Beam neural network architecture

Abstract

Certain aspects of the present disclosure relate to methods and apparatus for operating an artificial neural network (ANN) device. The method includes transmitting a plurality of phase-shifted RF signals corresponding to a plurality of input signals from at least one artificial input neuron of the ANN using a plurality of transmit antennas and a plurality of transmit beams; receiving, at at least one artificial output neuron of the ANN, a plurality of phase-shifted RF signals from at least one artificial input neuron using a plurality of receive antennas; and generating a plurality of digital output signals corresponding to the plurality of input signals at at least one artificial output neuron of the ANN.

Description

Cross-references to related applications

[0001] This application claims priority to Israeli Patent Application No. 305741, filed on September 7, 2023, which has been assigned to the assignee of this application and is expressly incorporated herein by reference in its entirety, as if fully set forth below and for all applicable purposes. Technical Field

[0002] This disclosure relates to various aspects of artificial neural networks, and more specifically to beam neural network architectures. Background Technology

[0003] Machine learning is generally the process of generating trained models (e.g., artificial neural networks, trees, or other structures) that represent a generalized fit to a training dataset. Applying the trained model to input data produces inferences that can be used to gain insights into the input data. In some cases, applying the model to input data is described as “running inference” or “performing inference” on the input data.

[0004] To train a model and perform inference on input data, various mathematical processing components are used to perform various mathematical operations. For example, the multiply-accumulate-add (MAC) unit can be used to perform these operations to train the model and then use the trained model to perform inference on the input data. However, it should be noted that the MAC unit can be used for a wide variety of mathematical operations and is not so limited to those related to training the model and performing inference on the input data. These mathematical operations can be performed on various types of numerical data with varying complexities. Generally speaking, the complexity of these operations scales with the bit size and type of data. For example, operations using 8-bit integers may not be as computationally complex as performing inference using larger integers such as 64-bit integers. Similarly, operations using integers of a given bit size may not be as computationally complex as operations using floating-point numbers of a given bit size (e.g., operations using 32-bit integers may not be as computationally complex as operations using 32-bit floating-point numbers, even if the data has the same bit size).

[0005] Power utilization, heat output, and processing time typically scale with computational complexity. That is, computationally less complex operations generally consume less power and complete faster than computationally more complex operations. Therefore, executing computationally more complex operations can lead to reduced battery life and delays in the ability to reallocate computing resources (e.g., computing cores on the processor, memory, etc.) to other tasks being performed on the device.

[0006] Therefore, there is a need for improved techniques to perform MAC operations more efficiently. Summary of the Invention

[0007] Certain aspects of this disclosure relate to an artificial neural network device. The artificial neural network device may include at least one artificial input neuron configured to transmit multiple phase-shifted RF signals corresponding to multiple input signals using multiple transmit antennas and multiple transmit beams. The artificial neural network device may also include at least one artificial output neuron configured to: receive multiple phase-shifted RF signals from the at least one artificial input neuron using multiple receive antennas; and generate multiple digital output signals corresponding to the multiple input signals.

[0008] Certain aspects of this disclosure relate to a method for operating an artificial neural network device. The method may include transmitting multiple phase-shifted RF signals corresponding to multiple input signals from at least one artificial input neuron of the ANN using multiple transmit antennas and multiple transmit beams; receiving the multiple phase-shifted RF signals from at least one artificial input neuron of the ANN using multiple receive antennas at at least one artificial output neuron of the ANN; and generating multiple digital output signals corresponding to the multiple input signals at at least one artificial output neuron of the ANN.

[0009] Other aspects include: processing systems configured to perform the foregoing methods and those methods described herein; a non-transitory computer-readable medium including instructions that, when executed by one or more processors of the processing system, cause the processing system to perform the foregoing methods and those methods described herein; a computer program product embodied on a computer-readable storage medium including code for performing the foregoing methods and those methods further described herein; and a processing system including components for performing the foregoing methods and those methods further described herein.

[0010] The following description and related figures illustrate certain exemplary features of one or more aspects. Attached Figure Description

[0011] Figure 1 Example artificial neurons are illustrated based on certain aspects presented in this article.

[0012] Figure 2 Examples Figure 1 Example block-level hardware implementation of artificial neurons.

[0013] Figure 3 This is a block diagram of an example artificial neural network device configured to implement a beam artificial neural network, based on certain aspects presented in this article.

[0014] Figure 4Another embodiment of an artificial neural network device comprising multiple nonlinear components, based on certain aspects presented herein, is illustrated.

[0015] Figure 5 Another embodiment of an artificial neural network device that uses different sets of phase shifts, based on certain aspects presented herein, is illustrated.

[0016] Figure 6 Another embodiment of an artificial neural network device comprising multiple signal attenuators, based on certain aspects presented herein, is illustrated.

[0017] Figure 7 Additional layers of artificial neural network devices based on certain aspects presented in this paper are illustrated.

[0018] Figure 8 This is a flowchart illustrating operations for operating an artificial neural network device, based on certain aspects presented in this article.

[0019] Figure 9 Examples of processing systems according to various aspects of this disclosure may include various components configured to perform operations for the techniques disclosed herein. Detailed Implementation

[0020] Certain aspects of this disclosure relate to beamed artificial neural network (BNN) architectures. For example, artificial neural networks can be used to make predictions or inferences about certain input data or signals, such as for facial recognition, autonomous driving, and medical diagnosis. Traditional artificial neural networks (ANNs) operate by multiplying different weights with a set of input signals, summing the weighted input signals, and applying a nonlinear function (called an activation function) to the summed weighted input signals to obtain one or more output signals Y representing inferences or predictions about that set of input signals.

[0021] However, these multiplication and summation operations consume significant power, posing a major obstacle to the large-scale adoption of artificial intelligence (AI). Therefore, aspects of this disclosure provide techniques for reducing power consumption associated with artificial neural networks. In some cases, these techniques include bounded neural networks (BNNs) configured to reduce power consumption by avoiding multiplication of input signals and weights, as well as summation of weighted inputs. For example, a BNN can be configured to use multiple phase shifters to individually shift the phase of the corresponding input signals, rather than implementing a power-intensive multiplier to multiply the weights and corresponding input signals to generate multiple weighted input signals. Multiple phase shifters include fewer logic components and consume significantly less power than a multiplier.

[0022] Additionally, instead of implementing a power-consuming adder to sum the weighted input signals, the phase-shifted input signals of a BNN can interact with each other when wirelessly transmitted via multiple transmit antennas. This interaction between the phase-shifted input signals allows for efficient summation in the air, thus avoiding the need for adders (and the included logic components) and reducing power consumption compared to conventional ANNs.

[0023] The detailed description below, taken in conjunction with the accompanying drawings, is intended as a description of various configurations and is not intended to represent the only configuration in which the concepts described herein can be practiced. To provide a comprehensive understanding of the various concepts, the detailed description includes specific details. However, it will be apparent to those skilled in the art that these concepts can be practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form to avoid obscuring such concepts.

[0024] Based on the teachings, those skilled in the art will recognize that the scope of this disclosure is intended to cover any aspect of this disclosure, whether implemented independently of or in combination with any other aspect of this disclosure. For example, an apparatus or method may be implemented using any number of the aspects described. Furthermore, the scope of this disclosure is intended to cover such apparatuses or methods practiced using other structures, functionalities, or structures and functionalities that complement or differ from the various aspects of this disclosure described. It should be understood that any aspect of this disclosure may be embodied by one or more elements of the claims.

[0025] The word “exemplary” is used in this document to mean “serving as an example, instance, or illustration.” Any aspect described as “exemplary” in this document is not necessarily to be construed as preferred or superior to other aspects.

[0026] While specific aspects are described herein, numerous variations and arrangements of these aspects fall within the scope of this disclosure. Although some benefits and advantages of preferred aspects have been mentioned, the scope of this disclosure is not intended to be limited to a particular benefit, use, or purpose. Rather, aspects of this disclosure are intended to be widely applicable to various technologies, system configurations, networks, and protocols, some of which are illustrated by way of example in the accompanying drawings and the following description of preferred aspects. The detailed description and drawings are merely illustrative and not limiting of this disclosure, the scope of which is defined by the appended claims and their equivalents.

[0027] In biological neurons, the output spike generated when a neuron fires is called an action potential. This electrical signal is a relatively fast, transient neural impulse with an amplitude of approximately 100 mV and a duration of approximately 1 ms. In a particular implementation of a neural system with a series of connected neurons, each action potential has substantially the same amplitude and duration, and therefore, the information in the signal can be represented solely by the frequency and number of spikes or the duration of the spikes, rather than by the amplitude. The information carried by an action potential can be determined by the spike, the neuron to which the spike originates, and the duration of the spike relative to one or more other spikes. The importance of a spike can be determined by the weights applied to the connections between neurons, as will be explained below.

[0028] Spike transmission from one level of neurons to another can be achieved through synaptic networks (or simply "synapses"). Neurons that precede a synapse are considered presynaptic neurons, while those that follow a synapse are considered postsynaptic neurons. Postsynaptic neurons can receive output signals (i.e., spikes) from presynaptic neurons and, based on adjustable synaptic weights w1... (i,i+1) ...w P (i,i+1) The signals are scaled, where P is the total number of synaptic connections between presynaptic and postsynaptic neurons, and i is an indicator of the neuron level. Furthermore, the scaled signals can be combined with the input signals of each postsynaptic neuron. Each postsynaptic neuron can generate an output spike based on the corresponding combined input signals. The output spike can be transmitted to neurons at another level using another synaptic connection network.

[0029] Biological synapses can mediate excitatory or inhibitory (hyperpolarization) effects in postsynaptic neurons and can also amplify neuronal signals. Excitatory signals depolarize the membrane potential (i.e., increase the membrane potential relative to the resting potential). If sufficient excitatory signals are received over a period of time to depolarize the membrane potential above a threshold, an action potential occurs in the postsynaptic neuron. Conversely, inhibitory signals typically hyperpolarize the membrane potential (i.e., decrease it). Inhibitory signals, if strong enough, can counteract the sum of excitatory signals and prevent the membrane potential from reaching a threshold. In addition to counteracting synaptic excitation, synaptic inhibition can exert strong control over spontaneously active neurons. Spontaneously active neurons are those that spike in power without further input, such as due to their dynamics or feedback. By inhibiting the spontaneous generation of action potentials in these neurons, synaptic inhibition can shape the firing patterns in neurons, a process often referred to as shaping. Depending on the desired behavior, various synapses can act as any combination of excitatory or inhibitory synapses.

[0030] Recently, non-biological components have been used to model biological neurons and synapses in artificial neural networks.

[0031] In some cases, such as, artificial neural networks can be simulated by a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a graphics processing unit (GPU) or other programmable logic device (PLD), discrete gate or transistor logic devices, discrete hardware components, software modules executed by a processor, or any combination thereof.

[0032] Figure 1 An example artificial neuron 100 that can be implemented in an artificial neural network (ANN) according to certain aspects presented herein is illustrated. As illustrated, the artificial neuron 100 can receive multiple inputs (e.g., X...). A -X N ), which can be summed to produce an output. Typically, different weights (e.g., W) can be used. A -W N The weighted inputs are applied individually to each input to modify the importance of each of these inputs. The sum of the weighted inputs can be processed by a nonlinear function (e.g., F) (called the activation function or transfer function) to produce the output Y of the artificial neuron 100.

[0033] Artificial neuron 100 can be one of many artificial neurons in an ANN, which can be organized into layers. By adjusting the weights of the connections between artificial neurons in different layers of an ANN during the training process, the ANN can learn to map input data to desired outputs, enabling it to perform reasoning or prediction, classify objects, or recognize patterns in various domains such as image and speech recognition, natural language processing, and even medical diagnosis.

[0034] In modern artificial intelligence (AI) / machine learning engines that use artificial neural networks, neurons can be implemented using a multiply-accumulate (MAC) component. For example, Figure 2 A block diagram of MAC component 200 is shown, which can be used to implement ANN. Figure 1 The illustrated artificial neuron 100. As illustrated, the MAC component 200 may include one or more multipliers 202a-202n, an adder 204, and a nonlinear activation function 206. Furthermore, as illustrated, each multiplier, such as multiplier 202a, may receive an input (e.g., X). A ) and weights (e.g., W) AMultiplier 202a can then multiply the input and weights to generate a weighted input. The weighted inputs from each multiplier 202a-202n can then be passed to adder 204, which sums all the weighted inputs. The sum of the weighted inputs can then be passed through nonlinear activation function 206 to generate the output of the MAC component 200 of the artificial neuron 100. Aspects related to beam neural networks

[0035] As mentioned above, ANNs can be used for reasoning or prediction on input data, such as for image and speech recognition, natural language processing, and even medical diagnosis. A major problem in operating ANNs today is the very high power consumption and hardware complexity associated with the multipliers 202a-202n and adders 204 of each artificial neuron 100 implemented within the ANN. Not only does the ANN itself consume a very large amount of power, but there is also significant power consumption associated with cooling the ANN to prevent overheating and eventual electrical failure. While significant progress has been made in the application of ANNs to solve real-world problems, high power consumption poses a major obstacle to the large-scale adoption of AI.

[0036] These power consumption issues have led to additional research aimed at finding more power-efficient artificial neural network solutions, one of which involves spike neural networks (SNNs) inspired by how the human brain works. For example, SNNs use integral and firing spike neurons to exchange information via spiking or spikes. Because (e.g., when spikes occur) SNNs are relatively sparsely active, power consumption in SNNs can be relatively low. However, the reduction in power consumption that SNNs can provide is currently inhibited by the need to allow dozens or hundreds of spikes before efficiently converging to a single neuron in an equivalent ANN.

[0037] Therefore, aspects of this disclosure provide techniques for reducing power consumption associated with artificial neural networks. In some cases, these techniques include beam-based artificial neural networks (BNNs) configured to reduce power consumption by avoiding the use of multipliers (e.g., multipliers 202a-202n) and adders (e.g., adder 204). For example, instead of implementing power-consuming multipliers to multiply weights and corresponding input signals to generate multiple weighted input signals, a BNN can be configured to use multiple phase shifters to individually shift the phase of corresponding input signals, and multiple phase shifters include fewer logic components and consume significantly less power than multipliers. These phase-shifted input signals can then be wirelessly transmitted using multiple transmit antennas and received by multiple receive antennas.

[0038] Furthermore, instead of implementing a power-consuming adder (e.g., adder 204) to sum the weighted input signals, the phase-shifted input signals of the BNN can interact with each other when wirelessly transmitted via multiple transmit antennas. This interaction between the phase-shifted input signals allows for efficient summation of the phase-shifted input signals in the air, thereby avoiding the need for adders (and the included logic components) and reducing power consumption compared to conventional ANNs.

[0039] Figure 3 A block diagram illustrating an example artificial neural network device 300 configured to implement a BNN according to certain aspects of this disclosure is shown. In some cases, the artificial neural network device 300 may be configured to make predictions or inferences about input data based on a trained model. For example, in some cases, the trained model of the artificial neural network device 300 may be trained to make predictions or inferences about: (1) image recognition and computer vision (e.g., image classification, object detection, face recognition, and image generation), (2) natural language processing (e.g., language translation, sentiment analysis, text generation, and speech recognition), (3) healthcare (e.g., medical image analysis, disease diagnosis, drug discovery, and prediction of patient outcomes from medical records), (4) autonomous transportation (e.g., interpreting sensor data, making real-time decisions, and autonomous navigation), etc.

[0040] As shown in the figure, the artificial neural network device 300 includes a simple two-layer BNN, which comprises an input layer 302 and an output layer 304. The input layer 302 can correspond to artificial input neurons, and the output layer 304 can correspond to artificial output neurons. Although in Figure 3 Only two layers and two artificial neurons are illustrated, but the artificial neural network device 300 may include a BNN with additional layers and artificial neurons.

[0041] As will be described in more detail below, the artificial input neurons in input layer 302 can be configured to transmit multiple phase-shifted radio frequency (RF) signals corresponding to multiple input signals using multiple transmit antennas and multiple transmit beams. Correspondingly, the artificial output neurons in output layer 304 can be configured to receive multiple phase-shifted RF signals from the artificial input neurons using multiple receive antennas; and generate multiple digital output signals corresponding to the multiple input signals.

[0042] In some cases, the artificial neural network device 300 can be implemented within an RF transceiver. For example, in some cases, the artificial input neurons in the input layer 302 can be implemented by one or more transmit chains in the RF transceiver, and the artificial output neurons in the output layer 304 can be implemented by one or more receive chains in the RF transceiver. In some cases, one or more transmit chains of the RF transceiver can be positioned adjacent to and facing one or more receive chains, such that phase-shifted RF signals wirelessly transmitted from one or more transmit chains can be directly received by one or more receive chains.

[0043] As shown in the figure, the artificial input neurons (e.g., the sending chain) in the input layer 302 include multiple digital-to-analog converters (DACs) 306. A-N First multiple local oscillators (LO) 308 A-N Multiple phase shifters 310 A-N and multiple transmitting antennas 312 A-N Where A and N are positive integers. Additionally, as shown in the figure, the artificial output neurons (e.g., receiver chains) in the output layer 304 include multiple receiver antennas 314. A-M The second multiple LO 316 A-M and multiple analog-to-digital converters (ADCs) 318 A-M , where A and M are positive integers.

[0044] As shown in the figure, multiple DAC 306 A-N Each DAC in the configuration is configured to receive multiple digital input signals X. A-N The corresponding digital input signal in the circuit. For example, as shown in the figure, the first DAC 306 A It can be configured to receive a first digital input signal X A Second DAC 306 B It can be configured to receive a second digital input signal X B And the Nth DAC 306 N It can be configured to receive the Nth digital input sequence X N , where N is a positive integer. In some cases, the digital input signal may include data that is to be predicted or inferred by a trained model of the BNN of the artificial neural network device 300. For example, in some cases, the digital input signal may include image or video data input to the BNN of the artificial neural network device 300 for use in facial recognition, medical diagnosis, etc., as described above.

[0045] Furthermore, after receiving the corresponding digital input signal, multiple DAC 306 A-NEach DAC in the system can be configured to convert a corresponding digital input signal into a corresponding analog radio frequency (RF) signal. For example, converting a digital input signal can involve converting binary values ​​(such as zero and / or one) into continuous analog RF signals according to a specific sampling rate.

[0046] Subsequently, multiple DAC 306 A-N Each DAC in the circuit can be configured to output a corresponding analog RF signal to multiple phase shifters 310. A-N The corresponding phase shifters in the circuit. For example, the first DAC 306. A It can be configured to output a first analog RF signal to a first phase shifter 310. A Second DAC 306 B It can be configured to output a second analog RF signal to the second phase shifter 310. B ,etc.

[0047] As described above, the artificial input neurons in the input layer 302 include a first plurality of LO 308 A-N The first multiple LO308 A-N It can be configured to consist of multiple DAC 306 A-N The corresponding analog RF signal output by each DAC in the DAC is up-converted relative to the signal output by multiple DACs 306. A-N The DAC outputs a higher frequency corresponding to the frequency of the analog RF signal. For example, in some cases, the first analog RF signal can be generated by the first DAC 306. A Output at the first frequency. Afterwards, the first LO 308 A It can output a continuous sinusoidal signal with a second frequency higher than the first frequency of the first analog RF signal. Then, mixer 320 can be used to mix the first analog RF signal with a signal generated by the first LO 308. A The output continuous sinusoidal signal is mixed to upconvert the frequency of the first analog RF signal to a third frequency, which represents the sum of the first frequency of the first analog RF signal and the second frequency of the continuous sinusoidal signal. Therefore, the third frequency is a higher frequency relative to the first frequency of the first analog RF signal.

[0048] After receiving the corresponding analog RF signal (up-converted), multiple phase shifters 310 A-N Each corresponding phase shifter is configured to apply a corresponding phase shift to a corresponding analog RF signal associated with that phase shifter to generate a corresponding phase-shifted RF signal among a plurality of phase-shifted RF signals. In some cases, multiple phase shifters 310 A-NEach phase shifter in the configuration may include one or more delay elements configured to introduce a controlled delay into the corresponding analog RF signal, thereby causing a phase shift in the corresponding analog RF signal. For example, the first phase shifter 310... A It can be configured to shift the first phase W A Applied to from the first DAC 306 A The first corresponding analog RF signal output, the second phase shifter 310 B It can be configured to shift the second phase W B Applied to from the second DAC 306 B The output of the second corresponding analog RF signal, etc.

[0049] In some cases, each phase shift in the corresponding phase shift can represent or correspond to a corresponding weight of the trained model of the artificial neural network device 300. As shown in the figure, there may be... N One phase shift or weight, where N It is a positive integer. In some cases, when a BNN implemented by an artificial neural network device is fully connected, the total number of phase shifts or weights can be equal to... N * M ,in M It is the total number of digital output signals Y.

[0050] As mentioned above, compared to using multipliers (such as...) Figure 2 The illustrated multipliers 202a-202n consume significantly less power by shifting the phase of the corresponding analog RF signal compared to multiplying the corresponding weights with the corresponding input signals. For example, this is achieved with multiple phase shifters 310 used to introduce controlled delays to shift the phase of the corresponding analog RF signal. A-N Compared to one or more delay elements within a single element, a multiplier can include significantly more logic components, which collectively consume significantly more power. As a result, relative to... Figure 2 The illustrated MAC component 200 and artificial neural network device 300 can reduce power consumption. The MAC component is configured using a multiplier and uses the multiplier to multiply the corresponding weights with the corresponding input signals.

[0051] After shifting the phase of the corresponding analog RF signal, multiple phase shifters 310 A-N Each corresponding phase shifter can output the corresponding phase-shifted RF signal associated with that phase shifter to multiple transmit antennas 312. A-N The corresponding transmitting antenna in the [element / concept]. For example, the first phase-shifted analog RF signal can be transmitted from the first phase shifter 310. A Output to the first transmitting antenna 312 A The second phase-shifted analog RF signal can be generated from the second phase shifter 310. B Output to the second transmitting antenna 312B ,etc.

[0052] Subsequently, multiple transmitting antennas 312 A-N Each corresponding transmit antenna is configured to use multiple transmit beams 322 A-N The corresponding transmit beam will wirelessly transmit the corresponding phase-shifted RF signal associated with the corresponding transmit antenna to multiple receive antennas 312. A-M For example, as shown in the figure, a first transmitting antenna 312 can be used. A and the first transmitting beam 322 A To wirelessly transmit the first phase-shifted analog RF signal, a second transmitting antenna 312 can be used. B Second transmission beam 322 B To wirelessly transmit second phase-shifted analog RF signals, and so on.

[0053] As shown in the figure, multiple transmission beams 322 can be transmitted in different spatial directions. A-N Each transmit beam in the spectrum, the spatial direction of which can be controlled based on a corresponding phase shift associated with a corresponding phase-shifted RF signal. For example, the first transmit beam 322 can be controlled based on a first phase shift associated with a first corresponding phase-shifted RF signal. A The spatial orientation can be controlled based on the second phase shift associated with the second corresponding phase shift RF signal, allowing the second transmit beam 322 to be controlled. B Spatial direction, etc.

[0054] In some cases, multiple transmitting antennas 312 can be used. A-N One of the transmitting antennas in the system simultaneously transmits each corresponding phase-shifted RF signal from multiple phase-shifted RF signals. As a result, the corresponding phase-shifted RF signals can be combined with each other in the air, effectively leading to a summation of the corresponding phase-shifted RF signals in the air. Because the corresponding phase-shifted RF signals can be summed in the air, the artificial neural network device 300 may not require an adder, such as... Figure 2 The illustrated adder 204 can further reduce the power consumption of the artificial neural network device 300 (e.g., relative to...). Figure 2 The illustrated MAC component 200).

[0055] As a result of this air summation and based on multiple transmitting antennas 312 A-N The transmitting antenna in the middle is relative to multiple receiving antennas 314 A-M The location of the receiving antennas, multiple receiving antennas 314 A-M Each receiving antenna in the configuration can be configured to receive signals from multiple transmitting antennas 312. A-N Receive different corresponding composite analog RF signals. For example, each different composite analog RF signal can be based on multiple transmitting antennas 312.A-N Different combinations of corresponding phase-shifted RF signals (and their associated phases) transmitted by the corresponding transmitting antennas. For example, due to multiple transmitting antennas 312 A-N Each transmitting antenna in the first receiving antenna 314 A The location of the first receiving antenna 314 A It can be configured to receive a first composite analog RF signal. Similarly, due to multiple transmit antennas 312 A-N Each transmitting antenna in the second receiving antenna 314 M The location of the second receiving antenna 314 M It can be configured to receive a second composite analog RF signal.

[0056] After receiving different corresponding composite analog RF signals, multiple receiving antennas 314 A-M Each corresponding receiving antenna in the array can output different corresponding composite analog RF signals associated with that antenna to multiple ADCs 318. A-M The corresponding ADC in the [database / array]. For example, the first receiving antenna 314. A It can be configured to output a first composite analog RF signal to a first ADC 318 A And the second receiving antenna 314 M It can be configured to output a second composite analog RF signal to a second ADC 318. M .

[0057] As described above, the artificial output neurons in output layer 304 include a second plurality of LO 316 A-M In some cases, the signal is obtained from multiple receiving antennas 314 by the corresponding ADC. A-M Before the corresponding composite analog RF signal output by each of the receiving antennas in the receiving antenna, the second multiple LO 316 A-M It can be configured to downconvert the corresponding composite analog RF signal to a frequency relative to the multiple receiving antennas 314 A-M The receiving antenna in the signal receives and outputs a lower frequency of the corresponding composite analog RF signal. For example, in some cases, the first composite analog RF signal can be received by the first receiving antenna 314. A Receive and output at the first frequency. Afterwards, the first LO 316 A It can output a continuous sinusoidal signal with a second frequency lower than the first frequency of the first composite analog RF signal. Then, mixer 324 can be used to mix the first composite analog RF signal with a signal from the first LO 316. AThe output continuous sinusoidal signal is mixed to downconvert the frequency of the first composite analog RF signal to a third frequency, which represents the sum of the first frequency of the first composite analog RF signal and the second frequency of the continuous sinusoidal signal. Therefore, the third frequency is a lower frequency relative to the first frequency of the first composite analog RF signal.

[0058] As shown in the figure, multiple ADC 318 A-M Each corresponding ADC in the array is configured to obtain (down-converted) a different corresponding composite analog RF signal from the receiving antenna associated with that corresponding ADC. Subsequently, multiple ADCs 318 A-M Each corresponding ADC in the array is configured to convert different corresponding composite analog RF signals (down-converted) obtained from the receiving antenna associated with that corresponding ADC into multiple digital output signals Y. A-M The corresponding digital output signal. In some cases, the corresponding digital output signal can represent multiple digital input signals X. A-N The prediction or inference of the multiple digital input signals X A-N It is generated by the application of the corresponding phase shift or weights of the trained model of the artificial neural network device 300.

[0059] In some cases, due to the above-described air summation, multiple digital output signals Y A-M Each digital output signal in the circuit can be represented by multiple phase shifters 310 A-N Different phase shift modifications for different applications and multiple receiving antennas 314 A-M The receiving antenna and multiple transmitting antennas 312 A-N The sum of digital input signals modified by the different relative distances between the transmitting antennas. For example, the first corresponding digital output signal Y A It can be used with multiple receiving antennas 314 A-M The first receiving antenna 314 A Related. Furthermore, as shown in the figure, the first receiving antenna 314 A With multiple transmitting antennas 312 A-N Each corresponding transmitting antenna in (e.g., 312) A 312 B ……312 N The different relative distances (e.g., d) A d B ...d N As a result, the first ADC 318 A The first corresponding digital output signal Y A It is due to different corresponding phase shifts (e.g., with the first phase shifter 310) A The associated first phase shifter and second phase shifter 310 B(associated with the second phase shift, etc.), and the first receiving antenna 314 A To multiple transmitting antennas 312 A-N Each corresponding transmitting antenna in (e.g., 312) A 312 B ……312 N The different relative distances (e.g., d) A d B ...d N Modify the corresponding digital input signal X A-N The sum of .

[0060] As mentioned above, in a conventional ANN, such as regarding Figure 1 and Figure 2 In an ANN (Application Not Node), the weighted input signal can be summed and passed through a nonlinear function (e.g., F), called the activation function or transfer function, to produce an output Y. The application of this nonlinear function is a key step that allows ANNs to capture and model complex relationships in the input data. For example, many real-world problems involve complex and nonlinear relationships between input signals and desired outputs. A linear combination of input signals might only model linear relationships, which may be insufficient to capture the complexity of the underlying data patterns in the input signals. Instead, a nonlinear activation function introduces the ability to represent more complex relationships, enabling ANNs to learn and generalize from data more effectively.

[0061] In some cases, this nonlinear function can be implemented in different ways within the BNN of the artificial neural network device 300. For example, in some cases, multiple ADCs 318 A-M Each corresponding ADC in the array can be configured to limit the different composite analog RF signals associated with that corresponding ADC, thereby introducing nonlinearity in the different composite analog RF signals associated with that corresponding ADC and the corresponding output signal. For example, in some cases, the first ADC 318 A It can be configured to frequency limit the first composite analog RF signal at a specific level, so as to combine the first composite analog RF signal and the first output signal Y. A Nonlinearity is introduced.

[0062] In some cases, this nonlinearity can be introduced by one or more nonlinear components of the artificial neural network device 300. For example, Figure 4 Another embodiment of the artificial neural network device 300 is illustrated. For example... Figure 4 As shown in the implementation scheme, the output layer 304 of the artificial neural network device 300 may include multiple nonlinear components 326 A-M In some cases, multiple nonlinear components 326 A-M It can be configured to output multiple digital signals YA-M Nonlinearity (e.g., a nonlinear function F similar to that described above) is introduced into the corresponding digital output signal. For example, nonlinear component 326 A It can be configured to be in the first ADC318 A Process the first composite analog RF signal and use it as the first output signal Y A Before output, a nonlinearity is introduced into the first composite analog RF signal. In some cases, multiple nonlinear components 326 A-M It can include multiple low-noise amplifiers (LNAs), multiple power amplifiers (PAs), or multiple ADCs 318. A-M At least one of them.

[0063] In some cases, multiple phase shifters 310 A-N A set of phase shifts or weights can be applied to multiple output signals Y. A-M All multiple output signals. In other cases, the phase shifts or weights of different groups can be generated by multiple phase shifters 310. A-N Application, and multiple output signals Y in different groups A-M Shared between. Figure 5 Examples of using different groups of phase shifts or weights are shown in the figure.

[0064] For example, Figure 3 The illustrated multiple phase shifters 310 A-N It may include at least a first set of phase shifters and a second set of phase shifters. For example, such as Figure 5 As shown, the first set of phase shifters may include a first phase shifter 310. A Second phase shifter 310 B The second set of phase shifters may include a third phase shifter 310. C and the fourth phase shifter 310 D In some cases, the first set of phase shifters 310 A 310 B It can be configured to shift the first set of phases W 1,A and W 1,B Applied to the first DAC 306 respectively A Second DAC 306 B The corresponding analog RF signals output, which correspond to the first set of digital input signals X A and X B For example, the first set of phase shifters 310 A 310 B It can be configured to shift the first set of phases W 1,A and W 1,B Applied to the corresponding analog RF signal to generate the above-mentioned... Figure 3The first group of phase-shifted RF signals in the description of multiple phase-shifted RF signals.

[0065] Similarly, the second set of phase shifters 310 C 310 D It can be configured to shift the second set of phases W 2,A and W 2,B Applied to the third DAC 306 respectively C and the fourth DAC 306 D The corresponding analog RF signals output, which correspond to the first set of digital input signals X C and X D For example, the second set of phase shifters 310 C 310 D It can be configured to shift the second set of phases W 2,A and W 2,B Applied to the corresponding analog RF signal to generate the above-mentioned... Figure 3 The second set of phase-shifted RF signals in a plurality of phase-shifted RF signals described. In some cases, the first set of phase shifts (e.g., W) 1,A and W 1,B ) can be different from the second set of phase shifts (e.g., W) 2,A and W 2,B In some cases, when using different sets of phase shifts or weights, the total number of trainable weights / phase shifts in a BNN implemented by the artificial neural network device 300 can include 2* N Each phase shift or weight.

[0066] like Figure 5 As shown, Figure 3 Multiple ADCs 318 A-M It can include the first ADC 318 A Second ADC 318 B Third ADC318 C And the fourth ADC 318 D Similarly, such as Figure 5 As shown, Figure 3 Multiple receiving antennas 314 A-M It may include a first receiving antenna 314 A Second receiving antenna 314 B Third receiving antenna 314 C and the fourth receiving antenna 314 DIn some cases, when using different sets of phase shifts or weights, the artificial neural network device 300 can selectively apply these different sets of phase shifts or weights to different sets of digital output signals. In some cases, selectively applying different sets of phase shifts or weights to different sets of digital output signals may involve the artificial neural network device 300 selectively putting different sets of ADCs into an on or off state. In some cases, the artificial neural network device 300 may selectively put different sets of ADCs into an on or off state based on control signaling from, for example, one or more processors associated with the artificial neural network device 300.

[0067] For example, in some cases, the first group of ADCs (such as the first ADC 318) A Second ADC 318 B It can be configured to receive from the first group of receiving antennas among a plurality of receiving antennas (such as the first receiving antenna 314). A Second receiving antenna 314 B The first set of phase-shifted RF signals is obtained. Additionally, a second set of ADCs (such as a third ADC 318) is used. C And the fourth ADC 318 D It can be configured to receive signals from a second group of multiple receiving antennas (such as a third receiving antenna 314). C and the fourth receiving antenna 314 D The second set of phase-shifted RF signals is obtained.

[0068] In such cases, when the first set of phase-shifted RF signals is obtained (e.g., it is based on the first set of phase-shifted W...), 1,A and W 1,B When generated, the first group of ADCs (e.g., the first ADC 318) A Second ADC 318 B The second ADC (e.g., the third ADC 318) can be configured to be in a powered-on state. C And the fourth ADC 318 D The system is configured to be in a power-off state. Similarly, when a second set of phase-shifted RF signals is obtained (e.g., it is based on a second set of phase-shifted W...), 2,A and W 2,B When generated), the second group of ADCs (e.g., the third ADC 318) C And the fourth ADC 318 D The first ADC (e.g., the first ADC 318) is configured to be powered on. A Second ADC 318 B It is configured to be in a power-off state.

[0069] Due to the different phase shifts or weights of the different groups and the selective placement of different groups of ADCs into power-on and power-off states, the first group of digital output signals (such as the first digital output signal Y) A Second digital output signal Y B It can be phase-shifted with the first group W. 1,A and W 1,B Related. Similarly, the second set of digital output signals (such as the third digital output signal Y) C and the fourth digital output signal Y D It can be phase-shifted with the second group W. 2,A and W 2,B Related.

[0070] In some cases, the artificial neural network device 300 may also include multiple signal attenuators that can act as an additional set of weights or trainable parameters (e.g., attenuation levels) for the BNN implemented in the artificial neural network device 300. For example, as Figure 6 As shown, at least one artificial output neuron in the output layer 304 also includes a first plurality of signal attenuators 330. A-M The first plurality of signal attenuators are configured to receive signals from the plurality of receiving antennas 314 A-M The received composite analog RF signal is processed by multiple ADCs 318 A-M The corresponding ADC in the input layer 302 attenuates the corresponding composite analog RF signal beforehand. In some cases, the artificial input neurons in the input layer 302 may also include a second or more signal attenuators 332. A-N The second plurality of signal attenuators can be placed in the plurality of phase shifters 310 A-N Previously, and was configured to be in use by multiple DAC 306 A-N After outputting the corresponding analog RF signal, at least one of the corresponding analog RF signals is attenuated, such as... Figure 6 As shown. In some cases, a second multiple signal attenuator 332 A-N It can be placed in multiple phase shifters 310 A-N Subsequently, and configured to operate using multiple phase shifters 310 A-N The corresponding phase shifter outputs the corresponding phase-shifted RF signal and then attenuates the corresponding phase-shifted RF signal.

[0071] As mentioned above, although in Figure 3 Only two layers and two artificial neurons are illustrated, but the artificial neural network device 300 may include a BNN with one or more intermediate layers. For example, as shown... Figure 7As illustrated, the artificial neural network device 300 may include an intermediate layer 702 disposed between the input layer 302 and the output layer 304. Although only one intermediate layer 702 is illustrated, it should be understood that the artificial neural network device 300 may include any number of intermediate layers.

[0072] As shown in the figure, the artificial input neurons in the input layer 302 are configured to use phase shifter 310. A Based on digital input signal X A A phase-shifted RF signal is generated. Afterwards, the transmitting antenna 312 can be used. A The phase-shifted RF signal is wirelessly transmitted to the intermediate artificial neuron in the intermediate layer 702. The intermediate layer 702 can receive the phase-shifted RF signal using the receiving antenna 704. The phase-shifted RF signal received from the input layer 302 can then be used as an input signal in the intermediate layer 702. For example, as shown, the phase-shifted RF signal can be input to a phase shifter 706, which is configured to apply another phase shift to the phase-shifted RF signal. In some cases, as shown, the phase shift applied by the phase shifter 706 in the intermediate layer 702 (e.g., W) A,2 (This can be different from the phase shifter 310 in the input layer 302) A The phase shift of the application (e.g., W) A,1 This includes phase shifts applied by phase shifters in other intermediate layers. Furthermore, as shown, before the phase-shifted RF signal is input to phase shifter 706, it can be input to power amplifier (PA) 708, which can be configured to amplify the power of the phase-shifted RF signal.

[0073] After that, although Figure 7 Not shown, but (dual) phase-shifted RF signals can be wirelessly transmitted to one or more other intermediate layers configured to apply additional phase shifts. However, for simplicity, Figure 7 An example is illustrated of a (double) phase-shifted RF signal wirelessly transmitted from intermediate layer 702 to an artificial output neuron in output layer 304. After receiving the (double) phase-shifted RF signal from intermediate layer 702, the artificial output neuron in output layer 304 can process the (double) phase-shifted RF signal using the techniques described above to output a digital output signal Y. A .

[0074] Figure 8 This is a flowchart illustrating operation 800 for operating an artificial neural network (ANN) device (e.g., for a neural network) according to various aspects of this disclosure. Operation 800 may, for example, be performed by a processing system 900, such as a user equipment (UE), or other computing device, on which operations are performed for various tasks. Figure 9The illustrated computing device performs these operations. These operations may include, for example, operations defined for an artificial neural network that generates one or more inferences from input data or signals.

[0075] As illustrated in the figure, operation 800 can begin at box 810, using multiple transmit antennas and multiple transmit beams to transmit multiple phase-shifted RF signals corresponding to multiple input signals from at least one artificial input neuron of the ANN.

[0076] Operation 800 continues in box 815, where multiple phase-shifted RF signals are received from at least one artificial input neuron at at least one artificial output neuron using multiple receiving antennas.

[0077] Operation 800 continues in box 820, wherein multiple digital output signals corresponding to multiple input signals are generated at at least one artificial output neuron of the ANN.

[0078] In some cases, at least one artificial input neuron includes multiple digital-to-analog converters (DACs), multiple phase shifters, and multiple transmit antennas. In some cases, at least one artificial output neuron includes multiple receive antennas and multiple analog-to-digital converters (ADCs).

[0079] In some cases, operation 800 further includes performing at least one of the following: attenuating the corresponding analog RF signal after it is output by the DAC, or attenuating the corresponding phase-shifted RF signal after it is output by the corresponding phase shifter. In some cases, each of a plurality of DACs can be configured to receive signals from one or more processors (such as those described below). Figure 9 The central processing unit (CPU) 902, graphics processing unit (GPU) 904, digital signal processor (DSP) 906 and / or neural processing unit (NPU) 908 of the described processing system 900 receive corresponding digital input signals.

[0080] In some cases, operation 800 also includes applying a corresponding phase shift to a corresponding analog RF signal associated with that corresponding phase shifter at each of the plurality of phase shifters to generate a corresponding phase-shifted RF signal among a plurality of phase-shifted RF signals, and outputting the corresponding phase-shifted RF signal associated with that corresponding phase shifter to a corresponding transmit antenna among a plurality of transmit antennas. In some cases, the corresponding phase shift corresponds to a set of weights of an ANN implemented by an ANN device. In some cases, the corresponding phase shift can be performed by one or more processors (such as those described below). Figure 9 The described processing system 900 (CPU 902, GPU 904, DSP 906, and / or NPU 908) provides power to each corresponding phase shifter. In some cases, the corresponding phase shift can be stored in memory, such as as described below. Figure 9 The memory 924 of the processing system 900 described.

[0081] In some cases, operation 800 also includes, at each of the plurality of transmit antennas, wirelessly transmitting a corresponding phase-shifted RF signal associated with that corresponding transmit antenna to a plurality of receive antennas using a corresponding transmit beam from a plurality of transmit beams.

[0082] In some cases, operation 800 further includes receiving different corresponding composite analog RF signals from multiple transmit antennas at each of the multiple receive antennas, wherein each different composite analog RF signal is based on a different combination of corresponding phase-shifted RF signals transmitted by the corresponding transmit antennas. In some cases, operation 800 further includes outputting different corresponding composite analog RF signals associated with the corresponding antenna to a corresponding ADC of multiple ADCs at each of the multiple receive antennas.

[0083] In some cases, operation 800 also includes, at each of the multiple ADCs, obtaining different corresponding composite analog RF signals from the receiving antenna associated with that corresponding ADC, and converting the different corresponding composite analog RF signals obtained from the receiving antenna associated with that corresponding ADC into a corresponding digital output signal among a plurality of digital output signals. In some cases, the corresponding digital output signal may be provided to another artificial input neuron of the ANN and / or may be provided to one or more processors (such as those described below). Figure 9 The described processing system 900 includes a CPU 902, a GPU 904, a DSP 906, and / or an NPU 908 for further processing.

[0084] In some cases, the first corresponding digital output signal is associated with a first receiving antenna among a plurality of receiving antennas. In some cases, the first receiving antenna has a different relative distance to each corresponding transmitting antenna among a plurality of transmitting antennas. In some cases, the first corresponding digital output signal is the sum of corresponding digital input signals modified by different corresponding phase shifts and different relative distances from the first receiving antenna to each corresponding transmitting antenna.

[0085] In some cases, operation 800 also includes, at each of the plurality of ADCs, clipping the different composite analog RF signals associated with that respective ADC to introduce nonlinearity in the different composite analog RF signals associated with that respective ADC.

[0086] In some cases, at least one artificial output neuron includes multiple nonlinear components configured to introduce nonlinearity within a corresponding digital output signal among multiple digital output signals. In some cases, the multiple nonlinear components include at least one of multiple low-noise amplifiers (LNAs), multiple power amplifiers (PAs), or multiple ADCs.

[0087] In some cases, the multiple phase shifters include at least a first set of phase shifters and a second set of phase shifters. In some cases, operation 800 further includes applying a first set of phase shifters to a corresponding analog RF signal corresponding to a first set of digital input signals to generate a first set of phase-shifted RF signals among a plurality of phase-shifted RF signals. In some cases, operation 800 further includes applying a second set of phase shifters to a corresponding analog RF signal corresponding to a second set of digital input signals to generate a second set of phase-shifted RF signals among a plurality of phase-shifted RF signals, wherein the first set of phase shifts is different from the second set of phase shifts. In some cases, the first set of phase shifts and the second set of phase shifts may be stored in a memory (such as memory 924) and processed by one or more processors (such as those described below). Figure 9 The CPU 902, GPU 904, DSP 906, and / or NPU 908 of the described processing system 900 are respectively provided to the first set of phase shifters and the second set of phase shifters.

[0088] In some cases, a first group of ADCs among multiple ADCs is configured to obtain a first group of phase-shifted RF signals from a first group of receiving antennas among multiple receiving antennas. In some cases, a second group of ADCs among multiple ADCs is configured to obtain a second group of phase-shifted RF signals from a second group of receiving antennas among multiple receiving antennas.

[0089] In some cases, when the first set of phase-shifted RF signals is received, the first set of ADCs is configured to be powered on, and the second set of ADCs is configured to be powered off. In some cases, when the second set of phase-shifted RF signals is received, the second set of ADCs is configured to be powered on, and the first set of ADCs is configured to be powered off. In some cases, one or more processors (such as those described below) Figure 9 The CPU 902, GPU 904, DSP 906 and / or NPU 908 of the described processing system 900 can be configured to output control signals to control whether the first group of ADCs and the second group of ADCs are in a powered-on or powered-off state.

[0090] In some cases, at least one artificial output neuron also includes multiple signal attenuators. In some cases, operation 800 also includes attenuating the corresponding composite analog RF signal before it is obtained from the corresponding ADC. In some cases, the attenuation level applied to the composite analog RF signal can be determined by one or more processors (such as those described below). Figure 9 The CPU 902, GPU 904, DSP 906 and / or NPU 908 of the described processing system 900 are controlled or configured.

[0091] In some cases, at least one artificial input neuron also includes multiple signal attenuators. In some cases, operation 800 further includes attenuating at least one of the corresponding analog RF signal or the corresponding phase-shifted RF signal after the corresponding analog RF signal is output from the DAC and after the corresponding phase-shifted RF signal is output from the corresponding phase shifter. In some cases, the attenuation level applied to the corresponding analog RF signal or the corresponding phase-shifted RF signal can be determined by one or more processors (such as those described below). Figure 9 The CPU 902, GPU 904, DSP 906 and / or NPU 908 of the described processing system 900 are controlled or configured.

[0092] In some cases, at least one artificial output neuron also includes multiple local oscillators (LOs). In some cases, operation 800 also includes downconverting the corresponding composite analog RF signal to a lower frequency relative to the frequency at which the corresponding composite analog RF signal is received before the corresponding composite analog RF signal is obtained from the corresponding ADC.

[0093] In some cases, at least one artificial input neuron also includes multiple local oscillators (LOs). In some cases, operation 800 also includes upconverting the corresponding analog RF signal to a higher frequency relative to the frequency at which the corresponding analog RF signal is output by the DAC.

[0094] Figure 9 An example processing system 900 is depicted. In some cases, the processing system can be configured to perform one or more operations for operating an artificial neural network device, such as regarding... Figure 8 The operation described is 800.

[0095] Processing system 900 includes CPU 902, which in some examples may be a multi-core CPU. Instructions executed at CPU 902 may be loaded, for example, from program memory associated with CPU 902, or may be loaded from memory 924. In some cases, instructions executed by CPU 902 may cause processing system 900 to perform one or more operations for operating an artificial neural network device, such as regarding... Figure 8 The operation described is 800.

[0096] The processing system 900 also includes additional processing components tailored for specific functions, such as GPU 904, DSP 906, NPU 908, multimedia component 910, and wireless connectivity component 912.

[0097] NPUs (such as the NPU 908) are generally configured to implement dedicated circuitry for implementing control and arithmetic logic for executing machine learning algorithms, such as those for processing artificial neural networks (ANNs), deep neural networks (DNNs), random forests (RFs), etc. NPUs are sometimes alternatively referred to as neural signal processors (NSPs), tensor processing units (TPUs), neural network processors (NNPs), intelligent processing units (IPUs), vision processing units (VPUs), or graphics processing units.

[0098] NPUs (such as the NPU 908) are configured to accelerate the execution of common machine learning tasks such as image classification, machine translation, object detection, and various other predictive models. In some examples, multiple NPUs may be instantiated on a single chip (such as a system-on-a-chip (SoC)), while in other examples, an NPU may be part of a dedicated neural network accelerator.

[0099] An NPU can be optimized for either training or inference, or in some cases configured to balance performance between the two. For an NPU capable of performing both training and inference, these two tasks can generally still be performed independently.

[0100] NPUs designed to accelerate training are typically configured to speed up the optimization of new models. This involves taking an existing dataset (usually labeled or sublabeled), iterating over the dataset, and subsequently tuning model parameters (such as weights and biases) to improve model performance—a highly computationally intensive operation. Generally, optimization based on incorrect predictions involves backtracking through the layers of the model and determining gradients to reduce prediction errors.

[0101] NPUs designed to accelerate inference are typically configured to operate on the full model. Therefore, such NPUs can be configured to take new data slices as input and rapidly process them through a trained model to generate model outputs (e.g., inference).

[0102] In one specific implementation, the NPU 908 is part of one or more of the CPU 902, GPU 904, and / or DSP 906.

[0103] In some examples, the wireless connectivity component 912 may include sub-components for, for example, third-generation (3G) connectivity, fourth-generation (4G) connectivity (e.g., 4G LTE), fifth-generation connectivity (e.g., 5G or NR), Wi-Fi connectivity, Bluetooth connectivity, and other wireless data transmission standards. The wireless connectivity component 912 is further connected to one or more antennas 914.

[0104] The processing system 900 may also include one or more sensor processing units 916 associated with any type of sensor, one or more image signal processors (ISPs) 918 associated with any type of image sensor, and / or a navigation component 920, which may include a satellite-based positioning system component (such as GPS or GLONASS) and an inertial positioning system component.

[0105] The processing system 900 may also include one or more input and / or output devices 922, such as screens, touch-sensitive surfaces (including touch-sensitive displays), physical buttons, speakers, microphones, etc.

[0106] In some examples, one or more processors in the processing system 900 may be based on the ARM or RISC-V instruction set.

[0107] Furthermore, as illustrated in the figure, the processing system 900 may include an artificial neural network device 300, which is configured to implement the above-mentioned... Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 The BNN is described. In some cases, an artificial neural network device 300, combined with one or more processors (such as CPU 902, GPU 904, DSP 906, and / or NPU 908), can be configured to perform operations related to... Figure 8 The operation described is 800.

[0108] The processing system 900 also includes a memory 924, which represents one or more static and / or dynamic memories, such as dynamic random access memory, flash-based static memory, etc. In some cases, the memory 924 may include executable instructions that, when executed by one or more of the aforementioned processors, cause the processing system 900 to perform one or more operations for operating the artificial neural network device 300, such as regarding... Figure 8 Operation 800 is described. Additionally, as described above, in some cases, one or more phase shifts of the artificial neural network device 300 may be stored in memory 924 and provided to the artificial neural network device 300 by one or more of the processors described above.

[0109] Generally speaking, the processing system 900 and / or its components can be configured to perform the methods described herein.

[0110] It is worth noting that features of the processing system 900 may be omitted in other aspects, such as when the processing system 900 is a server computer. For example, the multimedia component 910, wireless connection component 912, sensor processing unit 916, ISP 918, and / or navigation component 920 may be omitted in other aspects. Furthermore, aspects of the processing system 900 may be distributed, such as training a model and using the model to generate inferences. Example Terms

[0111] Specific implementation examples are described in the following numbered clauses.

[0112] Clause 1: A method for operating an artificial neural network (ANN) device, the method comprising: transmitting a plurality of phase-shifted RF signals corresponding to a plurality of input signals from at least one artificial input neuron of the ANN using a plurality of transmit antennas and a plurality of transmit beams; receiving the plurality of phase-shifted RF signals from at least one artificial output neuron of the ANN using a plurality of receive antennas; and generating a plurality of digital output signals corresponding to the plurality of input signals at at least one artificial output neuron of the ANN.

[0113] Clause 2: The method according to Clause 1, wherein: the at least one artificial input neuron comprises: a plurality of digital-to-analog converters (DACs); a plurality of phase shifters; and the plurality of transmitting antennas; and the at least one artificial output neuron comprises: the plurality of receiving antennas; and a plurality of analog-to-digital converters (ADCs).

[0114] Clause 3: The method according to Clause 2 further includes, at each of the plurality of DACs: obtaining a corresponding digital input signal among the plurality of digital input signals; converting the corresponding digital input signal into a corresponding analog radio frequency (RF) signal; and outputting the corresponding analog RF signal to a corresponding phase shifter among the plurality of phase shifters.

[0115] Clause 4: The method according to Clause 3 further includes, at each of the plurality of phase shifters, applying a corresponding phase shift to the corresponding analog RF signal associated with the corresponding phase shifter to generate a corresponding phase-shifted RF signal among the plurality of phase-shifted RF signals; and outputting the corresponding phase-shifted RF signal associated with the corresponding phase shifter to a corresponding transmit antenna among the plurality of transmit antennas.

[0116] Clause 5: The method according to Clause 4, wherein the corresponding phase shift corresponds to a set of weights of the ANN implemented by the ANN device.

[0117] Clause 6: The method according to Clause 4 further includes, at each of the plurality of transmit antennas, wirelessly transmitting the corresponding phase-shifted RF signal associated with the corresponding transmit antenna to the plurality of receive antennas using the corresponding transmit beam of the plurality of transmit beams.

[0118] Clause 7: The method according to Clause 6 further includes, at each of the plurality of receiving antennas: receiving different corresponding composite analog RF signals from the plurality of transmitting antennas, wherein each different composite analog RF signal is based on a different combination of corresponding phase-shifted RF signals transmitted by the corresponding transmitting antenna of the plurality of transmitting antennas; and outputting the different corresponding composite analog RF signals associated with the corresponding antenna to a corresponding ADC of the plurality of ADCs.

[0119] Clause 8: The method according to Clause 7 further includes, at each of the plurality of ADCs: obtaining the different corresponding composite analog RF signals from the receiving antenna associated with the corresponding ADC; and converting the different corresponding composite analog RF signals obtained from the receiving antenna associated with the corresponding ADC into a corresponding digital output signal among the plurality of digital output signals.

[0120] Clause 9: The method according to Clause 8, wherein: a first corresponding digital output signal is associated with a first receiving antenna of the plurality of receiving antennas; the first receiving antenna has a different relative distance to each corresponding transmitting antenna of the plurality of transmitting antennas; and the first corresponding digital output signal is the sum of the corresponding digital input signals modified by the different corresponding phase shifts and the different relative distances from the first receiving antenna to each corresponding transmitting antenna.

[0121] Clause 10: The method according to Clause 8 further includes, at each of the plurality of ADCs, limiting the different composite analog RF signals associated with the respective ADC to introduce nonlinearity in the different composite analog RF signals associated with the respective ADC.

[0122] Clause 11: The method according to Clause 8, wherein: the at least one artificial output neuron includes a plurality of nonlinear components configured to introduce nonlinearity within a respective digital output signal of the plurality of digital output signals; and the plurality of nonlinear components includes at least one of a plurality of low-noise amplifiers (LNAs), a plurality of power amplifiers (PAs), or the plurality of ADCs.

[0123] Clause 12: The method according to Clause 8, wherein: the plurality of phase shifters includes at least a first set of phase shifters and a second set of phase shifters; and the method further includes: applying a first set of phase shifts to the corresponding analog RF signal corresponding to a first set of digital input signals using the first set of phase shifters to generate a first set of phase-shifted RF signals among the plurality of phase-shifted RF signals; and applying a second set of phase shifts to the corresponding analog RF signal corresponding to a second set of digital input signals using the second set of phase shifters to generate a second set of phase-shifted RF signals among the plurality of phase-shifted RF signals, wherein the first set of phase shifts is different from the second set of phase shifts.

[0124] Clause 13: The method according to Clause 12, wherein: a first group of ADCs of the plurality of ADCs is configured to obtain a first group of phase-shifted RF signals from a first group of receiving antennas of the plurality of receiving antennas; and a second group of ADCs of the plurality of ADCs is configured to obtain a second group of phase-shifted RF signals from a second group of receiving antennas of the plurality of receiving antennas.

[0125] Clause 14: The method according to Clause 13, wherein: when the first set of phase-shifted RF signals is obtained, the first set of ADCs is configured to be in a powered-on state, and the second set of ADCs is configured to be in a powered-off state; and when the second set of phase-shifted RF signals is obtained, the second set of ADCs is configured to be in the powered-on state, and the first set of ADCs is configured to be in the powered-off state.

[0126] Clause 15: The method according to Clause 8, wherein: the at least one artificial output neuron further includes a plurality of signal attenuators; and the method further includes attenuating the corresponding composite analog RF signal using the plurality of signal attenuators of the at least one artificial output neuron before the corresponding composite analog RF signal is obtained by the corresponding ADC.

[0127] Clause 16: The method according to Clause 8, wherein: the at least one artificial input neuron further comprises a plurality of signal attenuators; and the method further comprises using the plurality of signal attenuators of the at least one artificial input neuron to perform at least one of the following: attenuating the corresponding analog RF signal after the corresponding analog RF signal is output by the DAC, or attenuating the corresponding phase-shifted RF signal after the corresponding phase-shifted RF signal is output by the corresponding phase-shifter.

[0128] Clause 17: The method according to Clause 8, wherein: the at least one artificial output neuron further comprises a plurality of local oscillators (LOs); and the method further comprises, before the corresponding composite analog RF signal is obtained by the corresponding ADC, using the plurality of LOs of the at least one artificial output neuron to downconvert the corresponding composite analog RF signal to a lower frequency relative to the frequency at which the corresponding composite analog RF signal is received.

[0129] Clause 18: The method according to Clause 8, wherein: the at least one artificial input neuron further includes a plurality of local oscillators (LOs); and the method further includes using the plurality of LOs of the at least one artificial input neuron to upconvert the corresponding analog RF signal to a higher frequency relative to the frequency at which the corresponding analog RF signal is output by the DAC.

[0130] Clause 19: An apparatus comprising: one or more memories including executable instructions; and one or more processors configured to execute the executable instructions and cause the apparatus to perform a method according to any one of Clauses 1 to 18.

[0131] Clause 20: An apparatus comprising components for performing the method according to any one of Clauses 1 to 18.

[0132] Clause 21: A non-transitory computer-readable medium comprising executable instructions that, when executed by one or more processors of a device, cause the device to perform a method according to any one of Clauses 1 to 18.

[0133] Clause 22: A computer program product embodied on a computer-readable storage medium, the computer program product comprising code for performing a method according to any one of Clauses 1 to 18. Additional Notes

[0134] The foregoing description is provided to enable any person skilled in the art to practice the various aspects described herein. The examples discussed herein do not limit the scope, applicability, or aspects set forth in the claims. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects. For example, the function and arrangement of the elements discussed may be changed without departing from the scope of this disclosure. Various processes or components may be omitted, substituted, or added as appropriate in various examples. For example, the described methods may be performed in a different order than described, and various actions may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined in some other examples. For example, any number of aspects set forth herein may be used to implement an apparatus or practice. Moreover, the scope of this disclosure is intended to cover such apparatuses or methods practiced using other structures, functionalities, or structures and functionalities that complement or replace the various aspects of this disclosure set forth herein. It should be understood that any aspect of the disclosure herein may be embodied by one or more elements of the claims.

[0135] The various exemplary logic blocks, modules, and circuits described in this disclosure can be implemented or executed using a general-purpose processor, digital signal processor (DSP), ASIC, field-programmable gate array (FPGA) or other programmable logic device (PLD), discrete gate or transistor logic device, discrete hardware component, or any combination thereof designed to perform the functions described herein. While the general-purpose processor may be a microprocessor, in alternative embodiments, the processor may be any commercially available processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors working in conjunction with a DSP core, a system-on-a-chip (SoC), or any other such configuration.

[0136] As used herein, "processor," "at least one processor," or "one or more processors" generally refers to a single processor configured to perform one or more operations, or multiple processors configured to collectively perform one or more operations. In the case of multiple processors, the execution of one or more operations may be divided among different processors, but one processor may perform multiple operations, and multiple processors may collectively perform a single operation. Similarly, "memory," "at least one memory," or "one or more memory" generally refers to a single memory configured to store data and / or instructions, or multiple memories configured to collectively store data and / or instructions.

[0137] As used in this article, the phrase “at least one of” in a list of items refers to any combination of these items, including a single member. As an example, “at least one of a, b, or c” is intended to cover: a, b, c, ab, ac, bc, and abc, as well as any combination with multiple identical elements (e.g., aa, aaa, aab, aac, abb, acc, bb, bbb, bbb, bbc, cc, and ccc, or any other ordering of a, b, and c).

[0138] As used herein, the term "determine" encompasses a wide variety of actions. For example, "determine" can include calculation, operation, processing, deduction, investigation, lookup (e.g., searching in a table, database, or other data structure), assertion, etc. Additionally, "determine" can include receiving (e.g., receiving information), accessing (e.g., accessing data in memory), etc. Furthermore, "determine" can include parsing, selecting, picking, building, etc.

[0139] The methods disclosed herein include one or more actions for implementing the methods. These actions may be interchanged without departing from the scope of the claims. In other words, unless a specific order of actions is specified, the order and / or use of a particular action may be modified without departing from the scope of the claims. Furthermore, the various operations of the methods described above may be performed by any suitable component capable of performing the corresponding function. This component may include various hardware and / or software components and / or modules, including but not limited to circuits, application-specific integrated circuits (ASICs), or processors.

[0140] The following claims are not intended to be limited to the aspects shown herein, but should be given the full scope consistent with the language of the claims. Within the claims, unless specifically stated otherwise, reference to the singular form of an element is not intended to mean “one and only one,” but rather “one or more.” Unless otherwise specifically stated, the term “some” refers to one or more. No element of any claim shall be interpreted in accordance with 35 USC §112(f) unless that element is explicitly stated using the phrase “for a component of.” All structural and functional equivalents of the elements throughout the various aspects described in this disclosure that are known to a person skilled in the art or will later be known are expressly incorporated herein by reference and are intended to be covered by the claims. Furthermore, nothing disclosed herein is intended to be offered to the public, whether or not such disclosure is explicitly recited in the claims.

Claims

1. An artificial neural network (ANN) device, said artificial neural network (ANN) device comprising: At least one artificial input neuron, the at least one artificial input neuron being configured to transmit multiple phase-shifted RF signals corresponding to multiple input signals using multiple transmit antennas and multiple transmit beams; and At least one artificial output neuron, said at least one artificial output neuron being configured to: The plurality of phase-shifted RF signals are received from the at least one artificial input neuron using multiple receiving antennas; and Generate multiple digital output signals corresponding to the multiple input signals.

2. The ANN device according to claim 1, wherein: The at least one artificial input neuron includes: Multiple digital-to-analog converters (DACs); Multiple phase shifters; and The plurality of transmitting antennas; and The at least one artificial output neuron includes: The plurality of receiving antennas; and Multiple analog-to-digital converters (ADCs).

3. The ANN device of claim 2, further comprising using each of the plurality of DACs to: Obtain the corresponding digital input signal from the plurality of digital input signals; Convert the corresponding digital input signal into a corresponding analog radio frequency (RF) signal; and The corresponding analog RF signal is output to the corresponding phase shifter among the plurality of phase shifters.

4. The ANN device of claim 3, wherein each of the plurality of phase shifters is configured to: The corresponding phase shift is applied to the corresponding analog RF signal associated with the corresponding phase shifter to generate the corresponding phase-shifted RF signal among the plurality of phase-shifted RF signals; and The corresponding phase-shifted RF signal associated with the corresponding phase shifter is output to the corresponding transmit antenna among the plurality of transmit antennas.

5. The ANN device of claim 4, wherein the corresponding phase shift corresponds to a set of weights of the ANN implemented by the ANN device.

6. The ANN device of claim 4, wherein each of the plurality of transmit antennas is configured to wirelessly transmit the corresponding phase-shifted RF signal associated with the corresponding transmit antenna to the plurality of receive antennas using the corresponding transmit beam of the plurality of transmit beams.

7. The ANN device of claim 6, wherein each of the plurality of receiving antennas is configured to: Different corresponding composite analog RF signals are received from the plurality of transmitting antennas, wherein each different composite analog RF signal is based on a different combination of corresponding phase-shifted RF signals transmitted by the corresponding transmitting antenna among the plurality of transmitting antennas; and The different corresponding composite analog RF signals associated with the corresponding antenna are output to the corresponding ADCs among the plurality of ADCs.

8. The ANN device of claim 7, wherein each of the plurality of ADCs is configured to: The different corresponding composite analog RF signals are obtained from the receiving antenna associated with the corresponding ADC; and The different corresponding composite analog RF signals obtained from the receiving antenna associated with the corresponding ADC are converted into a corresponding digital output signal among the plurality of digital output signals.

9. The ANN device according to claim 8, wherein: The first corresponding digital output signal is associated with the first receiving antenna among the plurality of receiving antennas; The first receiving antenna has a different relative distance to each corresponding transmitting antenna among the plurality of transmitting antennas; and The first corresponding digital output signal is the sum of the corresponding digital input signals modified by different corresponding phase shifts and different relative distances from the first receiving antenna to each corresponding transmitting antenna.

10. The ANN device of claim 8, wherein each of the plurality of ADCs is configured to limit the different composite analog RF signals associated with the respective ADC to introduce nonlinearity in the different composite analog RF signals associated with the respective ADC.

11. The ANN device according to claim 8, wherein: The at least one artificial output neuron includes a plurality of nonlinear components, the plurality of nonlinear components being configured to introduce nonlinearity within a corresponding digital output signal among the plurality of digital output signals; and The plurality of nonlinear components include at least one of a plurality of low-noise amplifiers (LNAs), a plurality of power amplifiers (PAs), or a plurality of ADCs.

12. The ANN device according to claim 8, wherein: The plurality of phase shifters includes at least a first set of phase shifters and a second set of phase shifters; The first set of phase shifters is configured to apply a first set of phase shifts to the corresponding analog RF signals corresponding to the first set of digital input signals to generate the first set of phase-shifted RF signals among the plurality of phase-shifted RF signals; The second set of phase shifters is configured to apply a second set of phase shifts to the corresponding analog RF signals corresponding to the second set of digital input signals to generate the second set of phase-shifted RF signals among the plurality of phase-shifted RF signals; and The first set of phase shifts is different from the second set of phase shifts.

13. The ANN device according to claim 12, wherein: The first group of ADCs among the plurality of ADCs is configured to obtain the first group of phase-shifted RF signals from the first group of receiving antennas among the plurality of receiving antennas; and The second group of ADCs among the plurality of ADCs is configured to obtain the second group of phase-shifted RF signals from the second group of receiving antennas among the plurality of receiving antennas.

14. The ANN device according to claim 13, wherein: When the first set of phase-shifted RF signals is obtained, the first set of ADCs is configured to be in a powered-on state, and the second set of ADCs is configured to be in a powered-off state. and When the second set of phase-shifted RF signals is obtained, the second set of ADCs is configured to be in the powered-on state, and the first set of ADCs is configured to be in the powered-off state.

15. The ANN device of claim 8, wherein the at least one artificial output neuron further comprises a plurality of signal attenuators configured to attenuate the respective composite analog RF signal before the respective composite analog RF signal is obtained by the respective ADC.

16. The ANN device of claim 8, wherein the at least one artificial input neuron further comprises a plurality of signal attenuators configured to perform at least one of the following: attenuating the corresponding analog RF signal after the corresponding analog RF signal is output by the DAC or attenuating the corresponding phase-shifted RF signal after the corresponding phase-shifted RF signal is output by the corresponding phase-shifter.

17. The ANN device of claim 8, wherein the at least one artificial output neuron further comprises a plurality of local oscillators (LOs) configured to downconvert the respective composite analog RF signal to a lower frequency relative to the frequency at which the respective composite analog RF signal is received before the respective composite analog RF signal is obtained by the respective ADC.

18. The ANN device of claim 8, wherein the at least one artificial input neuron further comprises a plurality of local oscillators (LOs) configured to upconvert the respective analog RF signal to a higher frequency relative to the frequency at which the respective analog RF signal is output by the DAC.

19. A method for operating an artificial neural network (ANN) device, the method comprising: Multiple phase-shifted RF signals corresponding to multiple input signals are transmitted from at least one artificial input neuron of the ANN using multiple transmit antennas and multiple transmit beams; At at least one artificial output neuron of the ANN, multiple receiving antennas are used to receive the multiple phase-shifted RF signals from the at least one artificial input neuron; as well as At the at least one artificial output neuron of the ANN, a plurality of digital output signals corresponding to the plurality of input signals are generated.

20. The method of claim 19, wherein: The at least one artificial input neuron includes: Multiple digital-to-analog converters (DACs); Multiple phase shifters; and The plurality of transmitting antennas; and The at least one artificial output neuron includes: The plurality of receiving antennas; and Multiple analog-to-digital converters (ADCs); and The method further includes, at each of the plurality of DACs: Obtain the corresponding digital input signal from the plurality of digital input signals; Convert the corresponding digital input signal into a corresponding analog radio frequency (RF) signal; and The corresponding analog RF signal is output to the corresponding phase shifter among the plurality of phase shifters.

21. The method according to claim 20, further comprising: At each of the plurality of phase shifters: The corresponding phase shift is applied to the corresponding analog RF signal associated with the corresponding phase shifter to generate the corresponding phase-shifted RF signal among the plurality of phase-shifted RF signals; as well as The corresponding phase-shifted RF signal associated with the corresponding phase shifter is output to the corresponding transmitting antenna among the plurality of transmitting antennas; At each of the plurality of transmitting antennas, the corresponding phase-shifted RF signal associated with the corresponding transmitting antenna is wirelessly transmitted to the plurality of receiving antennas using the corresponding transmitting beam of the plurality of transmitting beams; as well as At each of the plurality of receiving antennas: Different corresponding composite analog RF signals are received from the plurality of transmitting antennas, wherein each different composite analog RF signal is based on a different combination of corresponding phase-shifted RF signals transmitted by the corresponding transmitting antenna among the plurality of transmitting antennas; as well as The different corresponding composite analog RF signals associated with the corresponding antenna are output to the corresponding ADCs among the plurality of ADCs.

22. The method of claim 21, wherein the corresponding phase shift corresponds to a set of weights of an ANN implemented by the ANN device.

23. The method of claim 21, further comprising, at each of the plurality of ADCs: The different corresponding composite analog RF signals are obtained from the receiving antenna associated with the corresponding ADC; and The different corresponding composite analog RF signals obtained from the receiving antenna associated with the corresponding ADC are converted into a corresponding digital output signal among the plurality of digital output signals.

24. The method according to claim 23, wherein: The first corresponding digital output signal is associated with the first receiving antenna among the plurality of receiving antennas; The first receiving antenna has a different relative distance to each corresponding transmitting antenna among the plurality of transmitting antennas; and The first corresponding digital output signal is the sum of the corresponding digital input signals modified by different corresponding phase shifts and different relative distances from the first receiving antenna to each corresponding transmitting antenna.

25. The method of claim 23, wherein, At least one of the following conditions exists: The method further includes, at each of the plurality of ADCs, limiting the different composite analog RF signals associated with the respective ADC to introduce nonlinearity in the different composite analog RF signals associated with the respective ADC; or The at least one artificial output neuron includes a plurality of nonlinear components, the plurality of nonlinear components being configured to introduce nonlinearity within a corresponding digital output signal among the plurality of digital output signals; and The plurality of nonlinear components include at least one of a plurality of low-noise amplifiers (LNAs), a plurality of power amplifiers (PAs), or a plurality of ADCs.

26. The method of claim 23, wherein: The plurality of phase shifters includes at least a first set of phase shifters and a second set of phase shifters; and The method further includes: The first set of phase shifters is used to apply a first set of phase shifts to the corresponding analog RF signals corresponding to the first set of digital input signals to generate a first set of phase-shifted RF signals among the plurality of phase-shifted RF signals; and The second set of phase shifters is used to apply the second set of phase shifts to the corresponding analog RF signals corresponding to the second set of digital input signals to generate the second set of phase-shifted RF signals among the plurality of phase-shifted RF signals, wherein the first set of phase shifts is different from the second set of phase shifts.

27. The method according to claim 26, wherein: The first group of ADCs among the plurality of ADCs is configured to obtain the first group of phase-shifted RF signals from the first group of receiving antennas among the plurality of receiving antennas; The second group of ADCs among the plurality of ADCs is configured to obtain the second group of phase-shifted RF signals from the second group of receiving antennas among the plurality of receiving antennas; When the first set of phase-shifted RF signals is obtained, the first set of ADCs is configured to be in a powered-on state, and the second set of ADCs is configured to be in a powered-off state. and When the second set of phase-shifted RF signals is obtained, the second set of ADCs is configured to be in the powered-on state, and the first set of ADCs is configured to be in the powered-off state.

28. The method according to claim 23, wherein: The at least one artificial input neuron also includes multiple local oscillators (LO); The at least one artificial output neuron also includes multiple local oscillators (LO); The at least one artificial input neuron also includes multiple signal attenuators; The at least one artificial output neuron further includes multiple signal attenuators; and The method further includes: The corresponding analog RF signal is up-converted to a higher frequency relative to the frequency at which the corresponding analog RF signal is output by the DAC using the plurality of LOs of the at least one artificial input neuron; as well as Before the corresponding composite analog RF signal is obtained by the corresponding ADC, the corresponding composite analog RF signal is down-converted to a lower frequency relative to the frequency at which the corresponding composite analog RF signal is received using the plurality of LOs of the at least one artificial output neuron. The plurality of signal attenuators using the at least one artificial input neuron perform at least one of the following: attenuating the corresponding analog RF signal after the corresponding analog RF signal is output by the DAC, or attenuating the corresponding phase-shifted RF signal after the corresponding phase-shifted RF signal is output by the corresponding phase-shifter; and Before the corresponding composite analog RF signal is obtained from the corresponding ADC, the corresponding composite analog RF signal is attenuated using the plurality of signal attenuators of the at least one artificial output neuron.

29. An apparatus for operating an artificial neural network (ANN) device, the apparatus comprising: A component for transmitting multiple phase-shifted RF signals corresponding to multiple input signals from at least one artificial input neuron of the ANN using multiple transmitting antennas and multiple transmitting beams; A component for receiving the plurality of phase-shifted RF signals from the at least one artificial input neuron using a plurality of receiving antennas at at least one artificial output neuron of the ANN; and A component for generating a plurality of digital output signals corresponding to the plurality of input signals at at least one artificial output neuron of the ANN.

30. A non-transitory computer-readable medium for operating an artificial neural network (ANN) device, the non-transitory computer-readable medium comprising: Instructions, which, when executed by one or more processors of the device, cause the device to: Multiple phase-shifted RF signals corresponding to multiple input signals are transmitted from at least one artificial input neuron of the ANN using multiple transmit antennas and multiple transmit beams; At at least one artificial output neuron of the ANN, multiple receiving antennas are used to receive the multiple phase-shifted RF signals from the at least one artificial input neuron; as well as At the at least one artificial output neuron of the ANN, a plurality of digital output signals corresponding to the plurality of input signals are generated.

Images