Simulation system for edge artificial intelligence calculation

Abstract

A simulation system for edge artificial intelligence computing, comprising: a first plurality of simulated edge devices configured to receive an input analog signal and output a first plurality of output analog signals; a second plurality of analog edge devices configured to receive the first plurality of output analog signals and output a second plurality of output analog signals; and one or more memory devices in communication with the first plurality of simulated edge devices and the second plurality of simulated edge devices and configured to store weight parameters that are adjustable based on time constants of the first plurality of simulated edge devices or the second plurality of simulated edge devices, or both. The second plurality of output analog signals is multiplied by a weighting parameter to obtain a plurality of weighted analog signals.

Description

[0001] Cross-reference to related applications

[0002] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 684,581, filed August 19, 2024, and U.S. Patent Application No. 19 / 241,117, filed June 17, 2025, which are incorporated herein by reference in their entirety.

[0003] Government licensing rights

[0004] This invention was made with government support under IARPA designation 2022-21102100011. The government enjoys certain rights to this invention. Technical Field

[0005] This disclosure generally relates to a simulation system for edge artificial intelligence computing. Background Technology

[0006] Artificial intelligence (AI) computing at the edge is often constrained by size, weights, and power (SWaP). Existing digital computer-based methods are insufficient for edge AI computing. Power-intensive techniques used to train AI models in data centers are generally unsuitable for edge applications. Many edge applications require the interpretation and inference of analog data. Existing digital computer-based methods are highly inefficient due to the burden of analog-to-digital conversion (ADC) and digital-to-analog conversion (DAC) and other processing bottlenecks inherent in the von Neumann architecture. Instead of building better graphics processing units (GPUs) or smaller transistors, a new approach to designing AI hardware for edge computing is needed. Another issue of particular interest is in-situ training for edge applications. Traditional stochastic gradient descent (SGD) learning methods rely on backpropagation of errors through multi-layered neural networks, which is difficult to implement in analog hardware. Attached Figure Description

[0007] As illustrated in the accompanying drawings, features and advantages will become apparent from the following more specific description of various exemplary aspects, wherein similar reference numerals generally denote the same, functionally similar and / or structurally similar elements.

[0008] The complete and enabling disclosure of this description, including its best mode, is set forth in the specification with reference to the accompanying drawings, for those skilled in the art, wherein:

[0009] Figure 1 It is a network architecture of a microelectromechanical system (MEMS) device in a neural network for reducing noise in an input analog signal or for denoising an input analog signal, according to aspects of this disclosure.

[0010] Figure 2This describes an electronic circuitry illustrating the interconnection between multiple analog edge devices (e.g., MEMS) for implementing an analog system for edge artificial intelligence using a neural network and one or more field-programmable analog arrays (FPAAs) according to aspects of this disclosure.

[0011] Figure 3 This is a schematic diagram of an example MEMS device according to aspects of this disclosure.

[0012] Figure 4 This is a schematic diagram of a MEMS device used as a neuron implementing a neural network, according to aspects of this disclosure.

[0013] Figure 5 It is a graph depicting the noise input analog signal, the noise-reduced output analog signal, and the comparison with the actual ground analog signal according to aspects of this disclosure.

[0014] Figure 6 This is a flowchart of a machine learning algorithm (e.g., a continuous-time autoregressive recurrent neural network (CTRNN)) using aspects of this disclosure to solve ordinary differential equations.

[0015] Figure 7 This is a schematic diagram illustrating an example implementation of multi-level hierarchical differential target propagation for in-situ training of MEMS CTRNNs according to aspects of this disclosure.

[0016] Figure 8 This is a flowchart illustrating the details of the differential target propagation scheme used by CTRNN according to aspects of this disclosure. Detailed Implementation

[0017] The features, advantages, and aspects of this disclosure will be set forth or apparent from consideration of the following detailed description, accompanying drawings, and claims. Furthermore, the following detailed description is exemplary and intended to provide further explanation, without limiting the scope of the claimed disclosure.

[0018] Various aspects of this disclosure are discussed in detail below. While specific aspects are discussed, this is for illustrative purposes only. Those skilled in the art will recognize that other components and configurations can be used without departing from this disclosure.

[0019] A MEMS analog edge device is provided that, compared to existing digital computing architectures, can perform edge artificial intelligence (AI) computations with relatively low power and relatively high speeds (e.g., hundreds of times lower power and hundreds of times faster speeds). The term "MEMS device" is broadly used to include mechanical MEMS devices, photoacoustic MEMS devices, ferroelectric MEMS devices, etc. The term "edge" is used throughout this document to refer to computing devices at the edge or end of a network architecture near the user, such as cellular phones, sensors (e.g., image sensors, sound sensors, radiation or light sensors, etc.), controllers for machines, drones (UAVs), or other network-connected devices. Analog computing is particularly powerful for edge AI applications where size, weight, and power-as-a-P (SWaP) play a significant role. Furthermore, analog computing is particularly powerful for edge AI applications where most of the collected data is generated by analog edge devices (e.g., analog sensors). By performing AI computations in the analog domain, the need for bidirectional analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) can be eliminated. Moreover, analog continuous-time processing does not require constant data shifting in and out of memory, which addresses one of the major bottlenecks of digital architectures. As a result, by operating in the analog domain, fast (e.g., real-time, virtually no latency), ultra-low power (e.g., mW to μW), and low-cost AI computation can be achieved at the edge. Furthermore, in-situ training allows edge AI systems to mitigate drift and variations in analog hardware. Actual deployment of edge analog hardware may require in-situ learning.

[0020] In the first alternative to digital-based AI computing, digital microcontrollers can be miniaturized to reduce the SWaP (Power-Swap) used for edge AI computing. However, this first alternative is still based on a digital architecture and includes the additional computational burden of analog-to-digital conversion and the bottlenecks of the von Neumann architecture. Furthermore, analog operations are more power efficient than digital operations that rely on transistor-based operations (such as those involving AND and OR gates).

[0021] In a second alternative, the analog machine learning using the device can be employed as a programmable resistor or analog memory. However: 1) these resistor-based devices typically consume more power than aspects of this disclosure, including capacitor-based MEMS CTRNNs; 2) these resistor-based devices are “static” devices, i.e., they do not “learn” time information as in aspects of this disclosure; and 3) these resistor-based devices do not have in-situ learning capabilities.

[0022] Analog computers use fundamental physical phenomena as computational primitives and represent information through the relative values ​​of analog signals. In one aspect of this disclosure, the analog edge computing device can be a MEMS device. In this aspect of the disclosure, the MEMS device controlled by dynamic differential equations is the computational primitive. For example, a MEMS cantilever beam as a computational primitive provides a relationship between voltage input and mechanical displacement output. Both input and output can be continuous-time analog signals. The relationship between input and output, also known as a transfer function, can represent simple operations such as multiplication, addition, or integration, or more complex operations including filtering, nonlinear functions, harmonic oscillations, etc. Due to lower power consumption and real-time response, MEMS devices exhibit a power reduction of more than 300 times and a speed increase of more than 100 times compared to digital workstations used to solve the same artificial intelligence (AI) problems.

[0023] For example, MEMS devices used to implement machine learning (ML) algorithms (e.g., continuous-time autoregressive recurrent neural networks, CTRNNs) can directly process input analog signals and can learn temporal information using the recurrent feature of the neural network. A non-limiting aspect of the recurrent feature is achieved through a self-feedback "memory" enabled by, for example, electrostatic MEMS. A CTRNN is a deep learning model that interprets irregular observations by introducing a continuous evolution of hidden states between observations. This recurrent feature is achieved through a self-feedback memory enabled by an electrostatic MEMS device. In one aspect of this disclosure, a MEMS CTRNN is constructed by connecting multiple MEMS devices to form a fully connected network. For example, the output of a MEMS device (i.e., a MEMS neuron) is multiplied by weights. After multiplication with the corresponding weights, the outputs of the multiple MEMS neurons are summed to provide inputs to the MEMS neurons at different layers. In one aspect of this disclosure, different types of neurons can be used in the CTRNN, including MEMS neurons and other analog neurons. Examples of analog neurons include resistors, capacitors, operational amplifiers, programmable analog memories, and other electronic circuit elements.

[0024] In one aspect of this disclosure, the MEMS device (e.g., a MEMS processor) is controlled by electrostatic excitation to provide a capacitive output. Analog electronics for the MEMS device can be constructed using an analog front-end and a field-programmable analog array (FPAA). The MEMS device has an adjustable time constant that can be controlled by voltage (called a weighting parameter). The weighting parameter is set in the FPAA. The output from the MEMS device is multiplied by the weighting parameter and applied to the next neuron layer (MEMS or analog).

[0025] In one aspect of this disclosure, the adjustable time constant of the MEMS device enables the CTRNN to learn salient information in a time series. The CTRNN can also be constructed from MEMS devices with different time constants for time learning. Learning salient information from a time series involves essentially ingesting the time series (e.g., analog signals such as audio signals) and extracting actionable information from the signals. MEMS technology is well-suited for learning salient information because the MEMS device is configured to extract state or meaning from a time series. In one non-limiting example, aspects of this disclosure can be used for acoustic navigation and ranging (sonar) underwater target detection to detect and locate objects of interest or probe. In another non-limiting example, aspects of this disclosure can be used to detect anomalous signals from machine vibration characteristics and use this information for predictive maintenance. In yet another non-limiting example, applying aspects of this disclosure to seismic signals also enables the detection of seismic activity.

[0026] In one aspect of this disclosure, Differential Objective Propagation (DTP) can be used to update weight parameters at local layers, instead of the standard backpropagation algorithm which identifies weight gradients by propagating the global loss error back through the network. This allows for in-situ training at local layers, which greatly simplifies analog circuit design. Using local layers differs from the standard backpropagation algorithm in many ways. Machine learning neural network models can include several processing layers, where information is transformed layer by layer into an increasingly larger information space. The standard backpropagation algorithm takes the output error and propagates it backward through all layers of the network to update the parameters. The DTP method predicts what the objective should be for each layer and then directly computes the error signal for each layer. The computed layer-specific error signal is then used to update the weights. Thus, essentially, the standard backpropagation method computes network-level error, while the DTP method computes the error for each layer of the network. In contrast to the standard backpropagation algorithm, which identifies weight gradients by propagating the global loss error back through the network, DTP uses an autoencoder-type architecture to establish intermediate objectives. An autoencoder is a neural network that is trained to copy its inputs to its output, typically through a compressed latent representation. DTP has two main advantages: (1) the information required at the output of each layer of the neural network can be obtained by simply forwarding through the network, and (2) the update equation is local for each layer, which means that only local information is needed to update the weights of layer 'i'.

[0027] Another issue of particular interest is in-situ training for edge applications. To date, the analog computing community has largely neglected this problem. Traditional stochastic gradient descent (SGD) learning methods rely on backpropagation of error through multi-layered neural networks, which is difficult to implement in analog hardware. In gradient descent, all available data is used to generate predictions for all available data using a machine learning neural network model. The error is then computed and minimized by moving the network parameters in the direction that minimizes the error. In SGD, instead, a subset of the data is used. The error gradient is computed for the subset of data. The model is then updated based on the computed error.

[0028] In SGD, gradients can have associated noise. This noise may be due to the sampling process or inherent noise in the analog input / output signals. The presence of noise in SGD can actually help with convergence, as machine learning neural network models are less likely to get stuck in local minima.

[0029] In one aspect of this disclosure, in-situ training is used to address two problems of the limited analog computers of the past: (1) simulating hardware changes, wherein this technique allows the neural network to learn its parameters based on separate edge hardware, and (2) simulating hardware drift, wherein this technique allows the neural network (NN) to recalibrate itself by relearning, thereby mitigating drift caused by environmental factors (e.g., temperature, humidity, etc.).

[0030] Analog edge devices (e.g., MEMS devices or other analog devices) are indeed different from each other, and their characteristics can drift over time. In-situ training allows each of the edge AI systems to learn its parameters based on individual edge hardware. This also allows the edge AI system to recalibrate itself through relearning, thereby mitigating the drift of the analog hardware.

[0031] In-situ training of MEMS CTRNN methods provides adaptive techniques to mitigate the effects of component variations. This adaptation naturally leads to a system that understands its environment. MEMS CTRNNs with in-situ training can learn and provide updated weight parameters based on hardware components instantiated at very high speeds (e.g., a few minutes of training time) and with very low energy consumption (e.g., less than 1 watt-hour). MEMS CTRNNs are also more robust to component degradation and failure than more conventional systems.

[0032] As a result, one aspect of this disclosure is to provide a MEMS CTRNN that utilizes electrostatic MEMS neurons to perform network computation in continuous time and the analog domain. The technical effects of this aspect of the disclosure allow or enable in-situ training of the MEMS CTRNN method, which provides adaptive techniques to mitigate the effects of component variations, allows or enables more robust systems than more conventional systems to resist component degradation and failure, and allows or enables edge computing with low power and high performance. Another aspect of this disclosure also provides in-situ supervised training of CTRNNs, which allows learning directly on edge hardware using received analog information and self-correction for variations and drift in the analog hardware.

[0033] Low SWaP and fast response edge AI enable numerous applications both within and outside the aerospace industry. These applications can include signal filtering (adaptive), prediction, tracking, classification, anomaly detection, or decoding information based on continuous signals. The deployment of edge AI will rapidly increase for autonomous platforms requiring SWaP, such as drones, autonomous vehicles, and robots.

[0034] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations. Furthermore, unless specifically stated otherwise, all aspects disclosed herein should be considered exemplary.

[0035] As used herein, terms such as “first” and “second” are used interchangeably to distinguish one component from another and are not intended to indicate the location or importance of individual components.

[0036] As used herein, the term "upstream" refers to the direction opposite to the direction of fluid flow, and the term "downstream" refers to the direction in the same direction as the direction of fluid flow. The terms "forward" or "front" refer to something in front of it, and "backward" or "rear" refer to something behind it. For example, when used in relation to fluid flow, forward / front can mean upstream, and backward / rear can mean downstream.

[0037] Additionally, as used herein, the term "radial" or "radially" refers to a direction away from a common center. For example, in the general case of a gas turbine engine, radial refers to the direction of a ray extending between the engine's central longitudinal axis and its outer circumference. Furthermore, as used herein, the term "group" or a "set" of elements can refer to any number of elements, including only one.

[0038] Unless otherwise stated, connection references (e.g., attachment, coupling, fixing, fastening, joining, and engagement) are broadly interpreted and may include intermediate members between sets of elements and relative movement between elements. Thus, a connection reference does not necessarily imply that two elements are directly connected and fixed to each other. Exemplary drawings are for illustrative purposes only, and the dimensions, positions, orders, and relative sizes reflected in the accompanying drawings may vary.

[0039] As used herein, a “controller” or “controller module” can include a component configured or adapted to provide instructions, control, operation, or any form of communication to an operable component to enable its operation. A controller or controller module can include any known processor, microcontroller, or logic device, including but not limited to: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), full-authority digital engine control (FADECs), proportional controllers (P), proportional-integral controllers (PI), proportional-derivative controllers (PD), proportional-integral-derivative controllers (PID controllers), hardware-accelerated logic controllers (e.g., for encoding, decoding, code conversion, etc.), and combinations thereof. Non-limiting examples of controllers or controller modules may be configured or adapted to run, operate, or otherwise execute program code to achieve operational or functional results, including performing various methods, functionalities, processing tasks, calculations, comparisons, sensing or measuring values, etc., to enable or implement the technical operations or actions described herein. Operational or functional results may be based on one or more inputs, stored data values, sensed or measured values, true or false indications, etc. While described as "program code," non-limiting examples of operable or executable instruction sets may include routines, programs, objects, components, data structures, algorithms, etc., that have the technical effect of performing a particular task or implementing a particular abstract data type. In another non-limiting example, the controller or controller module may also include processor-accessible data storage components, including memory, whether transient, volatile, or non-transient, or non-volatile memory. Additional non-limiting examples of memory may include random access memory (RAM), read-only memory (ROM), flash memory, or one or more different types of portable electronic memory, such as discs, DVDs, CD-ROMs, flash drives, universal serial bus (USB) drives, etc., or any suitable combination of these types of memory. In one example, program code may be stored in memory in a machine-readable format accessible to the processor. Additionally, memory may store various types of data, sensed or measured data values, input, generated or processed data, etc., accessible to the processor when providing instructions, control, or operations to achieve a function or operable result, as described herein.

[0040] Furthermore, as used herein, while a sensor may be described as “sensing” or “measuring” a corresponding value, sensing or measuring may include determining a value that represents or is associated with the corresponding value, rather than directly sensing or measuring the value itself. The sensed or measured value may also be provided to additional components. For example, the value may be provided to a controller or processor, which may perform processing on the value to determine a representative value or electrical characteristic representing said value.

[0041] The controller described herein can also compare a first value with a second value and can operate or control the operation of additional components based on the satisfaction of this comparison. As used herein, the term "satisfied" or "satisfied" in the comparison between the first and second values ​​will mean determining whether the first value exceeds, does not exceed, or is equal to the second value, such that the comparison is "true" when satisfied. Furthermore, as used herein, the term "satisfied" or "satisfied" in the comparison between the first value and a value range means determining that the first value is within a value range, such that the comparison is "true" when satisfied. It should be understood that this determination can be easily changed by positive / negative comparisons or true / false comparisons. Example comparisons may include comparing a sensed or measured value with a threshold or threshold range. For example, when a sensed, measured, or provided value is compared with another value or range (including stored or predetermined values ​​or ranges), the satisfaction of this comparison can result in an action, function, or operation controllable by the controller.

[0042] Additionally, as used herein, an element described as "electrically connected," "electrically coupled," or "in signal communication" may include electrical transmissions or signals sent, received, or transmitted to or from such a connection or coupling. Furthermore, such an electrical connection or coupling may include wired or wireless connections, or combinations thereof.

[0043] As used herein, "reduction" can refer to a relative value that is less than or below the initial value being compared. Regarding signal "noise," this reduction can be quantified based on typical electrical noise measurement techniques, including but not limited to, measuring the root mean square (RMS) value of "noise" using a voltmeter, measuring the noise power spectrum at different frequencies using a spectrum analyzer, measuring the noise figure by measuring the degradation of the signal-to-noise ratio (SNR) as the signal passes through the device, and combinations thereof.

[0044] Now refer to the attached diagram, Figure 1This disclosure describes a network architecture of multiple analog edge devices (e.g., MEMS devices) in a neural network for reducing noise in or denoising input analog signals, according to aspects of this disclosure. In one aspect, the input analog signal can be an audio signal, such as speech. The neural network can be used, for example, for speech recognition (e.g., speech direction recognition). The input analog signal can also be a noisy signal, and the neural network can be used as an adaptive filter to operatively reduce or substantially remove noise from the noisy signal. The terminology used herein substantially refers to removing at least 50% of the noise (e.g., removing at least 90%) in most applications. The neural network can also be used to identify the type of one (or a set of simultaneously) input analog signals and determine the orientation (e.g., relative orientation or heading) of the input analog signal source. The input analog signal can also be, for example, a vibration signal (e.g., earth vibration), and the neural network can be used to identify whether the vibration signal is an earthquake or another type of vibration and determine the location or orientation of the vibration source. The input analog signal can be an image signal, and the neural network can be used for image recognition. Numerous applications can be implemented in the analog domain using this neural network.

[0045] Figure 1 An input signal 100 is shown being fed into a first plurality of analog edge devices 102 (e.g., a plurality of MEMS devices). The input signal 100 is, for example, an input analog signal with signal noise (e.g., a noisy audio signal). Each of the first plurality of analog edge devices 102 can be excited by the input signal 100 to convert the input signal 100 into mechanical motion. The mechanical motion in each of the first plurality of analog edge devices 102 is converted into a first output electrical signal to obtain a first plurality of output analog signals (collectively shown as 103). Each of the first plurality of analog signals 103 of each of the first plurality of analog edge devices 102 is input to an analog edge device in a second plurality of analog edge devices 104. In this aspect of the disclosure, the first plurality of analog edge devices 102 (e.g., MEMS device computational neurons) are MEMS CTRNN neurons. Figure 1In the diagram, squares represent a first plurality of analog edge devices 102 (e.g., MEMS devices), and circles represent a second plurality of analog edge devices 104 (e.g., analog neurons implemented by FPAA). In other aspects of this disclosure, the first plurality of analog edge devices 102 (e.g., MEMS device neurons) and the second plurality of analog edge devices 104 (e.g., analog neurons implemented by FPAA) may be placed in the same layer of the neural network. In another aspect of this disclosure, the first plurality of analog edge devices 102 (e.g., MEMS device neurons) in the first layer may replace the second plurality of analog edge devices 104 (e.g., analog neurons) in the second layer. Examples of analog neurons include resistors, capacitors, operational amplifiers, programmable analog memories, and other electronic device circuit elements.

[0046] For example, multiple logic signals from a first analog edge device 102A (e.g., a MEMS device) in a first plurality of analog edge devices 102 (e.g., MEMS device neurons) are input into a second plurality of analog edge devices 104 (e.g., analog neurons), and multiple logic signals from a second analog edge device 102B are input into the second plurality of analog edge devices 104 (e.g., analog neurons). Each analog edge device in the first plurality of analog edge devices 102 outputs a corresponding output analog signal 103, which is multiplied by a corresponding weighting parameter to obtain multiple weighted signals. For example, the first output analog signal 103 output by the first analog edge device 102A in the first plurality of analog edge devices 102 is multiplied by a first weighting parameter to obtain a first weighted signal, and the second output analog signal 103 output by the second analog edge device 102B in the first plurality of analog edge devices 102 is multiplied by a second weighting parameter to obtain a second weighted signal. The multiple weighted signals are input into the analog edge devices in the second plurality of analog edge devices 104. For example, a first weighted signal and a second weighted signal are input to a first analog edge device 104A in a second plurality of analog edge devices 104. The first plurality of analog edge devices 102 are in a first layer, and the second plurality of analog edge devices 104 are in a second layer.

[0047] Each of the second plurality of analog edge devices 104 outputs an analog signal 105, which is multiplied by weight parameters to obtain a plurality of weighted signals (collectively shown as 105). For example, a first output analog signal 105 output by a first analog edge device 104A in the second plurality of analog edge devices 104 is multiplied by a first weight parameter to obtain a first weighted signal, and a second output analog signal 105 output by a second analog edge device 104B in the second plurality of analog edge devices 104 is multiplied by a second weight parameter to obtain a second weighted signal. The plurality of weighted signals are input to an analog edge device 106 (e.g., a MEMS neuron). The analog edge device 106 is located within the third layer of the network. For example, the first weighted signal and the second weighted signal are input to the analog edge device 106 (e.g., a MEMS device neuron). The second plurality of analog edge devices 104 and the analog edge device 106 (e.g., a MEMS device neuron) can be implemented as an analog array. The second plurality of simulated edge devices 104 are in the second layer, and the simulated edge devices 106 (e.g., MEMS neurons) are in the subsequent third layer. Although in Figure 1 Three layers are shown, but the number of layers is not limited to two or three. Multiple layers (two or more) can be present, and the process can be repeated multiple times. Analog edge device 106 outputs an output signal 108. Output signal 108 corresponds to input signal 100, which is operatively corrected or operatively filtered to substantially remove signal noise from input signal 100. A neural network corresponding to a second plurality of analog edge devices 104 (neurons) and analog edge devices 106 (neurons) is trained to find appropriate weight parameters to multiply various signals to output output signal 108, which is substantially or operatively noise-free, or has reduced noise compared to input signal 100.

[0048] Figure 2 The electronic circuit 200 according to an aspect of this disclosure illustrates the interconnection between multiple analog edge devices (e.g., MEMS) for implementing an analog system for edge artificial intelligence using a neural network 201 and one or more field-programmable analog arrays (FPAAs). Figure 2As shown, the electronic circuit 200 includes a plurality of MEMS devices 202 (MEMS neurons). In one aspect of this disclosure, the plurality of MEMS devices 202 (MEMS neurons) can convert electrical analog inputs into mechanical signals within the plurality of MEMS devices 202, which in turn output electrical signals through analog front-end circuitry. In a non-limiting aspect of this disclosure, each of the plurality of MEMS devices 202 is electrically connected to a corresponding electronic board among a plurality of electronic boards 204. The electronic circuit 200 also includes a main electronic board 206 (e.g., a printed circuit board PCB) for electrically connecting the plurality of electronic boards 204. The electronic circuit 200 also includes one or more memory devices 208 (e.g., one or more field-programmable analog arrays (FPAAs)). The one or more memory devices 208 are also electrically connected to the main electronic board 206. The one or more memory devices 208 communicate with the plurality of MEMS devices 202 via the main electronic board 206. In one aspect of this disclosure, as Figure 2 As shown, one or more memory devices 208 in electronic circuit 200 are one or more FPAAs. While one or more FPAAs used in electronic circuit 200 can serve as at least one memory device 208, or one or more memory devices 208, one or more ferroelectric memory devices, or one or more resistive memory devices, or one or more programmable analog memory devices, etc., can also serve as one or more memory devices 208. One or more memory devices 208 communicate with multiple MEMS devices 202. One or more memory devices 208 are used to store the weight parameters discussed above. Electronic circuit 200 is powered via power line 209. Electronic circuit 200 includes multiple input ports 210 connected to a main electronic board 206. Input analog signals are input through the multiple input ports 210 and transmitted through the main electronic board 206 to the multiple MEMS devices 202 (MEMS neurons). Output signals from the multiple MEMS devices 202 are multiplied by the weight parameters stored in one or more memory devices 208 to implement a neural network, such as... Figure 1 As shown.

[0049] The electronic circuit 200 also includes an output port 212 connected to the main electronic board 206. Output signals from the neural network of multiple MEMS devices 202 are output through the output port 212. In one aspect of this disclosure, the neural network 201 may be, for example, a CTRNN. In the electronic circuit 200 for implementing the neural network, multiple MEMS devices 202 (MEMS neurons or MEMS CTRNNs) are used to implement multiple neurons in the neural network 201. For example, refer to... Figure 1 Input signal 100 (e.g., reference) Figure 1 The interpreted "noise" audio signal is input through one or more input ports 210, and the output signal (e.g.) is... Figure 1 The corrected or filtered output signal 108 is output through output port 212. Weighting parameters for multiplying various signals output by the first plurality of analog edge devices 102 (e.g., MEMS device neurons) and the second plurality of analog edge devices 104 (e.g., analog neurons) are stored in a plurality of memory devices 208 (e.g., FPAA). For example, by using a plurality of input ports 210 instead of a single input port, the directionality of the input analog signal (e.g., audio signal) can be determined by processing (e.g., comparing) the input analog signal using a neural network 201 (e.g., CTRNN).

[0050] The neural network 201 (e.g., CTRNN) can utilize multi-layered hierarchical MEMS devices 202 (MEMS neurons) and analog electronics. In one aspect of this disclosure, the multiple MEMS devices 202 (MEMS neurons) can have electrostatic excitation and capacitive output. The analog electronics 203 can be constructed using an analog front-end comprising multiple MEMS devices 202 and multiple memory devices 208 (e.g., FPAA). The multiple MEMS devices 202 (MEMS neurons) have adjustable time constants that can be controlled by voltage. Weight parameters are set in the multiple memory devices 208 (e.g., FPAA). The outputs from the multiple MEMS devices 202 (MEMS neurons) are multiplied by the weight parameters and applied to the next layer of neurons (MEMS devices or analog edge devices), as described above. Figure 1 As stated above.

[0051] The adjustable time constant of the MEMS device enables the neural network 201 (e.g., CTRNN) to learn salient information in a time series. In one aspect of this disclosure, the CTRNN can also be constructed from MEMS devices with different time constants for time learning. In a non-limiting example, seismic activity can generate different types of seismic waves traveling across the Earth at different speeds. The primary wave (“P-wave”) is the fastest and the first signal to arrive, followed by slower secondary waves (“S-waves”), and then surface waves. Non-limiting aspects of this disclosure may be included, where the MEMS neurons of an edge AI computer can adjust their time constants based on the different seismic waves being monitored by the respective MEMS neurons, as described herein.

[0052] Figure 3 This is a schematic diagram of an example MEMS device (MEMS neuron) 300 according to aspects of this disclosure. The MEMS device 300 is an electromechanical device that can be controlled by electrical input signals through various input ports by applying voltage to the MEMS device 300. For example, the MEMS device has an excitation port 302, a detection port 304, and a tuning port 306. (Reference: Back) Figure 1The input signal 100 is input through the excitation port 302, and the output analog signal 103 is output through the detection port 304. The output analog signal 103 is multiplied by the corresponding weight parameter and added to the output signals from other analog edge devices, such as the first plurality of analog edge devices, and multiplied by the corresponding weight parameter. The sum of these weighted output signals is fed to the input of the next neuron, such as 104. The MEMS neuron is characterized by a tunable time constant, which is various tuning voltages input through the tuning port 306. For example, various weight parameters and various tuning voltage parameters are stored in multiple memory devices 208 (e.g., Figure 2 In the FPAA shown, multiplication, summation, and nonlinear activation functions (e.g., sign modulus, hyperbolic tangent, ReLU (rectified linear unit)) are implemented in multiple memory devices 208 (e.g., FPAA). Other aspects of these functions may be analog units such as resistors, capacitors, operational amplifiers, etc.

[0053] Figure 4 This is a schematic diagram of a MEMS device 400 used as a neuron implementing a neural network according to an aspect of this disclosure. The MEMS device 400 includes a frame 402, one or more anchors 404, and a plurality of springs 406. The one or more anchors 404 secure the plurality of springs 406 to the frame. An electrostatic force is applied to the plurality of springs 406 by applying a voltage via an input signal. As a result, the plurality of springs 406 move, which changes the capacitance of the MEMS device 400, thereby ultimately changing the output voltage or output current of the output signal. Furthermore, the frame 402 can be moved by applying a voltage. An input signal 100 is applied via an excitation port 302, causing the frame 402 to move. In this aspect of the disclosure, the movement of the frame 402 is detected by a change in capacitance via a detection port 304. Finally, an amplifier in an analog front-end or interface circuit converts the capacitance change into an output voltage or output current of the output signal.

[0054] Figure 5 This is a graph depicting the input analog signal with noise, the output analog signal with noise reduction, and a comparison with a real ground analog signal according to aspects of this disclosure. The X-axis represents time, and the Y-axis represents voltage. Figure 5The curves shown are exemplary, where the noisy input analog signal 502 (e.g., input signal 100) includes high-frequency variation components above low-frequency variation components. The output analog signal 504 is a smoothed signal where the amplitude of the high-frequency noise components is significantly reduced. The term "substantially" as used herein means that the high-frequency components of the analog noise are reduced by at least 50% (e.g., at least 90%). The input analog signal 502 corresponds to the input signal (e.g., input signal 100) fed into the neural network. One objective is to train the neural network to fully learn the comparison between the output analog signal 504 (the denoised analog signal) and the ground-based true analog signal 506, thereby determining the optimal weighting parameters that minimize the error between the output analog signal 504 (the denoised analog signal) and the ground-based true analog signal 506. In one non-limiting example of this disclosure, the comparison of the output analog signal 504 with the ground-based real analog signal 506 can be operatively modified with weight parameters for continuous multiplication or weighting of the analog signal with the modified weight parameters, wherein the modification of the weight parameters reduces the error between the output analog signal 504 (the denoised analog signal) and the ground-based real analog signal 506. One objective of neural network inference is to process the input analog signal 502 through a network with learned weight parameters, such that the noise in the output analog signal 504 (the denoised analog signal) is significantly reduced.

[0055] The noise reduction of the input analog signal is performed using a neural network that uses multiple MEMS neurons (MEMS devices).

[0056] MEMS devices (neurons) can be modeled using the following mathematical equation (1):

[0057]

[0058] When term Z corresponds to a displacement of current and / or voltage that is converted into an analog signal associated with the output of the MEMS device, term V... i Corresponding to the input signal, w o Corresponding to the natural frequency of the MEMS device, ζ is the damping ratio, and V m It is the voltage applied to adjust the weighting parameters, where d is the electrode gap, and V is the voltage applied. b ε is the bias voltage, m is the mass, A is the exchange electrode gap, and ε is the vacuum permittivity.

[0059] By ignoring second-order terms Equation (1) can be approximated as equation (2):

[0060]

[0061] And the time delay t is determined by equation (3):

[0062]

[0063] Equation (3) shows that, in addition to other parameters, the time delay t also depends on the voltage V. m As a result, by adjusting the voltage V applied to adjust the weighting parameter... m The time delay or time constant t can be adjusted. MEMS devices have an initial set time constant determined by the design, geometry, and other physical parameters of the MEMS device's components. This initial set time constant can also vary with temperature or typical environmental conditions of the MEMS device (e.g., temperature, humidity, etc.). To provide more control over the time constant of the MEMS device, it can be adjusted or changed by applying a voltage or current to the MEMS device. Voltage V m Used to adjust the time constant. MEMS devices naturally solve equation (1) or equation (2) to provide output signals (voltage and / or current) by using input signals (voltage or current to control displacement) and control parameters (e.g., voltage or current to control time).

[0064] In one aspect of this disclosure, a CTRNN using a MEMS device can directly process input analog signals and can learn time information using the cyclic characteristics of the neural network. This cyclic characteristic is implemented through a self-feedback memory enabled by an electrostatic MEMS device.

[0065] Figure 6 This is a flowchart of a machine learning algorithm (e.g., CTRNN) using a neural network for solving ordinary differential equations, according to aspects of this disclosure. Figure 6 An ordinary differential equation solver (ODE) 600 implemented using a neural network 602 of an analog edge device (e.g., a MEMS device) is shown. The neural network 602 can implement a CTRNN model 610. The neural network 602 receives an analog signal input 604 and outputs an analog signal output 606. As described above, the time constant parameter of the analog edge device (e.g., the MEMS device) is controlled based on the environment 608 of the analog edge device.

[0066] Figure 7 This is a schematic diagram illustrating an example implementation of multi-level hierarchical differential target propagation for in-situ training of MEMS CTRNNs according to aspects of this disclosure. Figure 7As shown, a neural network 700 with a neuron layer 702 of analog edge devices (e.g., MEMS devices) is configured to receive one or more input analog signals 704, such as input signal 100. In one aspect of this disclosure, the neural network 700 is a CTRNN, which is trained in situ using one or more input analog signals 704 to provide an output analog signal 706 to an analog edge device 708 (e.g., MEMS neuron). The use of in-situ training can address two problems of past limited analog computers: (i) analog hardware variations and (ii) analog hardware drift. The neural network (e.g., CTRNN) can learn its parameters based on individual edge hardware. The neural network (e.g., CTRNN) can recalibrate itself through relearning, thereby mitigating drift caused by environmental factors (e.g., correcting environmental effects to change the time constant), such as... Figure 6 As shown.

[0067] The inference path may include a neural network 700 with neuron layers 702. Typically, the inference path uses weight parameters that include the time constant parameter set in the CTRNN. However, in the current case, the CTRNN is also trained in situ using one or more input analog signals 704. As a result, Differential Target Propagation (DTP) is used to allow updates to the weight parameters at local layers, instead of the standard backpropagation algorithm which identifies weight gradients by backpropagating the global loss error through the network. This allows for in-situ training at local layers, which greatly simplifies the analog circuit design. For example, the first layer 710A in neuron layer 710, which provides a prediction of the time constant T2, can be locally backpropagated to the last layer 702A in neuron layer 702. Similarly, the second layer 710B in neuron layer 710, which provides a prediction of the time constant T1, can be locally backpropagated to layer 702B in neuron layer 702. This feature, called Differential Target Propagation (DTP), is used to allow weight updates at local layers.

[0068] Figure 8 This is a flowchart illustrating details of the differential target propagation scheme used by the CTRNN according to aspects of this disclosure. Figure 8 As shown, the continuous-time autoregressive recurrent neural network uses training path 802 and forward path or inference path 804. Forward path 804 corresponds to... Figure 7 The neuron layer 702 shown, and the training path 802 corresponding to Figure 7The diagram shows neuron layer 710. Training path 802 has multiple target neuron layers (layer 0 and layer 1), and forward path 804 has multiple neuron layers (layer 0, layer 1, and layer 2). Each neuron layer is controlled by inputting weights using a weight estimator. Instead of the standard backpropagation algorithm that identifies weight gradients by backpropagating the global loss error through a neural network (e.g., a CTRNN), weights are updated in real-time and in situ at local layers (layer 0, layer 1, layer 2, etc.). This allows for in-situ training at local layers, which greatly simplifies analog circuit design.

[0069] Figure 8 A non-limiting example of aspects of this disclosure is described, such as the Differential Target Propagation (DTP) technique. The DTP method allows computation or generation of a target for each neuron in a network, and the estimation of weights using a weight estimator local to each neuron. This is done by constructing an additional neural network (NN) layer called the target layer, which is arranged as a "mirror" of the inference NN, such as... Figure 8 The dashed box 802 is shown inside. The inferred path and the target signal of layer N are used to generate the target for layer N-1, which in turn is used to generate the target for layer N-2, and so on. At each layer of the network, each weight estimator belonging to a neuron receives the target signal and then subtracts it from the neuron's output. The result of the subtraction is then multiplied by one of the inputs, integrated, and produces the weight value that needs to be applied to that given input.

[0070] In one aspect of this disclosure, differential objective propagation in a CTRNN can be used to update the weight parameters. The CTRNN uses an autoencoder-type architecture to construct intermediate objectives. Weight updates are based on local layer information to provide in-situ training to mitigate temporal drift across multiple simulated edge devices in the CTRNN.

[0071] A MEMS analog edge device is provided that enables edge artificial intelligence (AI) computation with relatively low power and relatively high speed (e.g., hundreds of times lower power and hundreds of times faster speed) compared to existing digital computing architectures. The term "edge" is used throughout this paper to refer to computing devices located at the edge or end of a network architecture near the user, such as cellular phones, sensors, or other network-connected devices. Analog computing is particularly powerful for edge AI applications where size, weight, and power-as-a-P (SWaP) play a significant role. Furthermore, analog computing is particularly powerful for edge AI applications where most of the collected data is generated by analog edge devices (e.g., analog sensors). By performing AI computation in the analog domain, the need for bidirectional analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) can be eliminated. Moreover, analog continuous-time processing does not require constant data shifting in and out of memory, addressing one of the major bottlenecks of digital architectures. As a result, by operating in the analog domain, fast (e.g., real-time, virtually no latency), ultra-low power (e.g., mW to μW), and low-cost AI computation can be achieved at the edge. Furthermore, in-situ training allows edge AI systems to mitigate drift and variations in analog hardware. The actual deployment of edge simulation hardware may require in-situ learning.

[0072] The simulation system described herein can be configured to perform AI computations using natural physical phenomena as computational primitives without analog-to-digital conversion. The simulation system can be configured to perform AI computations at least two orders of magnitude faster and with at least two orders of magnitude lower energy consumption compared to conventional digital techniques.

[0073] In one aspect of this disclosure, the simulated edge computing device can be a MEMS device. Compared to a digital workstation solving the same AI problem, the MEMS device exhibits a power reduction of more than 300 times and a speed increase of more than 100 times.

[0074] For example, MEMS devices used to implement machine learning (ML) algorithms (e.g., CTRNN) can directly process input analog signals and can learn temporal information using the cyclic features of neural networks. CTRNN is a deep learning model that interprets irregular observations by introducing a continuous evolution of hidden states between observations. This cyclic feature is achieved through a self-feedback memory enabled by an electrostatic MEMS device.

[0075] In one aspect of this disclosure, the MEMS device (e.g., a MEMS processor) is controlled by electrostatic excitation to provide a capacitive output. Analog electronics for the MEMS device can be constructed using an analog front-end and a field-programmable analog array (FPAA). The MEMS device has an adjustable time constant that can be controlled by voltage. Weighting parameters are set in the FPAA. The output from the MEMS device is multiplied by the weighting parameters and applied to the next neuron layer (MEMS or analog).

[0076] In one aspect of this disclosure, the adjustable time constant of the MEMS device enables the MEMS-CTRNN to learn salient information in a time series. The CTRNN can also be constructed from MEMS devices with different time constants for time learning.

[0077] In one aspect of this disclosure, differential target propagation (DPT) can be used to update weights at local layers, instead of the standard backpropagation algorithm which identifies weight gradients by propagating the global loss error back through the network. This allows for in-situ training at local layers, which greatly simplifies analog circuit design.

[0078] In one aspect of this disclosure, in-situ training is used to address two problems of the limited analog computers of the past: (1) simulating hardware changes, wherein this technique allows the neural network to learn its parameters based on separate edge hardware, and (2) simulating hardware drift, wherein this technique allows the neural network to recalibrate itself by relearning, thereby mitigating drift caused by the environment.

[0079] Another issue of particular interest is in-situ training for edge applications. To date, the analog computing community has largely neglected this problem. Traditional stochastic gradient descent (SGD) learning methods rely on backpropagation of errors through multi-layered neural networks, which is difficult to implement in analog hardware. Analog edge devices are indeed different from one another, and their characteristics can drift over time. In-situ training allows each component in an edge AI system to learn its parameters based on individual edge hardware. This also allows the edge AI system to recalibrate itself through relearning, thus mitigating the drift of analog hardware.

[0080] By performing AI computations in the analog domain, the need for bidirectional analog-to-digital conversion can be eliminated. Analog, continuous-time processing eliminates the need for constant data loading and unloading from memory, addressing one of the major bottlenecks in digital architectures. This enables fast (real-time, no latency), ultra-low-power (mW to μW), and low-cost AI computations at the edge. Analog computation is particularly powerful for edge AI applications that require small size, weights, and power-to-weight ratios (SWaP) and where most of the collected data is generated by analog sensors.

[0081] Therefore, one aspect of this disclosure provides a MEMS continuous-time recurrent neural network (CTRNN) that utilizes electrostatic MEMS neurons to perform network computation in continuous time and the analog domain. Another aspect of this disclosure provides in-situ supervised training of the CTRNN, which allows learning directly on edge hardware using received analog information and self-correction for variations and drift in the analog hardware.

[0082] This enables low-SWaP, fast-response edge AI for a wide range of applications both within and outside the aerospace industry. Applications can include signal filtering (adaptive), prediction, tracking, classification, or anomaly detection, or decoding information based on continuous signals. Edge AI can be deployed in autonomous platforms that require SWaP, such as UAVs, autonomous vehicles, and robots.

[0083] Further aspects are provided by the following items.

[0084] A simulation system for edge artificial intelligence (AI) computing includes: an input port configured to receive an input analog signal; a first plurality of analog edge devices connected to the input port and configured to receive the input analog signal and output a first plurality of output analog signals; a second plurality of analog edge devices communicating with the first plurality of analog edge devices and configured to receive the first plurality of output analog signals from the first plurality of analog edge devices and output a second plurality of output analog signals; at least one memory device communicating with the first plurality of analog edge devices and the second plurality of analog edge devices, the at least one memory device configured to store weight parameters adjustable based on at least one of: predetermined values ​​from offline training, or based on real-time updates through in-situ training; and an output port configured to output an output analog signal, wherein the second plurality of output analog signals are multiplied by the weight parameters to obtain a plurality of weighted analog signals, and the plurality of weighted analog signals are output as the output analog signal through the output port.

[0085] According to any of the preceding clauses of the simulation system, the at least one memory device includes one or more field-programmable analog arrays (FPAAs) or one or more programmable analog memories.

[0086] According to any of the preceding clauses of the simulation system, wherein the first plurality of analog edge devices and the second plurality of analog edge devices are trained to find the weight parameters, thereby operatively reducing the signal noise of the output analog signal.

[0087] The simulation system according to any of the foregoing clauses further includes: a plurality of electronic boards, each of the first plurality of analog edge devices and the second plurality of analog edge devices being connected to a corresponding electronic board among the plurality of electronic boards; and a main electronic board, wherein the plurality of electronic boards and the at least one memory device are electrically connected to the main electronic board.

[0088] According to any of the preceding clauses, the analog system wherein the input analog signal has signal noise, and the output analog signal corresponds to the input analog signal and the signal noise is significantly reduced.

[0089] In any of the preceding clauses, the analog input signal is an audio signal.

[0090] The simulation system according to any of the foregoing clauses, wherein the simulation system is configured to perform AI computations using natural physical phenomena as computational primitives without analog-to-digital conversion, and wherein the simulation system is configured to perform AI computations at least two orders of magnitude faster and at least two orders of magnitude lower in energy consumption compared to conventional digital techniques.

[0091] According to any of the preceding clauses, the first plurality of analog edge devices or the second plurality of analog edge devices or both comprise a plurality of microelectromechanical systems (MEMS) devices.

[0092] According to any of the preceding clauses, in the simulation system, the plurality of MEMS devices are configured to convert the input analog signal into a mechanical signal and the mechanical signal into an output analog signal, wherein each of the plurality of MEMS devices has an adjustable time constant that can be controlled by voltage.

[0093] According to any of the preceding clauses, the first plurality of simulated edge devices, the second plurality of simulated edge devices, and the at least one memory device are part of a neural network, wherein each of the first plurality of simulated edge devices and each of the second plurality of simulated edge devices is a neuron in the neural network.

[0094] According to any of the preceding clauses, the simulation system wherein the neural network is a continuous-time recurrent neural network (CTRNN).

[0095] According to any of the preceding clauses, in the simulation system, the first plurality of simulated edge devices and the second plurality of simulated edge devices of the CTRNN have different time constants to perform time learning and adjust the time constants of the different time constants.

[0096] According to the simulation system described in the foregoing clause, the weight parameters are updated using differential target propagation in the CTRNN, the differential target propagation using an autoencoder-type architecture to establish intermediate targets, and the weight parameters are updated based on local layer information to provide in-situ training of the CTRNN, thereby mitigating temporal drift in the first plurality of simulated edge devices and the second plurality of simulated edge devices of the CTRNN.

[0097] According to any of the preceding clauses, the in-situ training of the CTRNN allows the CTRNN to learn its parameters based on a separate simulated edge device.

[0098] According to any of the preceding clauses of the simulation system, the in-situ training of the CTRNN allows the CTRNN to recalibrate itself through relearning, thereby mitigating simulation hardware drift caused by environmental factors.

[0099] According to any of the preceding clauses of the simulation system, the simulation drift caused by the environmental factors includes the change in the time constant caused by the environmental factors.

[0100] A method for processing analog signals, the method comprising: receiving an input analog signal at an input port; receiving the input analog signal via a first plurality of analog edge devices connected to the input port; outputting a first plurality of output analog signals via the first plurality of analog edge devices; receiving the first plurality of output analog signals from the first plurality of analog edge devices via a second plurality of analog edge devices; outputting a second plurality of output analog signals via the second plurality of analog edge devices; multiplying the second plurality of output analog signals with weight parameters stored on at least one memory device communicating with the first plurality of analog edge devices and the second plurality of analog edge devices to obtain a plurality of weighted analog signals, the weight parameters being adjustable based on at least one of: predetermined values ​​from offline training, or based on real-time updates from in-situ training; outputting the plurality of weighted analog signals as output analog signals via an output port; and comparing the output analog signals with ground-based real analog signals to modify the weight parameters for continuous multiplication, wherein the modified weight parameters reduce the error between the output analog signals and the ground-based real analog signals.

[0101] According to any of the foregoing provisions of the method, wherein the input analog signal has signal noise, and the output analog signal corresponds to the input analog signal and the signal noise is significantly reduced.

[0102] The method according to any of the foregoing clauses further includes applying a voltage to control the time constant of the first plurality of analog edge devices or the second plurality of analog edge devices or both.

[0103] The method according to any of the foregoing clauses further includes using differential target propagation to update the weight parameters in a continuous time recurrent neural network (CTRNN) including the first plurality of simulated edge devices, the second plurality of simulated edge devices and the at least one memory device, and using local layer information based on differential target propagation to identify the weight parameters in order to provide in-situ training to mitigate temporal drift in the first plurality of simulated edge devices and the second plurality of simulated edge devices of the CTRNN.

[0104] While the foregoing description is directed to preferred embodiments of the present disclosure, other variations and modifications will be apparent to those skilled in the art and can be made without departing from the present disclosure. Furthermore, features described in connection with one embodiment of the present disclosure may be used in conjunction with other embodiments, even if not expressly stated above.

Claims

1. A simulation system for edge artificial intelligence (AI) computing, characterized in that, The simulation system includes: An input port, configured to receive an input analog signal; A plurality of analog edge devices are connected to the input port and configured to receive the input analog signal and output a plurality of output analog signals; A second plurality of analog edge devices, which communicate with the first plurality of analog edge devices, are configured to receive the first plurality of output analog signals from the first plurality of analog edge devices and output a second plurality of output analog signals. At least one memory device, communicating with the first plurality of simulated edge devices and the second plurality of simulated edge devices, is configured to store weight parameters that can be adjusted based on at least one of: predetermined values ​​from offline training, or based on real-time updates via in-situ training; and An output port is configured to output an analog signal, wherein the second plurality of output analog signals are multiplied by the weighting parameters to obtain a plurality of weighted analog signals, and the plurality of weighted analog signals are output as the output analog signal through the output port.

2. The simulation system according to claim 1, characterized in that, The at least one memory device includes one or more field-programmable analog arrays (FPAAs) or one or more programmable analog memories.

3. The simulation system according to claim 1, characterized in that, The first plurality of analog edge devices and the second plurality of analog edge devices are trained to find the weight parameters, thereby operatively reducing the signal noise of the output analog signal.

4. The simulation system according to claim 1, characterized in that, Further includes: Multiple electronic boards, each of the first plurality of analog edge devices and the second plurality of analog edge devices being connected to a corresponding electronic board among the plurality of electronic boards; and A main electronic board, wherein the plurality of electronic boards and the at least one memory device are electrically connected to the main electronic board.

5. The simulation system according to claim 1, characterized in that, The input analog signal has signal noise, and the output analog signal corresponds to the input analog signal with significantly reduced signal noise.

6. The simulation system according to claim 1, characterized in that, The input analog signal is an audio signal.

7. The simulation system according to claim 1, characterized in that, The simulation system is configured to perform AI computations using natural physical phenomena as computational primitives without analog-to-digital conversion, and the simulation system is configured to perform AI computations at least two orders of magnitude faster and with at least two orders of magnitude lower energy consumption compared to conventional digital technologies.

8. The simulation system according to claim 1, characterized in that, The first plurality of analog edge devices or the second plurality of analog edge devices or both include a plurality of microelectromechanical systems (MEMS) devices.

9. The simulation system according to claim 8, characterized in that, The plurality of MEMS devices are configured to convert the input analog signal into a mechanical signal and the mechanical signal into an output analog signal, wherein each of the plurality of MEMS devices has an adjustable time constant that can be controlled by voltage.

10. A method for processing analog signals, characterized in that, The method includes: Receives input analog signals at the input port; The input analog signal is received by a first plurality of analog edge devices connected to the input port; The first plurality of analog output signals are output through the first plurality of analog edge devices; The first plurality of output analog signals are received from the first plurality of analog edge devices through the second plurality of analog edge devices; The second plurality of analog output signals are output through the second plurality of analog edge devices; The second plurality of output analog signals are multiplied with weight parameters stored on at least one memory device communicating with the first plurality of analog edge devices and the second plurality of analog edge devices to obtain a plurality of weighted analog signals, the weight parameters being adjustable based on at least one of: predetermined values ​​from offline training, or based on real-time updates from in-situ training; The multiple weighted analog signals are output as output analog signals through the output port; and The output analog signal is compared with the ground-based real analog signal to modify the weighting parameters for continuous multiplication, wherein the modified weighting parameters reduce the error between the output analog signal and the ground-based real analog signal.

Images