Synaptic Weights Research Articles

Synchronized stepwise control of firing and learning thresholds in a spiking randomly connected neural network toward hardware implementation

Spiking randomly connected neural network (RNN) hardware is promising as ultimately low power devices for temporal data processing at the edge. Although the potential of RNNs for temporal data processing has been demonstrated, randomness of the network architecture often causes performance degradation. To mitigate such degradation, self-organization mechanism using intrinsic plasticity (IP) and synaptic plasticity (SP) should be implemented in the spiking RNN. Therefore, we propose hardware-oriented models of these functions. To implement the function of IP, a variable firing threshold is introduced to each excitatory neuron in the RNN that changes stepwise in accordance with its activity. We also define other thresholds for SP that synchronize with the firing threshold, which determine the direction of stepwise synaptic update that is executed on receiving a pre-synaptic spike. To discuss the effectiveness of our model, we perform simulations of temporal data learning and anomaly detection using publicly available electrocardiograms (ECGs) with a spiking RNN. We observe that the spiking RNN with our IP and SP models realizes the true positive rate of 1 with the false positive rate being suppressed at 0 successfully, which does not occur otherwise. Furthermore, we find that these thresholds as well as the synaptic weights can be reduced to binary if the RNN architecture is appropriately designed. This contributes to minimization of the circuit of the neuronal system having IP and SP.

Efficient High Heating Value estimation using Latin Hypercube Sampling and Artificial Neural Network-based approach.

To maximize energy recovery in waste-to-energy (WTE) systems, the High Heating Value (HHV) of municipal solid waste (MSW) must be accurately estimated. To forecast the HHV of MSW, this study proposes a unique method that combines an Artificial Neural Network (ANN) model with Latin Hypercube Sampling (LHS), with a focus on Solan City, Himachal Pradesh, India. In the present study, the elemental characteristics of waste have been used to deal with uncertainty and to find the suitable parameters responsible for the HHV of the MSW. Initially, Latin Hypercube Sampling (LHS) has been used to deal with uncertainty in the elemental composition of MSW, which includes carbon (C), hydrogen (H), nitrogen (N), sulfur (S), and oxygen (O) content. This elemental composition has been used as input parameters to the ANN model for predicting the HHV of MSW. The network 5-28-5-1 offered a minimum MAPE value of 2.18%, MSE, RMSE and R2 values are 0.012, 0.107 and 0.767, respectively. Thereafter, a synaptic weight approach was used to find the most significant parameters responsible for HHV in MSW. It was observed that carbon is the most suitable parameter for HHV of MSW. By dealing with the uncertainty in MSW characteristics, the integration of LHS strengthens the robustness of the model. The results offer an accurate and economical approach for HHV estimation, which will be useful for improving the MSW management and WTE conversion procedures.

CdSe quantum dots photoelectric memristors for simulating biological visual system behavior

The visual system is the most important in the biological neurosensory system, so simulation of the biological visual system is the key for developing artificial nervous system. This work incorporates a CdSe quantum dot layer into silk fibroin based memristor device, endowing the devices with light-responsive capability. Furthermore, interesting phenomena were observed when the device operated as artificial synapses: excessive voltage stimulation led to a reduction in synaptic weight compared to the responses observed under normal electrical stimulation. This behavior mirrors sensations associated with pain, neuroprotection, and potential injuries to the neural system. At the end, we designed a visual perception system, simulating processes of the biological light intensity perception and the visual degradation response under intense light stimulation. Our research demonstrates the feasibility of constructing an artificial visual nervous system using a hardware system based on memristors.

Tailoring enzyme modified starch through high-moisture extrusion: Unraveling structure-property relationships

The conventional method of producing enzyme modified starch (ES) involves using a dilute starch suspension, which necessitated substantial heat expenditure for drying the product and resulted in significant energy waste. Improved extrusion cooking technology (IECT) could extrude materials under high moisture, and it is a new physical modification technology. A comprehensive investigation was undertaken to produce enzyme modified starch (ES) with varying dextrose equivalent (DE) using IECT. The study systematically explored the effects of different screw speeds on the resulting IECT-ES properties. And principal component analysis and artificial neural network analysis were used to explore the correlation between data and the synaptic weight relationships. Results indicated that Maltese cross of IECT-ES disappeared, the R1047/1022 and R995/1022 decreased to 0.7305 and 0.9012 respectively, the full width at half maximum at 480 cm−1 increased to 26.81 %, light transmittance increased to 21.20 % and DE value rose to 14.19 % at 300 rpm. The properties evolution of IECT-ES was closely related to changes in structure and conditions. This study aimed to evaluate the effectiveness of IECT as an alternative method for producing enzyme modified starch. The findings provide valuable insights into more energy-efficient starch modification processes.

Dendritic spines of layer 5 pyramidal neurons of the aging somatosensory cortex exhibit reduced volumetric remodeling.

Impairments in synaptic dynamics and stability are observed both in neurodegenerative disorders and in the healthy aging cortex, which exhibits elevated dendritic spine turnover and decreased long-term stability of excitatory connections at baseline, as well as an altered response to plasticity induction. In addition to the discrete gain and loss of synapses, spines also change in size and strength both during learning and in the absence of neural activity, and synaptic volume has been associated with stability and incorporation into memory traces. Furthermore, intrinsic dynamics, an apparently stochastic component of spine volume changes, may serve as a homeostatic mechanism to prevent stabilization of superfluous connections. However, the effects of age on modulation of synaptic weights remain unknown. Using two-photon excitation (2PE) microscopy of spines during chemical plasticity induction in vitro and analyzing longitudinal in vivo 2PE images after a plasticity-inducing manipulation, we characterize the effects of age on volumetric changes of spines of the apical tuft of layer 5 pyramidal neurons of mouse primary somatosensory cortex. Aged mice exhibit decreased volumetric volatility and delayed rearrangement of synaptic weights of persistent connections, as well as greater susceptibility to spine shrinkage in response to chemical long-term depression. These results suggest a deficit in the aging brain's ability to fine-tune synaptic weights to properly incorporate and retain novel memories. This research provides the first evidence of alterations in spine volumetric dynamics in healthy aging and may support a model of impaired processing and learning in the aged somatosensory system.Significance Statement Aging is known to impact cognitive functions and sensory processing, yet the underlying mechanisms at the synaptic level remain unclear. This study investigates dendritic spine dynamics of layer 5 pyramidal neurons in the aging somatosensory cortex. By employing two-photon excitation microscopy and analyzing spine volume changes during plasticity induction, we reveal that aged mice exhibit decreased volumetric volatility and delayed synaptic weight rearrangement. These alterations may impair the brain's ability to fine-tune synaptic weights, which is crucial for incorporating and retaining new memories. This research provides the first evidence of altered spine volumetric dynamics in healthy aging, suggesting a potential mechanism for impaired processing and learning in the aged cortex. Understanding these changes could aid in mitigation of age-related cognitive decline.

Dynamic Analysis and FPGA Implementation of Fractional-Order Hopfield Networks with Memristive Synapse

Memristors have become important components in artificial synapses due to their ability to emulate the information transmission and memory functions of biological synapses. Unlike their biological counterparts, which adjust synaptic weights, memristor-based artificial synapses operate by altering conductance or resistance, making them useful for enhancing the processing capacity and storage capabilities of neural networks. When integrated into systems like Hopfield neural networks, memristors enable the study of complex dynamic behaviors, such as chaos and multistability. Moreover, fractional calculus is significant for their ability to model memory effects, enabling more accurate simulations of complex systems. Fractional-order Hopfield networks, in particular, exhibit chaotic and multistable behaviors not found in integer-order models. By combining memristors with fractional-order Hopfield neural networks, these systems offer the possibility of investigating different dynamic phenomena in artificial neural networks. This study investigates the dynamical behavior of a fractional-order Hopfield neural network (HNN) incorporating a memristor with a piecewise segment function in one of its synapses, highlighting the impact of fractional-order derivatives and memristive synapses on the stability, robustness, and dynamic complexity of the system. Using a network of four neurons as a case study, it is demonstrated that the memristive fractional-order HNN exhibits multistability, coexisting chaotic attractors, and coexisting limit cycles. Through spectral entropy analysis, the regions in the initial condition space that display varying degrees of complexity are mapped, highlighting those areas where the chaotic series approach a pseudo-random sequence of numbers. Finally, the proposed fractional-order memristive HNN is implemented on a Field-Programmable Gate Array (FPGA), demonstrating the feasibility of real-time hardware realization.

Linearly programmable two-dimensional halide perovskite memristor arrays for neuromorphic computing.

The exotic properties of three-dimensional halide perovskites, such as mixed ionic-electronic conductivity and feasible ion migration, have enabled them to challenge traditional memristive materials. However, the poor moisture stability and difficulty in controlling ion transport due to their polycrystalline nature have hindered their use as a neuromorphic hardware. Recently, two-dimensional (2D) halide perovskites have emerged as promising artificial synapses owing to their phase versatility, microstructural anisotropy in electrical and optoelectronic properties, and excellent moisture resistance. However, their asymmetrical and nonlinear conductance changes still limit the efficiency of training and accuracy of inference. Here we achieve highly linear and symmetrical conductance changes in Dion-Jacobson 2D perovskites. We further build a 7 × 7 crossbar array based on analogue perovskite synapses, achieving a high device yield, low variation with synaptic weight storing capability, multi-level analogue states with long retention, and moisture stability over 7 months. We explore the potential of such devices in large-scale image inference via simulations and show an accuracy within 0.08% of the theoretical limit. The excellent device performance is attributed to the elimination of gaps between inorganic layers, allowing the halide vacancies to migrate homogeneously regardless of grain boundaries. This was confirmed by first-principles calculations and experimental analysis.

A Dual‐Modal Memory Organic Electrochemical Transistor Implementation for Reservoir Computing

Neuromorphic computing devices offer promising solutions for next‐generation computing hardware, addressing the high throughput data processing demands of artificial intelligence applications through brain‐mimicking non‐von Neumann architecture. Herein, PEDOT:Tos/PTHF‐based organic electrochemical transistors (OECTs) with dual‐modal memory functions—both short‐term and long‐term—are demonstrated. By characterizing memory levels and relaxation times, the device has been efficiently manipulated and switched between the two modes through coupled control of pulse voltage and duration. Both short‐term and long‐term memory functions are integrated within the same device, enabling its use as artificial neurons for the reservoir unit and synapses in the readout layer to build up a reservoir computing (RC) system. The performance of the dynamic neuron and synaptic weight update are benchmarked with classification tasks on hand‐written digit images, respectively, both attaining accuracies above 90%. Furthermore, by modulating the device as both reservoir mode and synaptic mode, a full‐OECT RC system capable of distinguishing electromyography signals of hand gestures is demonstrated. These results highlight the potential of simplified, homogeneous integration of dual‐modal OECTs to form brain‐like computing hardware systems for efficient biological signal processing across a broad range of applications.

IMPLEMENTATION OF BACKPROPAGATION AND HYBRID ARIMA-NN METHODS IN PREDICTING ACCURACY LEVELS OF RAINFALL IN MAKASSAR CITY

Hybrid ARIMA-NN is a combined approach of the ARIMA model used to capture linear patterns in time series data and Artificial Neural Networks (ANN) to handle non-linear and stochastic patterns. Using a gradient descent algorithm, backpropagation adjusts synaptic weights based on the error between the network's prediction and actual training data values. In this study, a comparison was made between the Backpropagation method and Hybrid ARIMA-NN in forecasting rainfall in Makassar City. Rainfall data in Makassar City uses data from the rainfall measuring station at the Paotere Maritime Meteorological Station in Makassar. The activation functions used are ReLU and Leaky ReLU with epoch parameters set at 350, and learning rates of 0.01, 0.001, 0.0001, and 0.00001. The two best methods selected for further evaluation are Backpropagation with architecture 12-32-16-8-1 and Hybrid ARIMA-NN (ARIMA [4,0,1]-NN 12-256-128-64-1). The ARIMA model (4,0,1) with AIC values of 1303.4 and RMSE 162,369 is the best compared to other models, which aligns with the advantages of backpropagation architecture. The results showed that the Backpropagation method excelled with an RMSE value of 137.320 or 0.1149, indicating high accuracy in forecasting changes in seasonal trends and patterns. Hybrid ARIMA-NN gives good results with RMSE 145.834, as residues contain better nonlinearity compared to ARIMA models (4,0,1), although it shows a slightly higher error rate compared to Backpropagation.

Inhibiting the current spikes within the channel layer of LiCoO2-based three-terminal synaptic transistors

Synaptic transistors, which emulate the behavior of biological synapses, play a vital role in information processing and storage in neuromorphic systems. However, the occurrence of excessive current spikes during the updating of synaptic weight poses challenges to the stability, accuracy, and power consumption of synaptic transistors. In this work, we experimentally investigate the main factors for the generation of current spikes in the three-terminal synaptic transistors that use LiCoO2 (LCO), a mixed ionic-electronic conductor, as the channel layer. Kelvin probe force microscopy and impedance testing results reveal that ion migration and adsorption at the drain–source-channel interface cause the current spikes that compromise the device's performance. By controlling the crystal orientation of the LCO channel layer to impede the in-plane migration of lithium ions, we show that the LCO channel layer with the (104) preferred orientation can effectively suppress both the peak current and power consumption in the synaptic transistors. Our study provides a unique insight into controlling the crystallographic orientation for the design of high-speed, high-robustness, and low-power consumption nano-memristor devices.

Analog‐Digital Hybridity of Resistive Switching in Ion‐Irradiated BiFeO3 Memristor for Synergistic Neuromorphic Functionality and Artificial Learning

AbstractMemristors‐based neuromorphic devices represent emerging computing architectures to perform complex tasks by outpacing the traditional Von‐Neumann architectures in terms of speed, and energy efficiency. In this work, the resistive switching (RS) behavior of sol‐gel grown and ion‐irradiated BFO films is investigated under electrical stimulus. The Ag/BFO/FTO memristors emulate a combination of digital and analog RS behavior within a single device. The possible mechanism of analog digital hybridity is addressed by considering the formation of the conducting filament by oxygen vacancies, Ag+ ions and Schottky barrier height modulation. The ion‐irradiated BFO samples are analyzed using the Raman, XRD, and XPS studies. To uphold bioinspired synaptic actions, crucial synaptic functionalities like pair‐pulse facilitation and long‐term potentiation/depression are effectively achieved. More intricate synaptic behaviors are also demonstrated such as spike‐time‐dependent plasticity and Pavlovian classical conditioning, which represent the prominent attributes of both learning and forgetting behavior. Additionally, high pattern recognition accuracy (96.1%) is achieved in an artificial neural network simulation by using the synaptic weights of the memristors. This synergistic effect of digital and analog RS in ion‐irradiated BFO can be beneficial for the emulation of complex learning behavior as well as its incorporation into low‐power neuromorphic computing.

Computation with Sequences of Assemblies in a Model of the Brain.

Even as machine learning exceeds human-level performance on many applications, the generality, robustness, and rapidity of the brain's learning capabilities remain unmatched. How cognition arises from neural activity is the central open question in neuroscience, inextricable from the study of intelligence itself. A simple formal model of neural activity was proposed in Papadimitriou etal. (2020) and has been subsequently shown, through both mathematical proofs and simulations, to be capable of implementing certain simple cognitive operations via the creation and manipulation of assemblies of neurons. However, many intelligent behaviors rely on the ability to recognize, store, and manipulate temporal sequences of stimuli (planning, language, navigation, to list a few). Here we show that in the same model, sequential precedence can be captured naturally through synaptic weights and plasticity, and, as a result, a range of computations on sequences of assemblies can be carried out. In particular, repeated presentation of a sequence of stimuli leads to the memorization of the sequence through corresponding neural assemblies: upon future presentation of any stimulus in the sequence, the corresponding assembly and its subsequent ones will be activated, one after the other, until the end of the sequence. If the stimulus sequence is presented to two brain areas simultaneously, a scaffolded representation is created, resulting in more efficient memorization and recall, in agreement with cognitive experiments. Finally, we show that any finite state machine can be learned in a similar way, through the presentation of appropriate patterns of sequences. Through an extension of this mechanism, the model can be shown to be capable of universal computation. Taken together, these results provide a concrete hypothesis for the basis of the brain's remarkable abilities to compute and learn, with sequences playing a vital role.

Synaptic weight dynamics underlying memory consolidation: Implications for learning rules, circuit organization, and circuit function

Systems consolidation is a common feature of learning and memory systems, in which a long-term memory initially stored in one brain region becomes persistently stored in another region. We studied the dynamics of systems consolidation in simple circuit architectures with two sites of plasticity, one in an early-learning and one in a late-learning brain area. We show that the synaptic dynamics of the circuit during consolidation of an analog memory can be understood as a temporal integration process, by which transient changes in activity driven by plasticity in the early-learning area are accumulated into persistent synaptic changes at the late-learning site. This simple principle naturally leads to a speed-accuracy tradeoff in systems consolidation and provides insight into how the circuit mitigates the stability-plasticity dilemma of storing new memories while preserving core features of older ones. Furthermore, it imposes two constraints on the circuit. First, the plasticity rule at the late-learning site must stably support a continuum of possible outputs for a given input. We show that this is readily achieved by heterosynaptic but not standard Hebbian rules. Second, to turn off the consolidation process and prevent erroneous changes at the late-learning site, neural activity in the early-learning area must be reset to its baseline activity. We provide two biologically plausible implementations for this reset that propose functional roles in stabilizing consolidation for core elements of the cerebellar circuit.

Adaptive Synaptic Adjustment Mechanism to Improve Learning Performances of Spiking Neural Networks

ABSTRACTSpiking Neural Networks (SNNs) are currently attracting researchers' attention due to their efficiencies in various tasks. Spike‐timing‐dependent plasticity (STDP) is an unsupervised learning process that utilizes bio‐plausibility based on the relative timing of pre/post‐synaptic spikes of neurons. Integrated with STDP, SNNs perform well consuming less energy. However, it is hard to ensure that synaptic weights always converge to values guaranteeing accurate prediction because STDP does not change synaptic weights with supervision. To address this limitation, researchers have proposed mechanisms for inducing STDP to converge synaptic weights on the proper values referring to current synaptic weights. Thus, if the current weights fail to describe proper synaptic connections, they cannot induce STDP to update synaptic weights properly. To solve this problem, we propose an adaptive mechanism that helps STDP to converge synaptic weights directly based on input data features: Adaptive synaptic template (AST). AST leads synaptic weights to describe synaptic connections according to the data features. It prevents STDP from changing synaptic weights based on abnormal weights that fail to describe the proper synaptic connections. This is because it does not use the current synaptic weights that can disturb proper weight convergence. We integrate AST with an SNN and conduct experiments to compare it with a baseline (the SNN without AST) and benchmarks (previous works to improve STDP). Our experimental results show that the SNN using AST classifies various data sets with 6%–39% higher accuracy than the baseline and benchmarks.

GeTe/MoTe2 Van der Waals Heterostructures: Enabling Ultralow Voltage Memristors for Nonvolatile Memory and Neuromorphic Computing Applications.

Advanced electronic semiconducting Van der Waals heterostructures (HSs) are promising candidates for exploring next-generation nanoelectronics owing to their exceptional electronic properties, which present the possibility of extending their functionalities to diverse potential applications. In this study, GeTe/MoTe2 HS are explored for nonvolatile memory and neuromorphic-computing applications. Sputter-deposited Ag/GeTe/MoTe2/Pt HS cross-point devices are fabricated, and they demonstrate memristor behavior at ultralow switching voltages (VSET: 0.15V and VRESET: -0.14V) with very low energy consumption (≈30 nJ), high memory window, long retention time (104 s), and excellent endurance (105 cycles). Resistive switching is achieved by adjusting the interface between the Ag top electrode and the heterojunction switching layer. Cross-sectional transmission electron microscope images and conductive atomic force microscopy analysis confirm the presence of a conducting filament in the heterojunction switching layer. Further, emulating various synaptic functions of a biological synapse reveals that GeTe/MoTe2 HS can be utilized for energy-efficient neuromorphic-computing applications. A multilayer perceptron is implemented using the synaptic weights of the Ag/GeTe/MoTe2/Pt HS device, revealing high pattern accuracy (81.3%). These results indicate that HS devices can be considered a potential solution for high-density memory and artificial intelligence applications.

Ion-migration-induced synaptic plasticity in high ion-storable fluorine graphdiyne for information processing

Neuromorphic devices encounter constraints associated with complementary metal–oxide–semiconductor technology, such as intricate electronics and high power consumption. Recently, solid polymer electrolytes (SPE) have gained attraction in electronic devices for constructing artificial synapses, primarily due to their low-voltage operation enabled by ion-assisted signal transmission. This study presents a two-terminal artificial synapse fabricated using fluorine graphdiyne (F-GDY) synthesized through chemical vapor deposition and ion-doped SPE. The device architecture consists of Au/SPE/F-GDY/Si, which operates by ion migration induced by external potential pulses. Electrical measurements are conducted to depict the synaptic plasticity of the devices. Among devices employing different ion-doped SPEs (Li-SPE, Na-SPE, K-SPE, and Ca-SPE), the one utilizing Li-SPE demonstrates the highest synaptic weight. This is attributed to the superior Li+ ion storage capacity of F-GDY; hence, Li-SPE/F-GDY was chosen for in-depth examinations of synaptic activity presenting long-term potentiation, short-term depression, and pair-pulse facilitation. The device has good durability and stability and can respond to low action potential with power consumption at the pJ level. Furthermore, Modified National Institute of Standards and Technology simulations reveal its learning and pattern recognition capability with 91.3 % accuracy after 10 training epochs. Finally, flexible substrate-based SPE/F-GDY devices integrated with gate electrodes were fabricated to record physiological signals, showing promise for in-sensing applications. This work shows the potential of F-GDY/Li-SPE devices for low-power neuromorphic computing, though real-time parallel processing and multimodal integration (e.g., visual, auditory) remain unexplored. However, the biocompatibility and efficient ion-assisted transmission of F-GDY suggest promising future applications in these areas.

Asymmetric Spatiotemporal Online Learning for Deep Spiking Neural Networks

Although the precise symmetric forward and feedback connections between neurons are thought to be impossible in the brain, most existing deep spiking neural networks (SNNs) spatio-temporal learning algorithms still require such a strong architectural constraint to complete tasks. Besides that, as an expected feature for specialized neuromorphic hardware to be deployed, effective online learning of deep SNNs for training spatio-temporal data are still lacking. To address these issues, an effective biologically plausible asymmetric spatio-temporal online learning algorithm called ASTOL is proposed for training deep SNNs in real-time. ASTOL could learn the spatial and temporal features simultaneously in real-time, without the precise weights symmetric backpropagation and the need to know the duration of spatio-temporal data in advance. Experimental results show that the proposed ASTOL algorithm achieves comparable performance in real-time learning on the rate coding MNIST dataset and the temporal coding music MedleyDB dataset and speech TIDIGITS dataset with other state-of-the-art algorithms with simpler learning mechanism way as it avoids transport of synaptic weights. Besides that, the experiment on MNIST datasets also shows that ASTOL can use simpler network structures to achieve good performance faster.

An Electrolyte‐Gated InGaZnO Phototransistor that Emulates Visual Experience‐Dependent Plasticity

AbstractInspired by neurobiological learning rules, bionic devices that simulate the fundamental functions of synapses and neurons provide a highly effective approach to neuromorphic computing. Among various learning rules, the Bienenstock‐Cooper‐Munro (BCM) learning rule can explain the threshold sliding effect of synaptic weight modification in the visual cortex, which is difficult to explain with the classical Hebb's rule. Existing research mainly focuses on exploiting electrical stimulation to implement the BCM rule, while the optical implementation is still unexplored. In this paper, the light‐history‐dependent BCM learning rule is implemented with electrolyte‐gated InGaZnO (IGZO) transistors. The channel conductance can be modulated through light illumination and electrical stimulation. By utilizing the light‐history‐dependent property of the IGZO electrolyte‐gated transistor and following the triplet‐spike‐timing‐dependent plasticity (STDP) rules, the BCM learning rule is successfully emulated in a single device. Moreover, the light‐history‐dependent property enables a variety of bionic vision functions including image edge detection and associative memory. This work provides a paradigm for the novel implementation of the BCM rule and paves the way for further development of machine vision systems.

Practical Approach to Studying Evolutionary Methods for Setting Weight Coefficients of Artificial Neural Networks

The article describes the problems of developing neurocontrollers for controlling dynamic objects, including the complexity of forming training data sets. It is indicated that one of the known methods for training an artificial neural network controlling an object is the neuroevolutionary approach, which involves using a genetic algorithm to adjust the synaptic weighting coefficients of an artificial neural network. The idea of using a means of demonstrating the evolutionary approach to adjusting the weighting coefficients of an artificial neural network for practical training of students in the basics of the neuroevolutionary approach is proposed. Software has been developed to demonstrate the neuroevolutionary approach using the example of the evolution of an artificial neural network of a given structure intended to control a simplified computer model of an autonomous vehicle. A method for resolving the problem of stagnation when using the evolutionary approach to training an artificial neural network is described. Options for using the developed software in teaching students the basics of artificial intelligence technologies and evolutionary methods of multicriteria optimization are proposed.

Wake-Sleep Consolidated Learning.

We propose wake-sleep consolidated learning (WSCL), a learning strategy leveraging complementary learning system (CLS) theory and the wake-sleep phases of the human brain to improve the performance of deep neural networks (DNNs) for visual classification tasks in continual learning (CL) settings. Our method learns continually via the synchronization between distinct wake and sleep phases. During the wake phase, the model is exposed to sensory input and adapts its representations, ensuring stability through a dynamic parameter freezing mechanism and storing episodic memories in a short-term temporary memory (similar to what happens in the hippocampus). During the sleep phase, the training process is split into nonrapid eye movement (NREM) and rapid eye movement (REM) stages. In the NREM stage, the model's synaptic weights are consolidated using replayed samples from the short-term and long-term memory and the synaptic plasticity mechanism is activated, strengthening important connections and weakening unimportant ones. In the REM stage, the model is exposed to previously-unseen realistic visual sensory experience, and the dreaming process is activated, which enables the model to explore the potential feature space, thus preparing synapses for future knowledge. We evaluate the effectiveness of our approach on four benchmark datasets: CIFAR-10, CIFAR-100, Tiny-ImageNet, and FG-ImageNet. In all cases, our method outperforms the baselines and prior work, yielding a significant performance gain on continual visual classification tasks. Furthermore, we demonstrate the usefulness of all processing stages and the importance of dreaming to enable positive forward transfer (FWT). The code is available at: https://github.com/perceivelab/wscl.