs to generate the at least one redundant input; The second inference result is decoded to obtain a third inference result.

36. A device, characterized in that include: processor; a memory storing instructions that, when executed by the processor, cause the apparatus to: Performing a reasoning process on the input data to obtain a first reasoning result; A second reasoning result is determined based on the first reasoning result and probability information, wherein the probability information indicates, for each potential result among a plurality of potential results obtainable from the reasoning process, a probability of obtaining a corresponding potential result and another potential result from the plurality of potential results.

37. A non-transitory computer readable medium, characterized in that A program for execution by a processor is stored, the program comprising instructions for executing the method according to any one of claims 1 to 18.

38. A computer program product, characterized in that The computer program product comprises instructions which, when the program is executed by a computer, cause the computer to perform the method according to any one of claims 1 to 18.

39. A device, characterized in that The device comprises a processor configured to cause the device to perform the method according to any one of claims 1 to 18.

40. A processor of a device, characterized in that The processor is configured to cause the apparatus to execute the method according to any one of claims 1 to 18.

41. A system, characterized in that: include: A first device for obtaining a first inference result as part of a distributed inference process representing a machine learning process; A second device in communication with the first device, the second device being used to obtain a second inference result as part of the distributed inference process, the second inference result being based on the first inference result and probability information, wherein the probability information indicates, for each potential result of a plurality of potential results obtainable from the inference process, a probability of obtaining the corresponding potential result and another potential result from the plurality of potential results.

Images