Distributed reasoning method and device

By distributing machine learning processes to multiple devices in a wireless communication network and introducing redundant inputs, the problems of wireless link unreliability and latency are solved, and inference performance and reliability are improved. Low-cost hardware suitable for IoT devices are suitable for IoT devices. reasoning.

CN118922842APending Publication Date: 2024-11-08HUAWEI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202280093962.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2022-04-07
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

In wireless communication networks, the inference performance of machine learning processes is affected by the quality, reliability and latency of transmission between devices, core networks and remote data centers, especially when IoT devices connect through wireless links, resulting in data transmission not being possible Reliability and severe latency limit the application of machine learning services.

Method used

By distributing machine learning processes across multiple devices, introducing redundant inputs to generate redundant outputs, and performing the same component inference process in the processing device, ensuring that the redundant output is recognizable on each device without the need for special training of redundancy Reasoning unit, improves the reliability and efficiency of reasoning.

Benefits of technology

Reduces the computational and traffic load per device, allows the use of low-cost hardware to perform inference, improves inference performance and reliability, and enhances the inference capability of IoT devices, especially in wireless communication environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118922842A_ABST
    Figure CN118922842A_ABST
Patent Text Reader

Abstract

Aspects of the invention relate to reasoning, in particular to distributed reasoning representing machine learning processes. Predicted reasoning will become a key service in a 6G wireless network. Aspects of the present application relate to applying aspects of coding theory to distributed reasoning to introduce redundancy that improves the accuracy and robustness of reasoning. A method of decoding output from a distributed reasoning process is also provided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to reasoning, and more particularly to distributed reasoning for representing machine learning processes. Background Art

[0002] Future wireless communication systems (e.g., sixth-generation or "6G" wireless communication) are expected to develop toward increasingly diverse application scenarios. Many applications are expected to use artificial intelligence (AI) such as machine-learning (ML) to provide services for 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 can 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, such 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.

[0004] However, this can severely limit the use cases to which machine learning can be applied. Many devices connect to the network via wireless links, which are unreliable and have high latency. This problem is particularly acute for wireless links in Internet of Things (IoT) networks. As a result, it can be extremely challenging for machine learning services to provide data to or use data from IoT devices over wireless networks. Summary of the invention

[0005] Systems and methods are provided to facilitate distributing inference work to multiple devices, such that each device performs one or more tasks as part of a machine learning process. This can reduce the computational load on each device compared to a situation where one device performs the entire inference work, while also reducing the amount of data that each device needs to transmit as part of the machine learning process (e.g., reducing the traffic load). Because the computational load and traffic load on each device are reduced, low-complexity devices such as IoT devices can be used to perform inference. This means that inference can be performed using low-cost hardware that can even be powered by batteries.

[0006] By distributing the machine learning process across multiple devices in this manner, inference can be performed closer to the data source (e.g., closer to the client device providing the data for inference). For example, the machine learning process can be implemented by multiple devices (e.g., client devices) in an access network or core network.

[0007] However, since the device performing the inference may be a low-cost and / or low-power device, the processing and transmission capabilities of the device may be limited. This means that the inference task performed by a specific device may fail due to computational and / or transmission errors, which may affect the performance of the entire inference process.

[0008] Various aspects of the present invention relate to distributed reasoning processes representing machine learning processes. Redundancy is introduced into the distributed reasoning process by encoding multiple inputs to the distributed reasoning process to generate one or more redundant inputs. As part of the distributed reasoning process, multiple inputs and one or more redundant inputs are input into the same component reasoning process. This method of generating redundant inputs is independent of the machine learning process, which means that the same component reasoning process can be implemented at each device in the devices involved in the distributed reasoning process. Therefore, for example, the same component reasoning process trained using the same machine learning method and training data can be implemented at each processing device in the corresponding processing device, which is a significant practical advantage compared to traditional encoded reasoning methods that use specially trained redundant reasoning units.

[0009] On the one hand, a method is provided. The method includes: obtaining multiple inputs of a distributed reasoning process representing a machine learning process, wherein each of the multiple inputs is used for the same component reasoning process of the distributed reasoning process. The method also includes: encoding the multiple inputs to generate one or more redundant inputs, so that each of the one or more redundant inputs includes a cascade of data from corresponding at least two of the multiple inputs. Each of the one or more redundant inputs is used for the same component reasoning process of the distributed reasoning process. The method also includes: as part of the distributed reasoning process, for each of the multiple inputs and each of the one or more redundant inputs, sending the corresponding input to a corresponding processing device for executing the same component reasoning process.

[0010] In another aspect, sending the respective input to the respective processing device may include sending the respective input to the respective processing device via a wireless communication link.

[0011] On the other hand, the method may further include: for each redundant input of the one or more redundant inputs, adjusting the size of the corresponding redundant input so that the corresponding redundant input has the same dimension as at least one input of the multiple inputs.

[0012] The resizing may include at least one of: cropping, downsampling, interpolation, or padding.

[0013] On the other hand, the plurality of inputs may form an ordered data set, and for each redundant input of the one or more redundant inputs, the corresponding at least two inputs of the plurality of inputs may be adjacent in the ordered data set.

[0014] In another aspect, the machine learning process comprises a deep neural network.

[0015] In another aspect, a method is provided. The method includes obtaining multiple outputs of a distributed reasoning process representing a machine learning regression process. The multiple outputs include multiple results and multiple redundant results. Each of the multiple results is determined by the same component reasoning process of the distributed reasoning process based on a corresponding input in a plurality of inputs. Each of the one or more redundant results is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, and the corresponding redundant input includes a cascade of data from corresponding at least two inputs of the multiple inputs. The method also includes: performing one or more linear operations on at least two results to decode the multiple outputs to obtain reasoning data. The at least two results come from the multiple results and the one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results.

[0016] On the other hand, performing the one or more linear operations to decode the plurality of outputs to obtain inference data may include performing the one or more linear operations to determine missing outputs in the inference process.

[0017] On the other hand, the reasoning process may also represent a machine learning classification process. The multiple outputs may also include multiple labels and one or more redundant labels. The method may also include: performing one or more set operations on at least two labels to decode the multiple outputs to obtain the reasoning data. The at least two labels may include at least one redundant label of the one or more redundant labels.

[0018] On the other hand, a method is provided. The method includes obtaining multiple outputs from a distributed reasoning process representing a machine learning classification process. The multiple outputs include multiple labels and one or more redundant labels. Each of the multiple labels is determined by the same component reasoning process of the distributed reasoning process based on a corresponding input in a plurality of inputs. Each of the one or more redundant labels is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, and the corresponding redundant input includes a cascade of data from corresponding at least two inputs of the multiple inputs. The method also includes: performing one or more set operations on at least two labels to decode the multiple outputs to obtain reasoning data. The at least two labels come from the multiple labels and the one or more redundant labels, and the at least two labels include at least one redundant label of the one or more redundant labels.

[0019] On the other hand, performing the one or more set operations to decode the plurality of outputs to obtain inference data may include performing the one or more set operations to determine missing outputs in the inference process.

[0020] On the other hand, the reasoning process may also represent a machine learning regression process. The multiple outputs may also include multiple results and one or more redundant results. The method may also include: performing one or more linear operations on at least two results to decode the multiple outputs to obtain the reasoning data. The at least two results may come from the multiple results and the one or more redundant results, and the at least two results may include at least one redundant result of the one or more redundant results.

[0021] On the other hand, performing the one or more set operations to decode the plurality of outputs to obtain the inference data may include: performing a belief propagation process to decode the plurality of outputs to obtain the inference data.

[0022] In another aspect, a device for executing any one of the above methods is provided. In another aspect, a memory is provided. The memory contains instructions, and when the instructions are executed by a processor, the processor executes any one of the above methods.

[0023] On the other hand, a device is provided. The device includes a memory storing instructions and a processor. By executing the instructions, the processor: obtains multiple inputs of a distributed reasoning process representing a machine learning process; encodes the multiple inputs to generate one or more redundant inputs, so that each of the one or more redundant inputs includes a cascade of data from corresponding at least two of the multiple inputs. Each of the multiple inputs and each of the one or more redundant inputs are used for the same component reasoning process of the distributed reasoning process. By executing the instructions, the processor is also: as part of the distributed reasoning process, for each of the multiple inputs and each of the one or more redundant inputs, sends the corresponding input to a corresponding processing device for executing the same component reasoning process.

[0024] On the other hand, by executing the instructions, the processor may also be caused to send the corresponding input to the corresponding processing device via a wireless communication link, thereby sending the corresponding input to the corresponding processing device.

[0025] On the other hand, by executing the instructions, the processor can also: for each redundant input of the one or more redundant inputs, adjust the size of the corresponding redundant input so that the corresponding redundant input has the same dimension as at least one input of the multiple inputs.

[0026] On the other hand, the processor may also adjust the size of the corresponding redundant input by at least one of the following: cropping, downsampling, interpolating, or padding the corresponding redundant input.

[0027] On the other hand, the multiple inputs may form an ordered data set. By executing the instructions, the processor may encode the multiple inputs by encoding the multiple inputs to generate one or more redundant inputs, so that for each redundant input in the one or more redundant inputs, the corresponding at least two inputs in the multiple inputs may be adjacent in the ordered data set.

[0028] In another aspect, the machine learning process may include a deep neural network.

[0029] In another aspect, a device is provided. The device includes a memory storing instructions and a processor. By executing the instructions, the processor obtains multiple outputs of a distributed reasoning process representing a machine learning regression process, and the multiple outputs include multiple results and multiple redundant results. By executing the instructions, the processor is also caused to perform one or more linear operations on at least two results to decode the multiple outputs to obtain reasoning data. Each of the multiple results is determined by the same component reasoning process of the distributed reasoning process based on a corresponding input in a plurality of inputs. Each of the one or more redundant results is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, and the corresponding redundant input includes a cascade of data from corresponding at least two inputs in the multiple inputs. The at least two results come from the multiple results and the one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results.

[0030] On the other hand, by executing the instructions, the processor may also be caused to perform the one or more linear operations to determine the missing output in the reasoning process.

[0031] On the other hand, the reasoning process may also represent a machine learning classification process. The multiple outputs may also include multiple labels and one or more redundant labels. By executing the instructions, the processor may also perform one or more set operations on at least two labels to decode the multiple outputs to obtain the reasoning data. The at least two labels may come from the multiple labels and the one or more redundant labels, and the at least two labels may include at least one redundant label of the one or more redundant labels.

[0032] On the other hand, a device is provided. The device includes a memory storing instructions and a processor. By executing the instructions, the processor obtains multiple outputs from a distributed reasoning process representing a machine learning classification process. The multiple outputs include multiple labels and one or more redundant labels. By executing the instructions, the processor is also caused to perform one or more set operations on at least two labels to decode the multiple outputs to obtain reasoning data. Each of the multiple labels is determined by the same component reasoning process of the distributed reasoning process based on a corresponding input in a plurality of inputs. Each of the one or more redundant labels is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, and the corresponding redundant input includes a cascade of data from corresponding at least two inputs in the multiple inputs. The at least two labels come from the multiple labels and the one or more redundant labels, and the at least two labels include at least one redundant label of the one or more redundant labels.

[0033] On the other hand, by executing the instructions, the processor may also be caused to perform the one or more collective operations to determine the missing outputs in the reasoning process.

[0034] On the other hand, the reasoning process may also represent a machine learning regression process. The multiple outputs may also include multiple results and one or more redundant results. By executing the instructions, the processor may also perform one or more linear operations on at least two results to decode the multiple outputs to obtain the reasoning data. The at least two results may come from the multiple results and the one or more redundant results, and the at least two results may include at least one redundant result of the one or more redundant results.

[0035] On the other hand, by executing the instructions, the processor can also perform the one or more set operations to obtain the inference data by performing a belief propagation process to decode the multiple outputs, thereby decoding the multiple outputs to obtain the inference data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more fully understand the embodiments of the present invention and its advantages, the following description is made by way of example with reference to the accompanying drawings, in which:

[0037] Figure 1 is a schematic diagram of a communication system in which embodiments of the present invention may be implemented;

[0038] Figure 2 is another schematic diagram of a communication system in which embodiments of the present invention may be implemented;

[0039] Figure 3 is a block diagram illustrating units or modules of a device in which embodiments of the present invention may be implemented;

[0040] Figure 4 is a block diagram illustrating units or modules of a device in which embodiments of the present invention may be implemented;

[0041] Figure 5 An exemplary manner of combining input images to form redundant input images according to an embodiment of the present invention is shown;

[0042] Figure 6 An example of a system for implementing an encoding reasoning process is shown;

[0043] Figure 7 is a flowchart of a method according to an embodiment of the present invention;

[0044] Figure 8 is a flowchart of a method according to an embodiment of the present invention;

[0045] Fig. 9 An example of an input image and a redundant image according to an embodiment of the present invention is shown;

[0046] Fig.10 An example of a distributed reasoning process according to an embodiment of the present invention is shown;

[0047] Fig.11 and Fig.12 The object detection rate of the distributed reasoning process executed according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0048] The operation and structure of the current exemplary embodiment are discussed in detail below. However, it should be understood that many applicable inventive concepts provided by the present invention can be embodied in a variety of specific situations. The specific embodiments discussed only illustrate the specific structure of the present invention and the specific way to use the present invention, but do not limit the scope of the present invention.

[0049] refer to Figure 1 , a simplified schematic diagram of a communication system is provided as an illustrative and non-limiting example. 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 (electric devices, ED) 110a to 120j (generally referred to as 110) may be interconnected with each other or connected to one or more network nodes (170a, 170b, generally 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.

[0050] Figure 2An exemplary communication system 100 is shown. In general, 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 delivery and mobility, etc.). The communication system 100 may provide high 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 produce a heterogeneous network including multiple layers. Compared with traditional communication networks, heterogeneous networks may achieve 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.

[0051] 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. The non-terrestrial communication network 120c includes an access node 120c, which can be generally referred to as a non-terrestrial transmit and receive point (NT-TRP) 172.

[0052] Any ED 110 may alternatively or additionally be used to connect, access, or communicate with any other T-TRP 170a and 170b, NT-TRP 172, Internet 150, core network 130, PSTN 140, other network 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.

[0053] 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 and / or non-orthogonal dimensions.

[0054] Air interface 190c enables communication between ED 110d and one or more NT-TRPs 172 via a wireless link or a simple link. In some examples, the link is a dedicated connection for unicast transmissions, a connection for broadcast transmissions, or a connection between a group of EDs and one or more NT-TRPs for multicast transmissions.

[0055] 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 use 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 or the EDs 110a, 110b, and 110c, or both, 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 radio links using different radio technologies and / or protocols. Instead of (or in addition to) wireless communications, EDs 110a, 110b, and 110c may communicate with a service provider or switch (not shown) and with the Internet 150 via a wired communication channel. PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). Internet 150 may include computer networks and / or subnets (intranets), and include protocols such as Internet Protocol (IP), Transmission Control Protocol (TCP), and User Datagram Protocol (UDP). EDs 110a, 110b, and 110c may be multi-mode devices capable of operating according to a variety of wireless access technologies, and include a plurality of transceivers required to support these technologies.

[0056] 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 wearable devices, smart transportation, smart city, drones, robots, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, etc.

[0057] Each ED 110 represents any suitable end-user equipment 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) device, personal digital assistant (PDA), smartphone, laptop, computer, tablet, wireless sensor, consumer electronic device, smart book, vehicle, car, truck, bus, train or IoT device, industrial equipment, or devices in the above devices (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, hereinafter referred to as T-TRP 170. Also in Figure 3 , the NT-TRP is hereinafter referred to as NT-TRP 172. Each ED 110 connected to T-TRP 170 and / or NT-TRP 172 may be dynamically or semi-statically started (i.e., established, activated, or enabled), shut down (i.e., released, deactivated, or disabled), and / or configured in response to one or more of connection availability and connection necessity.

[0058] ED 110 includes a transmitter 201 and a receiver 203 coupled to one or more antennas 204. Only one antenna 204 is shown. Alternatively, one, part or all of the antennas may be panels. Transmitter 201 and receiver 203 may be integrated into a transceiver or the like. The transceiver is used to modulate data or other content so that it can be sent through at least one antenna 204 or a network interface controller (NIC). The transceiver is also used to demodulate data or other content received through 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 by wireless or wired means. Each antenna 204 includes any suitable structure for sending and / or receiving wireless signals or wired signals.

[0059] 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.

[0060] ED 110 may also include one or more input / output devices (not shown) or interfaces (e.g., 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.

[0061] ED 110 also includes a processor 210 for performing operations related to preparing uplink transmissions to be sent to NT-TRP 172 and / or T-TRP 170, processing downlink transmissions received from NT-TRP 172 and / or T-TRP 170, and processing sidelink transmissions to and from another ED 110. Processing operations related to preparing uplink transmissions 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 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 using, for example, a reference signal received from the NT-TRP 172 and / or the T-TRP 170.

[0062] Although not shown, the processor 210 may be part of the transmitter 201 and / or part of the receiver 203. Although not shown, the memory 208 may be part of the processor 210.

[0063] 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, respectively, which are used to execute instructions stored in a 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 a programmed field-programmable gate array (FPGA), a graphical processing unit (GPU), or an application-specific integrated circuit (ASIC).

[0064] In some implementations, the T-TRP 170 may be represented by other names, 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 transmitting / receiving 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 above-mentioned device, or may refer to a device in the above-mentioned device (eg, a communication module, a modem, or a chip).

[0065] In some embodiments, various parts of T-TRP 170 can be distributed. For example, some modules in T-TRP 170 can be remote 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 a network-side module that performs the following processing operations: for example, determining the location of ED 110, resource allocation (scheduling), message generation, and encoding / decoding, which modules 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.

[0066] 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. Alternatively, one, some or all of the antennas may 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 related to preparing downlink transmissions sent to the ED 110, processing uplink transmissions received from the ED 110, preparing backhaul transmissions sent to the NT-TRP 172, and processing transmissions received from the NT-TRP 172 via the backhaul. Processing operations related to preparing to send downlink transmissions or backhaul transmissions may include operations such as encoding, modulation, precoding (e.g., MIMO precoding), transmit beamforming, and generating symbols for transmission. Processing operations related to processing received transmissions in the uplink or on the 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 the scheduler 253 may schedule for transmission. The processor 260 may perform other network-side processing operations described herein, such as determining the location of the ED 110, determining the location at which the NT-TRP 172 is deployed, etc. In some embodiments, the processor 260 may generate signaling to configure one or more parameters of the ED 110 and / or one or more parameters of the NT-TRP 172, etc. Any signaling generated by the processor 260 is sent by the transmitter 252. Note that "signaling" as 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), while static or semi-static higher-layer signaling may be included in a data packet sent in a data channel such as a physical downlink shared channel (PDSCH).

[0067] The scheduler 253 may be coupled to the processor 260. The scheduler 253 may be included within the T-TRP 170 or may operate separately from the T-TRP 170. The T-TRP 170 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 that are used to implement some or all of the functionality and / or embodiments described herein and are executed by the processor 260.

[0068] Although not shown, the processor 260 may be a part of the transmitter 252 and / or a part of the receiver 254. In addition, although not shown, the processor 260 may implement the scheduler 253. Although not shown, the memory 258 may be a part of the processor 260.

[0069] The processor 260, the scheduler 253, and the processing components of the transmitter 252 and the receiver 254 may be implemented by the same or different one or more processors, which are used to 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 a dedicated circuit such as an FPGA, a GPU, or an ASIC.

[0070] Although the NT-TRP 172 is shown as a drone by way of example only, the NT-TRP 172 may be implemented in any suitable non-ground form. In addition, in some implementations, the NT-TRP 172 may be known by other names, 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. Alternatively, one, part, or all of the antennas may 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 related to preparing downlink transmissions sent to the ED 110, processing uplink transmissions received from the ED 110, preparing backhaul transmissions sent to the T-TRP 170, and processing transmissions received from the T-TRP 170 via backhaul. Processing operations related to preparing to send a downlink transmission or a 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 received transmissions in the uplink or on the backhaul may include operations such as receive beamforming, demodulation, and decoding received symbols. In some embodiments, the processor 276 implements transmit beamforming and / or receive beamforming based on beam direction information (e.g., BAI) received from the T-TRP 170. In some embodiments, the processor 276 may generate signaling to configure one or more parameters of the ED 110, etc. In some embodiments, the NT-TRP 172 implements physical layer processing, but does not implement high-level functions, such as functions of the medium access control (MAC) layer or the radio link control (RLC) layer. Since this is just an example, more generally, in addition to physical layer processing, the NT-TRP 172 may also implement high-level functions.

[0071] NT-TRP 172 also includes a memory 278 for storing information and data. Although not shown, processor 276 may be part of transmitter 272 and / or part of receiver 274. Although not shown, memory 278 may be part of processor 276.

[0072] Processor 276 and the processing components of transmitter 272 and receiver 274 may be implemented by the same or different one or more processors, respectively, which are used to 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 may be implemented using a programmed FPGA, GPU, or ASIC or other dedicated circuit. In some embodiments, NT-TRP 172 may actually be multiple NT-TRPs that operate together to serve ED 110 through coordinated multi-point transmission or the like.

[0073] The T-TRP 170, NT-TRP 172, and / or ED 110 may include other components, but these components are omitted for clarity.

[0074] One or more steps of the exemplary methods provided herein may be performed by Figure 4 The corresponding unit or module provided is executed. Figure 4 Units or modules in devices such as ED 110, T-TRP 170, or NT-TRP 172 are shown. For example, a signal can be sent by a sending unit or a sending module. 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. The signal can be processed by a processing unit or a processing module. Other steps can be performed by an artificial intelligence (AI) module or a machine learning (ML) module. The corresponding units or modules can be implemented using hardware, one or more components or devices that execute software, or a combination thereof. For example, one or more units or modules can be programmed integrated circuits such as FPGAs, GPUs, or ASICs. It should be understood that if these modules are implemented using software for execution by a processor, etc., these modules can 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.

[0075] Other details about ED 110, T-TRP 170 and NT-TRP 172 are known to those skilled in the art. Therefore, these details are omitted here.

[0076] In cases where a centralized machine learning process, such as a centralized DNN, performs inference using data from a set of distributed devices, the reliability of the machine learning process can depend on the quality, reliability, and latency of the transmission between the machine learning process and the devices.

[0077] This can be illustrated by considering an example of deploying a machine learning process in a remote data center that is connected to a core network of a communication network. The communication network includes multiple devices that are connected to the core network through a TRP or base station in an access network. The machine learning process at the remote data center is used to receive an inference request from the core network, where the request includes data from multiple client devices.

[0078] However, interruptions in wireless connections, etc. between a client device and its corresponding TRP may result in packet loss or delays in receiving data from the client device at the remote data center. Even if the client device is stably connected to the network, the network path between the client device and the remote data center may be long and / or may include many hops, thereby increasing the risk of delays and packet loss. In addition, network congestion caused by the transmission of large data, such as image data or video data, may further reduce the reliability of transmission between the device, the core network, and the remote data center.

[0079] Therefore, the inference performance of machine learning processes may be affected by the quality, reliability, and latency of transmission between devices, core networks, and remote data centers.

[0080] These issues can be mitigated by performing inference locally on the data source. Thus, for example, machine learning processes can be deployed closer to the device. However, machine learning processes can be computationally intensive. For example, a DNN can have as many as 10 billion to 100 billion neurons. Therefore, deploying inference performed using machine learning processes on a single client device is challenging.

[0081] In contrast, by distributing the machine learning process across multiple devices, the machine learning process can be implemented using low-cost and low-power devices. Distributed reasoning can 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 process across multiple devices, the machine learning process can be implemented in or near the access network, thereby reducing the risk of input data loss or delay for the machine learning process.

[0082] 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, thereby recovering missing results.

[0083] To introduce this redundancy, some people have proposed methods of encoding reasoning. However, existing methods of introducing interference redundancy involve using redundant interference units and significantly increase the training and reasoning complexity of the redundant reasoning unit 606. In addition, some proposed methods involve using a parity model to generate redundant inputs, but learning a parity model is more difficult because features from more queries are packed into a single parity query.

[0084] Various aspects of the present invention provide a method that includes obtaining multiple inputs of a distributed reasoning process representing a machine learning process. Each of the multiple inputs is used for the same component reasoning process of the distributed machine learning process. The method also includes: encoding the multiple inputs to generate one or more redundant inputs, so that each of the one or more redundant inputs includes a cascade of data from corresponding at least two of the multiple inputs. Each of the one or more redundant inputs is used for the same component reasoning process of the distributed reasoning process. The method also includes: as part of the distributed reasoning process, for each of the multiple inputs and each of the one or more redundant inputs, sending the corresponding input to a corresponding processing device for executing the same component reasoning process.

[0085] Thus, according to various aspects of the present invention, redundant inputs for a distributed reasoning process may be generated by concatenating (e.g., connecting without mixing) data from two or more of the plurality of inputs. This process may be referred to as encoding, similar to encoding theory. Thus, the redundant inputs may be generated by or according to a code that indicates data used to form the redundant inputs and how the data is connected (e.g., in what order) to form the redundant inputs. Thus, the code may be applied to the plurality of inputs such that data from two or more of the plurality of inputs are concatenated to form the redundant inputs.

[0086] This method of generating redundant inputs is independent of the machine learning process, which means that the same component reasoning process can be implemented at each of the devices involved in the distributed reasoning process. Therefore, for example, the same component reasoning process trained using the same machine learning method and training data can be implemented at each of the processing devices in the corresponding processing devices, which is a significant practical advantage over traditional coded reasoning methods that use specially trained redundant reasoning units. Since the same reasoning process is deployed on each of the processing devices in the processing devices, the redundant inputs of the same process should produce redundant outputs that contain elements of the outputs of the component reasoning at other devices. This is likely to be true because each of these devices uses the same process.

[0087] This method can be achieved by Figure 5504, 506, and 508, which are used in a digit detection process developed using machine learning. The digit detection process is used to process images with distinguishable digits, such as inputs 502, 504, 506, and 508. Given input images 502, 504, 506, and 508, one way to combine the input images to form a redundant input image is to overlap or superimpose the images to form image 510. However, the digits in image 510 are indistinguishable, which means that the digit detection process may not be able to detect the digits in image 510. Therefore, the digit detection process requires further training to process image 510.

[0088] In contrast, according to the present invention, input images 502, 504, 506, 508 may be concatenated to form image 512, etc. By concatenating images 502, 504, 506, 508 together without blending (e.g., superimposing) the images, digits are still distinguishable in combined image 512, which means that a digit detection process may still detect digits in image 512. This means that the same digit detection process used to process images 502, 504, 506, 508 may be used to process image 512 without further modification.

[0089] Although this example involves image data, those skilled in the art will appreciate that, in general, the disclosure herein can be applied to any data for inference.

[0090] Figure 6 1 shows an example of a system 600 for implementing a distributed reasoning process according to various aspects of the present invention. The system 600 includes a first reasoning unit 602 and a second reasoning unit 604. As part of the distributed reasoning process, the first reasoning unit 602 is used to perform a first input X 1 Execute the component reasoning process to obtain the first result Y 1 =f(X 1 As part of the distributed reasoning process, the second reasoning unit 604 is used to perform the second input X 2 Perform the same component reasoning process to obtain the second result Y 2 =f(X 2 ).

[0091] The system 600 further includes a redundant reasoning unit 606, which is configured to receive a redundant input X obtained by encoding the first input and the second input. 3 =h(X 1 ,X 2 According to one aspect of the present invention, the first input X 1 and the second input X 2 Encode to generate the third input X3 , so that the third input X 3 Includes the first input X 1 and the second input X 2 Cascade of data.

[0092] By generating the third input X in this way 3 , the third input X 3 The component reasoning processes implemented at the first reasoning unit 602 and the second reasoning unit 604 are still identifiable. This means that the same component reasoning processes implemented at the first reasoning unit 602 and the second reasoning unit 604 can also be implemented at the redundant reasoning unit 606. For example, the off-the-shelf machine learning processes can be implemented on each of the first reasoning unit 602, the second reasoning unit 604, and the redundant reasoning unit 606 to perform distributed reasoning with redundancy. More generally, this means that redundancy can be introduced into distributed reasoning without the need for dedicated training of the redundant reasoning unit 606.

[0093] The redundant reasoning unit 606 is used to perform a component reasoning process on the redundant input to output a third result Y 3 =f(h(X 1 ,X 2 )), the third result can be used to restore one of the first result or the second result.

[0094] Thus, for example, if the first reasoning unit 602 fails to return the first result, then the second result Y 2 (from the second reasoning unit 604) and the third result Y 3 (from redundant input 606) to determine the first result Similarly, if the second reasoning unit 604 fails to return the second result, it can be based on the first result Y 1 (from the first reasoning unit 602) and the third result Y 3 (from redundant input 606) to determine the second result Estimates.

[0095] Therefore, various aspects of the present invention can introduce redundancy into the distributed reasoning process without requiring dedicated training of the redundant reasoning units.

[0096] Figure 7 A flow chart of a method 700 according to an embodiment of the present invention is shown.

[0097] The method 700 may be performed by an apparatus. In a specific example, the method 700 may be performed by an apparatus in a communication system, for example, Figures 1 to 4For example, the method 700 may be performed by a core network of the communication system (eg, the core network of the communication system described above). Figures 1 to 4 As another example, the method 700 may be performed by an access network of the communication system (e.g., the core network 130 described above). Figures 1 to 4 The method is performed by an apparatus in any one of the RANs 120a and 120b described above.

[0098] In step 702, a plurality of inputs to a distributed inference process representing a machine learning process (eg, an algorithm) are obtained.

[0099] The distributed reasoning process is distributed across multiple processing devices. The processing device may include any suitable device, for example, one or more communication electronic devices (for example, the above combined Figures 1 to 4 any of the communicating EDs 110a to 110j described above) and / or one or more network nodes (e.g., Figures 1 to 4 Thus, for example, the processing device may access the access network (e.g., any of the network nodes 170a, 170b, 172 described above). Figures 1 to 4 In certain embodiments, the processing device may be deployed in a core network of the communication system (e.g., the core network of the communication system described above). Figures 1 to 4 In the core network 130 described above.

[0100] In a specific example, the plurality of processing devices may include one or more Internet of Things (IoT) devices. For example, the processing device may include a device for performing machine-to-machine (M2M) communication.

[0101] The input may include any data about which reasoning may be performed. Thus, for example, the input may include one or more of image data, audio data, video data, measurement data, network data of a communication network (e.g., indicating traffic, usage, performance, or any other network parameter), user data, or any suitable data.

[0102] Multiple inputs may be included in a single data set. For example, step 702 may include obtaining (e.g., receiving) a single data set that includes multiple inputs. Step 702 may also include extracting the multiple inputs from the single data set (e.g., by segmenting or dividing the single data set). For another example, step 702 may include receiving multiple inputs from one or more devices.

[0103] Each of the multiple inputs is used in the same component reasoning process of the distributed reasoning 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 the distributed reasoning process. The component 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 support vector machine, or any other suitable machine learning process. The component reasoning process may include a regression process, a classification process (e.g., a classifier), or a combination of a regression process and a classification process, etc. It will be appreciated by those skilled in the art that a machine learning process is typically selected based on a specific reasoning task. For example, a reasoning task may include image classification, and a component process may include a neural network such as a deep neural network trained to classify images.

[0104] The same component reasoning process is executed at each of the processing devices. In this case, the same component reasoning process means that the component reasoning process implemented at each of the processing devices includes the same architecture and is trained in the same manner (e.g., using the same training data). Therefore, the same code can be implemented at each of the processing devices in the processing devices, which, when executed by the processing device, causes the corresponding processing device to execute the same component reasoning process. This can be equivalently referred to as a corresponding instance of the same component reasoning process deployed at each of the processing devices.

[0105] In step 704, the method includes encoding the plurality of inputs to generate one or more redundant inputs. The one or more redundant inputs are redundant because they include data that is also included in the plurality of inputs to input the component process. Thus, for example, the one or more redundant inputs may be used to recover missing outputs from the distributed reasoning process.

[0106] Process multiple inputs so that each of the one or more redundant inputs includes a concatenation of data from at least two of the multiple inputs. Since this process provides redundancy in a manner similar to coding theory, it can be called coding. In this case, concatenation can refer to connecting 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 three or more data sets, etc. can be tiled. Tiling is particularly suitable for data with two or more dimensions.

[0107] Those skilled in the art will appreciate that there are various codes (e.g., error correction codes) that can be used to generate one or more redundant inputs. The code indicates the inputs used to form the redundant inputs, and the manner in which these inputs are connected to form the redundant outputs. Generally, the code can be a systematic code so that each redundant input contains data from each of the corresponding two or more inputs. For example, the redundant inputs may only include data from each of the two or more inputs indicated by the code.

[0108] A code can be represented by its generator matrix, parity check matrix, or code graph. Thus, for example, the generator matrix of the (7,4) Hamming matrix for t=1 can be written as:

[0109]

[0110] Where each row represents one of the multiple inputs, and each column (or variable node) represents a processing device. The 1 in each column indicates the input used for the component reasoning process at the corresponding processing device. Therefore, in this example, four inputs are used to generate three redundant inputs, thereby generating a total of seven inputs for the distributed reasoning process. The corresponding parity check matrix can be written as:

[0111]

[0112] Where each column represents a processing device and each row represents a check node connected to the corresponding processing device. In this case, t is a measure of the error correction capability of the code. Thus, by generating redundant inputs according to a code with an error correction capability of t, up to t erroneous outputs can be corrected (e.g., if the processing device returns an incorrect result), or up to 2t erased outputs can be corrected (e.g., if the processing device fails to return a result).

[0113] For another example, multiple inputs can be encoded according to the (21,12) degree 2 / 3 code. The parity check matrix of the (21,12) degree 2 / 3 code can be expressed as:

[0114]

[0115] Since the code is a (21, 12) code, a set of 12 inputs can be used to generate nine redundant inputs, generating a total of 21 inputs. Since the code is a degree 2 / 3 code, each of the redundant inputs generated according to the code includes data from two or three inputs, and the number of inputs used to generate a particular redundant input is determined according to the degree distribution.

[0116] Thus, one or more redundant inputs may be generated according to any suitable code. In a specific example, a code may be determined (eg, selected) according to a set of one or more rules (eg, criteria or factors). Examples of rules are as follows.

[0117] In one example, a code may be determined such that each redundant input includes only data from multiple inputs that satisfy a maximum threshold (e.g., less than the maximum threshold). For example, this may be referred to as limiting the degree of each redundant input. Thus, a code may be selected based on, for example, the sparsity of the code. By limiting the degree of each redundant input, decoding complexity may be reduced while improving reasoning performance. For example, in image detection, sparse codes may improve detection rates.

[0118] In addition, the code can also be determined based on the order or adjacency of two or more inputs in the multiple inputs. This may be particularly applicable when the multiple inputs form an ordered data set (e.g., image, audio data, video data, time series or any other suitable ordered data set). In these examples, the multiple inputs can be encoded so that for each redundant input in one or more redundant inputs, the corresponding at least two inputs in the multiple inputs are adjacent in the ordered data set.

[0119] Thus, in an example where the input includes component images extracted from a composite image, the multiple inputs may be encoded such that the redundant input includes pixels from two adjacent component images in the composite image. For another example, when the input includes frames extracted from video data, the multiple inputs may be encoded such that the redundant input includes data from two subsequent frames in the video data. Maintaining the contiguity of data in the redundant inputs may increase the probability of performing meaningful reasoning on the redundant inputs.

[0120] The code can also be determined based on the degree distribution of the code. In this case, the degree represents the number of connections associated with an input (e.g., one of multiple inputs or redundant inputs) or a constraint between inputs. Thus, for example, the degree of a particular input among multiple inputs can indicate the number of redundant inputs that include data from the particular input. The degree distribution of a code represents the probability mass function of the degree of the code. The degree distribution of the code can be optimized to improve decoding performance and / or reduce decoding complexity. It will be appreciated by those skilled in the art that the degree distribution can be optimized in a variety of ways. For example, the optimal degree distribution can be identified by computer simulation. Then, the code can be selected based on the computer simulation.

[0121] It will be appreciated by those skilled in the art that determining a code according to one or more of the above rules may include determining a type of code. There are various types of codes suitable for implementation in method 700. By way of example only, suitable codes may include Hamming codes, low-density parity-check (LDPC) codes, polar codes (e.g., systematic polar codes), BCH (Bose–Chaudhuri–Hocquenghem, BCH) codes, Reed-Muller (RM) codes, Golay codes (e.g., binary Golay codes), or any other suitable code.

[0122] For example, the code may be determined to be an LDPC code based on the sparsity rules described above. LDPC codes are sparse, so LDPC codes may be selected to limit the degree of each redundant input. Therefore, the type of code may be selected based on one or more of the above rules.

[0123] In addition, the characteristics of the code can also be determined according to one or more of the above-mentioned rules. In one example, the generator matrix of a particular type of code can be determined based on the above-mentioned rules. For example, if it has been determined that the code is an LDPC code, the generator matrix of the LDPC code can be determined based on the desired order or adjacency of two or more inputs in a plurality of inputs.

[0124] Those skilled in the art will appreciate that the type of code and / or the characteristics of the code may also be determined based on one or more performance constraints of the distributed reasoning process. For example, it may be determined that the code satisfies one or more of the following constraints: complexity, latency, resource availability (e.g., the number of available processing devices), and memory occupancy of the distributed reasoning process. For another example, the code length may be determined based on one or more performance constraints. In general, the code length should not be too long (e.g., the length may be less than 50).

[0125] Although examples of codes are provided herein, those skilled in the art will appreciate that the present invention is not limited thereto and, in general, any suitable code may be used.

[0126] The method 700 also includes, at step 706, for each of the plurality of inputs and each of the one or more redundant inputs, sending a corresponding input to a corresponding processing device of the plurality of processing devices. Each corresponding input is sent to a corresponding processing device for executing the same component reasoning process as part of the distributed reasoning process.

[0127] Therefore, in the above (7,4) Hamming code example, step 706 may include sending 4 inputs and 3 redundant inputs to the corresponding processing devices for executing the same component reasoning process. Therefore, the component reasoning process may have seven instances, each running on a different processing device.

[0128] The corresponding input may be sent directly or indirectly to the processing device. Thus, for example, the corresponding input may be sent to the processing device via one or more intermediate devices. One or more corresponding inputs may be sent to the corresponding processing device via a wireless link. For example, method 700 may be performed in a communication system, for example, in conjunction with Figures 1 to 4 In the communication system 100 described above, multiple processing devices may be connected to the communication system via corresponding wireless links. Due to the lower reliability and higher latency of wireless links, distributed reasoning may be particularly susceptible to data loss when the processing devices are connected via wireless links. Therefore, the redundancy provided by method 700 may be particularly advantageous when the processing devices are connected via wireless links.

[0129] Method 700 may also include adjusting the size of the redundant inputs so that each redundant input has a dimension that is the same as or comparable to the dimension of at least one of the multiple inputs. Since the redundant inputs are formed by cascading data from two or more inputs, the redundant inputs may have dimensions that are different from the dimensions of the inputs. For example, a redundant input formed by tiling four images having N×N pixels may have dimensions of 2N×2N. This may make it difficult for the component reasoning process to process the redundant inputs because the component reasoning process may be used to perform reasoning on inputs having a specific dimension. Therefore, the size of the redundant inputs may be adjusted so that the processing device can process more easily. In a specific example, the size of the redundant inputs may be adjusted to have the same dimension as at least one of the multiple inputs. For example, each of the redundant inputs and each of the multiple inputs may have the same dimension.

[0130] Those skilled in the art will appreciate that various techniques may be used to adjust the size of the redundant input. For example, resizing may include reducing one or more dimensions of the redundant input by cropping (e.g., pruning) or downsampling (e.g., reducing the size) the redundant input. Reducing the size of the redundant input may reduce the memory requirements of the processing device. In addition, resizing may also include interpolating and / or padding the redundant input. For example, the redundant input formed by tiling three N×N data sets may be padded to a 2N×2N dimension and then reduced to an N×N dimension.

[0131] The resizing may be performed on the same device that performs the encoding. The resizing may also be performed elsewhere. For example, the redundant inputs may be sent to respective processing devices, and each respective processing device may resize the received redundant inputs according to any of the above techniques.

[0132] It will be appreciated by the skilled person that reasoning is typically performed on large data sets. In some embodiments, the multiple inputs mentioned in the above method 700 may include a subset of the data set on which reasoning is to be performed. The data set on which reasoning is to be performed may be divided into multiple groups (e.g., multiple batches), wherein each group includes a subset of data from the data set. Each group of data may be encoded separately so that data from two groups are not encoded together. For example, the input images for an encoded object detection task may be grouped into multiple batches, wherein each batch contains two or more images. Each batch may include K input images, and N–K redundant images are generated according to the (N, K) encoding. Therefore, images from two different batches are not encoded together.

[0133] Therefore, in some embodiments, the data to be reasoned can be divided into groups containing corresponding multiple inputs. Method 700 can be performed for each corresponding multiple input. Therefore, the data from each group can be encoded separately, so that data from two different groups are not encoded together. This can simplify decoding the output from the distributed reasoning process.

[0134] Therefore, an embodiment of the present invention provides a method for encoding the input of a distributed reasoning process to generate one or more redundant inputs. By generating one or more redundant inputs according to the above method 700, the redundant inputs can be independent of the machine learning process, which means that the same component reasoning process can be implemented at each processing device in the processing devices involved in the distributed reasoning process. This means that redundancy can be provided for the distributed reasoning process without the need for special training for processing redundant inputs.

[0135] Figure 8 A flowchart of a method 800 for decoding a plurality of outputs from a distributed reasoning process representing a machine learning process is shown. The distributed reasoning process may be a combination of the above Figure 7 The method 800 may be performed by the same device that performs the method 700. Thus, for example, encoding and decoding may be performed by the same device. However, those skilled in the art will appreciate that this need not always be the case, and in general, the method 800 may be performed by any suitable device.

[0136] The machine learning process may include a regression process. In this case, the machine learning regression process may be a process that attempts to obtain a quantity (e.g., a continuous result rather than a discrete result) as an inference result, wherein the process is developed using machine learning. Alternatively, the machine learning process may include a classification process (e.g., the machine learning process may include a classifier). The machine learning classification process may be any process developed using machine learning that is intended to classify or label data. Those skilled in the art will appreciate that in certain examples, the machine learning process may include both regression and classification.

[0137] In step 802, method 800 includes obtaining a plurality of outputs. The plurality of outputs includes a plurality of results and a plurality of redundant results, wherein each of the plurality of results is determined by a same component reasoning process of the distributed reasoning process based on a corresponding input of the plurality of inputs. For example, the same component reasoning process may be the same component reasoning described above in conjunction with method 700.

[0138] Each of the one or more redundant outputs is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, the corresponding redundant input comprising a concatenation of data from corresponding at least two of the plurality of inputs. The redundant input may be based on the above in combination with Figure 7 The method 700 described is generated.

[0139] The multiple outputs may be obtained by receiving, for each of the multiple outputs, a corresponding output from a corresponding device executing the same component reasoning process. Thus, for example, step 802 may include collating the multiple outputs from multiple devices. Alternatively, for example, the multiple outputs may be received from a single device that collates the multiple outputs. In general, the multiple outputs may be received from one or more devices.

[0140] In examples where the machine learning process includes a classification process, the multiple results may include multiple labels, and the one or more redundant results may include one or more redundant labels. Thus, for example, each processing device may output a corresponding label from a component reasoning process. Each label may indicate a class or category of a corresponding input. Thus, for example, a label may indicate a class or category of a feature of a corresponding input. For example, an image classification process may output a label indicating that one image contains a monkey and another label indicating that another image contains a bear. Thus, labels may include classes, categories, class labels, or any other suitable way of classifying information, etc.

[0141] Alternatively, in an example where the machine learning process includes regression and classification, the multiple outputs may include multiple labels and one or more redundant labels in addition to the multiple results and one or more redundant results. Thus, for example, each processing device may output a corresponding label and a corresponding result from a component reasoning process.

[0142] In step 804, the plurality of outputs are decoded to obtain inference data. Decoding may include performing one or more linear operations and / or one or more set operations.

[0143] Thus, for example, step 804 may include performing one or more linear operations on at least two results to decode multiple outputs to obtain inference data. This may be particularly applicable in examples where the machine learning process includes a regression process. The at least two results come from multiple results and one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results. Therefore, the at least two results include at least one redundant result of the one or more redundant results, but may also include one or more results of the multiple results. In other words, the at least two results include at least one redundant result of the one or more redundant results, and zero or more results of the multiple results. Therefore, for example, decoding can be performed based on two or more redundant results. Alternatively, at least one redundant result of the redundant results or one or more results of the multiple results can be decoded.

[0144] The acquired reasoning data may include an estimate of missing results from an instance of the same component reasoning process (e.g., a result that should have been returned by the device but failed to be returned), etc. Since multiple results and redundant results are generated by the same component reasoning process, the missing results can be recovered by linear operations. Even if no data is lost in the distributed reasoning process, it is still advantageous to decode the results and redundant results by linear operations.

[0145] This can be done by referring to Fig. 9 Describe the example to illustrate. Fig. 9 An image 900 is shown being processed by a distributed inference process. As part of the distributed inference process, two input images 902, 904 are extracted from the image 900. A redundant input 906 is generated that includes data from the two input images 902, 904 and is populated with additional data from the image 900. The location and size of a person shown in the image 900 is represented by a bounding box 908 that is characterized by four quantities (x, y, w, h), where x and y are the coordinates of the center of the rectangle and w and h are the width and height of the rectangle.

[0146] Since a person is detected in both the second input image 904 and the redundant input image 906, the results of the component inference process performed on these images 904, 906 can be combined to obtain a more accurate result. This can be achieved by considering the following: the position (x) of the input image 904 in the image 900 i ,y i 、w i、h i ) and the position (x) of the redundant input image 906 in the image 900 r ,y r 、w r 、h r ), and the position of the person in the input image 904 returned by the component inference process (x pi ,y pi 、w pi 、h pi ) and the position (x pr ,y pr 、w pr 、h pr ). The position and size of the person in the image 900 (e.g., the position and size of the bounding box 908) can be determined according to the following:

[0147] x p =weighted_mean(x i +w i x pi ,x r +w r x pr )

[0148] y p =weighted_mean(y i +h i y pi ,y r +h r y pr )

[0149] w p =weighted_mean(w i w pi ,w r w pr )

[0150] h p =weighted_mean(h i h pi ,h r h pr )

[0151] The weights of the weighted mean (weighted_mean) can be based on the importance or confidence of each result, etc. Each result in the plurality of results and each redundant result in the one or more redundant results can be associated with a corresponding confidence indicator (e.g., a confidence score) indicating the likelihood of the presence of an entity or object in the bounding box. The weights of the weighted mean can be based on the confidence indicators associated with the results. Thus, for example, a heavier (or larger) weight can be assigned to a result associated with a stronger confidence indicator (e.g., a higher likelihood of the presence of an entity or object).

[0152] Thus, the results returned by executing the component reasoning processes on the input image 904 and the redundant input image 906 can be decoded by linear operations to obtain a more accurate estimate of the position and size of the bounding box. Thus, even when the distributed reasoning process returns results (e.g., no missing results) from each instance of the component reasoning process, the redundancy provided by the methods described herein can improve the accuracy of reasoning. Fig. 9 The example discussed involves image bounding box detection, and it should be understood that Figure 8 The described decoding method 800 is not limited thereto. In general, the decoding method 800 can be used to decode the output of any suitable distributed reasoning process.

[0153] Returning to method 800, one skilled in the art will appreciate that the one or more linear operations that may be used to decode the plurality of outputs may vary depending on the outputs, the distributed reasoning process, and / or the reasoning sought, etc. In this case, the linear operation is any operation that preserves vector addition and scalar multiplication operations. Thus, the one or more linear operations may include any operation f(.) that satisfies the following formula:

[0154] f(x+y)=f(x)+f(y);

[0155] f(ax)=af(x)

[0156] x and y can take any value, and a can take any constant.

[0157] As described above, step 804 may include performing one or more collective operations in addition to, or in lieu of, one or more linear operations.

[0158] Therefore, in some examples, one or more set operations may be performed in step 804 to decode the plurality of outputs to obtain inference data. In examples where the machine learning process includes a classification process such that the plurality of outputs include a plurality of labels and one or more redundant labels, performing one or more set operations is particularly appropriate.

[0159] Therefore, step 804 may include performing one or more set operations on at least two tags to decode multiple outputs to obtain inference data. The at least two tags are from multiple tags and one or more redundant tags, and the at least two tags include at least one redundant tag from the one or more redundant tags. Therefore, the at least two tags include at least one redundant tag from the one or more redundant tags, but may also include one or more tags from the multiple tags. In other words, the at least two tags include at least one redundant tag from the one or more redundant tags and zero or more tags from the multiple tags. For example, decoding can be performed based on two or more redundant tags. Alternatively, at least one redundant tag from the redundant tags or one or more tags from the multiple tags can be decoded.

[0160] Decoding can be used to obtain missing outputs (e.g., missing labels) from a distributed reasoning process. Additionally or alternatively, decoding can also be used to improve the accuracy of the reasoning process.

[0161] The multiple outputs can be decoded using set operations in a variety of ways. In some examples, the multiple outputs can be decoded using a belief propagation process (e.g., an algorithm). This can be illustrated by considering the following example, where the redundant labels form a set R, the multiple labels form a set S, and N is a neighboring set. Multiple outputs i can be decoded to obtain multiple inference labels j by performing one or more of the following steps:

[0162] {class j→i}={class i' ,i'∈N(j)∩R}-union({class i”}for all i”∈N(j)∪Sand i”≠i)

[0163] {class i→j}={class i}∪union({class j'→i}for all j'∈N(i)and j'≠j)

[0164] {class i}={class i}∪union({class j'→i}for all j'∈N(i))

[0165] Where ∪ represents the union of two classes, "union" represents the union of more than two classes, "–" represents set difference, and ∩ represents set intersection. The neighbor set N of a particular inference label j may include every label in the labels for inference label j. This particular belief propagation process may reduce the complexity of decoding because multiple inference labels may be determined without performing an exhaustive search. The belief propagation process is particularly suitable for encoding using sparse codes because it converges faster for sparse codes.

[0166] Although this example of the belief propagation algorithm uses union, intersection, and difference set operations, in general, any suitable set operation may be used to decode the multiple outputs. Thus, for example, the one or more set operations may include one or more of: union, intersection, complement, and difference.

[0167] Thus, the present invention provides a method for decoding multiple outputs from a distributed inference process.

[0168] As described above, the distributed reasoning process can represent a regression process and a classification process. In some examples, the distributed reasoning process can include using a regression process to identify an aspect in an image and using a classification process to classify the identified aspect. For example, the distributed reasoning process can be used to return the coordinates of an object in an image and identify the object as a person.

[0169] In embodiments where the distributed reasoning process includes regression and classification, redundancy can be provided for regression and classification. Thus, as described above, the multiple outputs can include multiple results, one or more redundant results, multiple labels, and one or more redundant labels. According to various aspects of the present invention, the multiple outputs can be jointly decoded to obtain reasoning data.

[0170] This can be illustrated by considering the example of a distributed inference process that involves performing object detection on an image. In this example, the encoded inference process for each corresponding input may return the following results:

[0171] object i = {bounding box i ,kind i}

[0172] Among them, the bounding box i Indicates the objects detected in the corresponding image i The position and dimensions of the bounding box of class i Indicated as an object iThus, for example, each instance of an encoded inference process may return the output [x, y, w, h, p], where [x, y] is the position of the bounding box, [w, h] are the width and height of the bounding box, and p is a probability (e.g., likelihood) indicating that the detected object belongs to one or more of each corresponding class.

[0173] According to the present invention, the output [x, y, w, h, p] from each of the component inference processes can be jointly decoded. Thus, for example, information obtained when decoding the regression result can be fed into the classification result, and vice versa. Thus, at least one of the multiple results and one of the redundant results can be decoded based on information obtained during decoding of at least one of the multiple labels and one of the redundant labels. Similarly, at least one of the multiple labels and one of the redundant labels can be decoded based on information obtained during decoding of at least one of the multiple results and one of the redundant results. By jointly performing the bounding box recovery and class recovery processes, missing inference data can be recovered.

[0174] Through a belief propagation process, such as the one described above, an initial estimate of the class of a detected object can be obtained. Alternatively, a detected object can be assigned the class associated with the highest probability (e.g., the highest p), or any class with a probability exceeding a threshold.

[0175] It can then be determined whether the object detected in the image and the redundant image is the same object. For example, this can be determined by the intersection of the union. According to this method, if the class assigned to the first object A is the same as the class assigned to the second object B, and the intersection of the union of A and B is greater than a threshold, it can be determined that A and B are the same object. If neither of these two conditions is met, it can be determined that A and B are two independent objects. This process can be used to remove duplicate detected objects.

[0176] To recover the bounding box, one or more linear operations may be performed on the coordinates and size of the bounding box determined based on the input image and the one or more redundant images. Thus, for example, the coordinates [x, y] and size [w, h] of the bounding box may be determined according to the following formula:

[0177]

[0178] Among them, x Δ ,y Δis the position offset of the detected object between the input image and the redundant image, [x',y'] and [w',h'] are the coordinates and size of the bounding box in the redundant image, and α is the scaling factor of the object size between the input image and the redundant image. Thus, for example, α can indicate how much the redundant image is scaled down for input to the component inference process.

[0179] By jointly performing bounding box recovery and class recovery processes, missing inference data can be recovered and the accuracy of the distributed inference process can be improved.

[0180] Fig.10 An example of a distributed reasoning process according to an embodiment of the present invention is shown. In this example, the distributed reasoning process is used to detect numbers in input images. Each of the input images has the same dimensions (N, M, P), where N is the width of the image, M is the height of the image, and P is the number of bits per pixel.

[0181] The inputs 1002a, 1002b, 1002c, 1002d (collectively referred to as 1002) of the distributed reasoning process are encoded to obtain three redundant inputs 1004a, 1004b, 1004c (collectively referred to as 10004). Figure 7 The step 704 described above is performed to encode. These inputs are encoded according to a code having a parity check matrix:

[0182]

[0183] Thus, the first redundant input 1004a includes the first input 1002a, the second input 1002b, and the third input 1002c. The second redundant input 1004b includes the first input 1002a, the second input 1002b, and the fourth input 1002d. The third redundant input 1004c includes the first input 1002a, the third input 1002c, and the fourth input 1002d.

[0184] Each of redundant inputs 1004 is generated by tiling a corresponding input from multiple inputs 1004. The tiled inputs are padded and scaled so that each of redundant inputs 1004 has the same dimension as input 1002. Thus, padding bits are added to the tiled inputs to form redundant inputs 1004 having the same shape as input 1002. Redundant inputs 1004 are then downsampled to have the same size as input 1002 (e.g., having dimensions (N, M, P)).

[0185] Each of inputs 1002 is sent to a corresponding one of inference units 1006a, 1006b, 1006c, 1006d (collectively 1006) for component inference processing. Thus, for example, first input 1002a may be sent to first inference unit 1006a.

[0186] Each of the redundant inputs 1004 is sent to a corresponding one of the redundant reasoning units 1008a, 1008b, 1008c (collectively 1008) for component reasoning processing. Thus, for example, the first redundant input 1004a may be sent to the first redundant reasoning unit 1008a.

[0187] As part of the distributed reasoning process, the reasoning unit 1006 and the redundant reasoning unit 1008 each execute the same component reasoning process. Thus, for example, the reasoning unit 1006 and the redundant reasoning unit 1008 each may execute the same code to detect numbers in the respective inputs of these units. For example, the reasoning unit 1006 and the redundant reasoning unit 1008 may be combined with the above Figure 7 The processing device operates in the same manner as described.

[0188] The reasoning unit 1006 and the redundant reasoning unit 1008 are used to provide the corresponding outputs from the component reasoning processes. In this example, the corresponding output includes labels indicating the digits detected in the input image 1002 and the redundant image 1004. Therefore, the first reasoning unit 1006a, the second reasoning unit 1006b, the third reasoning unit 1006c, and the fourth reasoning unit 1006d are used to perform component reasoning processes using the corresponding first input 1002a, the second input 1002b, the third input 1002c, and the fourth input 1002d to obtain the corresponding labels The first redundant reasoning unit 1008a, the second redundant reasoning unit 1008b and the third redundant reasoning unit 1008c are used to perform component reasoning processes using the corresponding first redundant input 1004a, the second redundant input 1004b and the third redundant input 1004c to obtain corresponding labels

[0189] However, as shown in the figure, the second reasoning unit 1006b and the fourth reasoning unit 1006b fail to return the corresponding labels of the two reasoning units. There are many reasons for this situation. For example, when one or both of the second reasoning unit 1006b and the fourth reasoning unit 1006d attempt to send their corresponding results For another example, when executing the component reasoning process, a calculation error may have occurred at one or both of the second reasoning unit 1006b and the fourth reasoning unit 1006d.

[0190] Recover missing labels by decoding labels received from reasoning unit 1006 and redundant reasoning unit 1008 Since the second input 1002 (e.g., the input to the second reasoning unit 1006b) is included in the first redundant input 1004a and the second redundant input 1004b, based on the labels from the first redundant reasoning unit 1008a and the second redundant reasoning unit 1008b Determine the second label The second label The estimate of can also be determined based on at least one of the following: the label from the third redundant reasoning unit 1008c Tags from the first reasoning unit 1006a and the third reasoning unit 1006c

[0191] Since the fourth input 1004 (eg, the input to the fourth reasoning unit 1006d) is included in the second redundant input 1004b and the third redundant input 1004c, based on the labels from the first redundant reasoning unit 1008b and the second redundant reasoning unit 1008c Confirm the fourth label Estimation of. Fourth label The estimate of can also be determined based on at least one of the following: the label from the first redundant reasoning unit 1008a Tags from the first reasoning unit 1006a and the third reasoning unit 1006c

[0192] Therefore, even if the second reasoning unit 1006a and the fourth reasoning unit 1006c fail to return the corresponding labels of the two reasoning units It is also possible to recover these missing labels from the labels returned by the distributed inference process. Figure 8 Therefore, one or more set operations can be used to recover the missing tags. For example, the belief propagation process described above in conjunction with step 804 of method 800 may be used to recover the missing labels.

[0193] Although described in the case of digital recognition Fig.10 The examples shown are not specific, but those skilled in the art will appreciate that more generally, the various aspects described are applicable to reasoning processes. Fig.10Various aspects of the described examples can be applied to, among other things, distributed inference processes representing machine learning classification processes, or even more generally, to distributed inference processes representing machine learning processes.

[0194] Fig.11 Detection rate of an object detection process performed on an image according to an embodiment of the present invention is shown. Inference is performed on images from the COCO-val2017 dataset (Lin, Tsung-Yi et al., "Microsoft coco: Common objects in context". Springer, Kamm, 2014) to detect 36,781 labeled objects. The detection rate refers to the proportion of objects in an image that are correctly detected and labeled by the inference process.

[0195] exist Fig.11 In each example shown, a YOLOv3 model (Farhadi, Ali, and Joseph Redmon. "Yolov3: An incremental improvement". 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". Springer, Kamm, 2014).

[0196] The lower dashed line shows the detection rate of the inference process performed without any redundancy (e.g., without the encoding described herein). The upper solid line shows the detection rate of object detection performed by the distributed inference process according to an embodiment of the present invention (e.g., where encoding is performed according to method 700 and decoding is performed according to the joint decoding process described above). In this example, the images in the COCO-train2017 dataset are grouped into batches of four images, and each batch generates three redundant images using a (7,4) Hamming code with a parity check matrix:

[0197]

[0198] Therefore, the (7,4) Hamming code is used to generate 7 inputs including 4 input images and 3 redundant input images. The generator matrix of this Hamming code can be expressed as:

[0199]

[0200] The input image and redundant input image are respectively fed into corresponding instances of the trained YOLOv3 model, which outputs bounding box estimates and class predictions for one or more objects detected in the image. These outputs are decoded according to the joint decoding process and belief propagation process described above.

[0201] This process is repeated for erasure probabilities ranging from 0 to 0.8, where the erasure probability indicates the likelihood that each instance of the YOLOv3 model returns its corresponding output. So, for example, an erasure probability of 0 means that all outputs of the YOLOv3 model are returned. Fig.11 As shown, under all tested erasure probabilities (including zero), the distributed reasoning process implementing various aspects of the present invention provides a high object detection rate. Therefore, various aspects of the present invention provide an improved object detection rate during image detection.

[0202] Fig.12 The object detection rate of the reasoning process performed without any redundancy (e.g., without the encoding described herein) is shown compared to the object detection rate of the reasoning process implemented using different codes described herein. The lower dashed line (with circular marks) shows the object detection rate of the reasoning process performed without any redundancy. Fig.11 The object detection rate for the described implementation using the (7,4) Hamming code is represented by the middle solid line with asterisk markers. The upper solid line with square markers shows the object detection rate for the implementation using the (24,12) degree-2 code. 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 in each batch are input to an instance of YOLOv3. Since the (24,12) code is a degree-2 code, each redundant image contains data from two images. The parity check matrix for the (24,12) degree-2 code can be expressed as:

[0203]

[0204] like Fig.12 As shown, encoding and decoding each of the distributed reasoning processes according to various aspects of the present invention provides a detection rate that exceeds the detection rate of the reasoning process without redundancy. In addition, Fig.12 shows that the highest detection rate is provided by the degree 2 code, which indicates that it can be combined with the above Figure 2 Describes one or more rules for selecting code to optimize performance.

[0205] It should be understood that one or more steps in 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 (FPGAs) or application specific integrated circuits (ASICs). 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.

[0206] Although combinations of features are shown in the illustrated embodiments, it is not necessary to combine all 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 features shown in any one of the drawings or in all parts schematically shown in the drawings. In addition, selected features of an exemplary embodiment can be combined with selected features of other exemplary embodiments.

[0207] Although the present invention has been described with reference to illustrative embodiments, this description is not to be interpreted in a limiting sense. Those skilled in the art will appreciate various modifications and combinations of the illustrative embodiments and other embodiments of the present invention with reference to this description. Therefore, the appended claims are intended to cover any such modifications or embodiments.

Claims

1. A method comprising: Obtaining a plurality of inputs representing a distributed inference process of a machine learning process, wherein each of the plurality of inputs is for a same component inference process of the distributed inference process; encoding the plurality of inputs to generate one or more redundant inputs, such that each of the one or more redundant inputs comprises a concatenation of data from a corresponding at least two of the plurality of inputs, wherein each of the one or more redundant inputs is used for the same component inference process of the distributed inference process; As part of the distributed reasoning process, for each of the plurality of inputs and each of the one or more redundant inputs, the corresponding input is sent to a corresponding processing device for performing the same component reasoning process.

2. The method according to claim 1, characterized in that Sending the respective input to the respective processing device includes sending the respective input to the respective processing device over a wireless communication link.

3. The method according to any one of claims 1 and 2, further comprising: For each redundant input of the one or more redundant inputs, the corresponding redundant input is resized such that the corresponding redundant input has the same dimension as at least one input of the plurality of inputs.

4. The method according to claim 3, characterized in that The resizing includes at least one of: cropping, downsampling, interpolation, or padding.

5. The method according to any one of claims 1 to 4, characterized in that The plurality of inputs form an ordered data set, and for each redundant input of the one or more redundant inputs, the respective at least two inputs of the plurality of inputs are adjacent in the ordered data set.

6. The method according to any one of claims 1 to 5, characterized in that The machine learning process includes a deep neural network.

7. A method comprising: Obtaining a plurality of outputs representing a distributed inference process of a machine learning regression process, the plurality of outputs comprising a plurality of results and a plurality of redundant results, wherein: Each of the plurality of results is determined by a same component reasoning process of the distributed reasoning process based on a corresponding input of the plurality of inputs; Each of the one or more redundant outputs is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, the corresponding redundant input comprising a concatenation of data from a corresponding at least two of the plurality of inputs; One or more linear operations are performed on at least two results to decode the plurality of outputs to obtain inference data, wherein the at least two results are from the plurality of results and the one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results.

8. The method according to claim 7, characterized in that Performing the one or more linear operations to decode the plurality of outputs to obtain inference data includes performing the one or more linear operations to determine missing outputs in the inference process.

9. The method according to claim 7 or 8, characterized in that: The reasoning process further represents a machine learning classification process, the plurality of outputs further include a plurality of labels and one or more redundant labels, and the method further includes: One or more set operations are performed on at least two tags to decode the multiple outputs to obtain the inference data, wherein the at least two tags are from the multiple tags and the one or more redundant tags, and the at least two tags include at least one redundant tag among the one or more redundant tags.

10. A method comprising: Obtaining a plurality of outputs from a distributed inference process representing a machine learning classification process, wherein the plurality of outputs comprises a plurality of labels and one or more redundant labels, wherein: Each of the plurality of results is determined by a same component reasoning process of the distributed reasoning process based on a corresponding input of the plurality of inputs; Each of the one or more redundant outputs is determined by the same component reasoning process of the distributed reasoning process based on a corresponding redundant input, the corresponding redundant input comprising a concatenation of data from a corresponding at least two of the plurality of inputs; One or more set operations are performed on at least two tags to decode the multiple outputs to obtain inference data, wherein the at least two tags are from the multiple tags and the one or more redundant tags, and the at least two tags include at least one redundant tag among the one or more redundant tags.

11. The method according to claim 10, characterized in that Performing the one or more set operations to decode the plurality of outputs to obtain inference data includes performing the one or more set operations to determine missing outputs in the inference process.

12. The method according to claim 10 or 11, characterized in that: The reasoning process further represents a machine learning regression process, the plurality of outputs further comprises a plurality of results and one or more redundant results, and the method further comprises: One or more linear operations are performed on at least two results to decode the multiple outputs to obtain the inference data, wherein the at least two results are from the multiple results and the one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results.

13. The method according to any one of claims 10 to 12, characterized in that Performing the one or more set operations to decode the plurality of outputs to obtain the inference data includes performing a belief propagation process to decode the plurality of outputs to obtain the inference data.

14. An apparatus comprising: a memory for storing instructions; The processor executes the instructions to cause the processor to perform the following operations: Obtaining a plurality of inputs representing a distributed inference process of a machine learning process, wherein each of the plurality of inputs is for a same component inference process of the distributed inference process; encoding the plurality of inputs to generate one or more redundant inputs, such that each of the one or more redundant inputs comprises a concatenation of data from a corresponding at least two of the plurality of inputs, wherein each of the one or more redundant inputs is used for the same component inference process of the distributed inference process; As part of the distributed reasoning process, for each of the plurality of inputs and each of the one or more redundant inputs, the corresponding input is sent to a corresponding processing device for performing the same component reasoning process.

15. The device according to claim 14, characterized in that Execution of the instructions also causes the processor to send the corresponding input to the corresponding processing device via a wireless communication link, thereby sending the corresponding input to the corresponding processing device.

16. The device according to any one of claims 14 and 15, characterized in that Execution of the instructions also causes the processor to adjust the size of each redundant input of the one or more redundant inputs so that the corresponding redundant input has the same dimension as at least one input of the plurality of inputs.

17. The device according to claim 16, characterized in that The processor is also caused to adjust the size of the corresponding redundant input by at least one of: cropping, downsampling, interpolating, or padding the corresponding redundant input.

18. The device according to any one of claims 14 to 17, characterized in that The multiple inputs form an ordered data set; by executing the instructions, the processor encodes the multiple inputs by encoding the multiple inputs to generate one or more redundant inputs, so that for each redundant input of the one or more redundant inputs, the corresponding at least two inputs of the multiple inputs are adjacent in the ordered data set.

19. The device according to any one of claims 14 to 18, characterized in that The machine learning process includes a deep neural network.

20. An apparatus comprising: a memory for storing instructions; The processor executes the instructions to cause the processor to perform the following operations: Obtaining a plurality of outputs representing a distributed inference process of a machine learning regression process, the plurality of outputs comprising a plurality of results and a plurality of redundant results, wherein: Each of the plurality of results is determined by a same component reasoning process of the distributed reasoning process based on a corresponding input of the plurality of inputs; Each of the one or more redundant outputs is determined by the same component inference process of the distributed inference process based on a corresponding redundant input, wherein the corresponding redundant input includes a cascade of data from corresponding at least two of the multiple inputs; one or more linear operations are performed on at least two results to decode the multiple outputs to obtain inference data, wherein the at least two results are from the multiple results and the one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results.

21. The device according to claim 20, characterized in that Execution of the instructions also causes the processor to perform the one or more linear operations to determine missing outputs in the inference process.

22. The device according to claim 20 or 21, characterized in that The inference process also represents a machine learning classification process, and the multiple outputs further include multiple labels and one or more redundant labels; by executing the instructions, the processor is also caused to perform the following operations: One or more set operations are performed on at least two tags to decode the multiple outputs to obtain the inference data, wherein the at least two tags are from the multiple tags and the one or more redundant tags, and the at least two tags include at least one redundant tag among the one or more redundant tags.

23. An apparatus comprising: a memory for storing instructions; The processor executes the instructions to cause the processor to perform the following operations: Obtaining a plurality of outputs from a distributed inference process representing a machine learning classification process, wherein the plurality of outputs comprises a plurality of labels and one or more redundant labels, wherein: Each of the plurality of results is determined by a same component reasoning process of the distributed reasoning process based on a corresponding input of the plurality of inputs; Each of the one or more redundant outputs is determined by the same component inference process of the distributed inference process based on a corresponding redundant input, wherein the corresponding redundant input includes a cascade of data from corresponding at least two of the multiple inputs; one or more set operations are performed on at least two tags to decode the multiple outputs to obtain inference data, wherein the at least two tags are from the multiple tags and the one or more redundant tags, and the at least two tags include at least one redundant tag among the one or more redundant tags.

24. The device according to claim 23, characterized in that Execution of the instructions also causes the processor to perform the one or more collective operations to determine missing outputs in the reasoning process.

25. The device according to claims 23 and 24, characterized in that The reasoning process also represents a machine learning regression process, and the multiple outputs also include multiple results and one or more redundant results; by executing the instructions, the processor is also caused to perform the following operations: One or more linear operations are performed on at least two results to decode the multiple outputs to obtain the inference data, wherein the at least two results are from the multiple results and the one or more redundant results, and the at least two results include at least one redundant result of the one or more redundant results.

26. The device according to any one of claims 23 to 25, characterized in that Execution of the instructions further causes the processor to perform the one or more collective operations to obtain the inference data by performing a belief propagation process to decode the plurality of outputs, thereby decoding the plurality of outputs to obtain the inference data.

27. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 13.

28. An apparatus comprising a processor for performing the method according to any one of claims 1 to 13.