Network of evolvable neural units can learn synaptic learning rules and spiking dynamics

Bertens, Paul; Lee, Seong Whan

doi:10.1038/s42256-020-00267-x

Cited by 11 publications

(7 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Neurons are powerful information processing units and can perform highly complex nonlinear computations even in individual cells. [ 50–52 ] Hardware implementation of artificial neurons with similar capability is of great significance for the construction of neuromorphic systems; in this aspect, memristive devices are crucial for implementing neuronal functions in artificial neurons. [ 6,53,54 ]…”

Section: Resultsmentioning

confidence: 99%

“…Neurons are powerful information processing units and can perform highly complex nonlinear computations even in individual cells. [50][51][52] Hardware implementation of artificial neurons with similar capability is of great significance for the construction of neuromorphic systems; in this aspect, memristive devices are crucial for implementing neuronal functions in artificial neurons. [6,53,54] Among various neuron models that have been established to describe the complex dynamics of biological neurons, the LIF neuron is most widely used for both algorithms and hardware implementations of SNNs.…”

Section: Lif Neuronsmentioning

confidence: 99%

See 1 more Smart Citation

A Bio‐Inspired Neuromorphic Sensory System

Wang

Wen

et al. 2022

Advanced Intelligent Systems

View full text Add to dashboard Cite

The advent of the intelligent society leads to the exponential growth of information, imposing urgent requirements in a time‐ and energy‐efficient way to process information where data are generated. This issue can be addressed by the neuromorphic paradigm of computing inspired by biological sensory systems that build up the association between external stimuli and the response of an organism in real‐time; in the paradigm, a neuromorphic system is integrated with sensors to form an artificial sensory system. Herein, a neuromorphic sensory system with integrated capabilities of gas sensing, data storage, and processing is demonstrated. Leaky integrate‐and‐fire (LIF) neurons, the basic computing units in the system, are realized with volatile memristive device Pt/Ag/TaOx/Pt; sensory neurons, i.e., the LIF neurons connected with an array of gas sensors, detect gases and convert the chemical information of gases into neural spikes; synapses based on nonvolatile memristive device Pt/Ta/TaOx/Pt transmit the signals from sensory neurons to relay neurons according to synaptic weights, which are trained by the supervised spike‐rate dependent plasticity; relay neurons then process the signals from the synapses and classify gases. The approach of this work can also be applied to emulate other biological perceptions through the integration with different sensors.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Lif Neuronsmentioning

confidence: 99%

A Bio‐Inspired Neuromorphic Sensory System

Wang

Wen

et al. 2022

Advanced Intelligent Systems

View full text Add to dashboard Cite

show abstract

“…They also test their rule's ability to control networks with more hidden neurons. More closely related to our method, is the work of Bertens and Lee [4]. They evolve a set of recurrent neural network cells and use them as basic units to form a network between them.…”

Section: Related Workmentioning

confidence: 99%

“…Closely related to the approach of Bertens and Lee [4] is the work of Kirsch et al [21]. They evolve a shallow network structure where all connections consist of recurrent neural networks with shared parameters.…”

Section: Related Workmentioning

confidence: 99%

Minimal Neural Network Models for Permutation Invariant Agents

Pedersen,

Risi

2022

Preprint

View full text Add to dashboard Cite

Organisms in nature have evolved to exhibit flexibility in face of changes to the environment and/or to themselves. Artificial neural networks (ANNs) have proven useful for controlling of artificial agents acting in environments. However, most ANN models used for reinforcement learning-type tasks have a rigid structure that does not allow for varying input sizes. Further, they fail catastrophically if inputs are presented in an ordering unseen during optimization. We find that these two ANN inflexibilities can be mitigated and their solutions are simple and highly related. For permutation invariance, no optimized parameters can be tied to a specific index of the input elements. For size invariance, inputs must be projected onto a common space that does not grow with the number of projections. Based on these restrictions, we construct a conceptually simple model that exhibit flexibility most ANNs lack. We demonstrate the model's properties on multiple control problems, and show that it can cope with even very rapid permutations of input indices, as well as changes in input size. Ablation studies show that is possible to achieve these properties with simple feedforward structures, but that it is much easier to optimize recurrent structures. CCS CONCEPTS• Computing methodologies → Artificial intelligence.

show abstract

“…Bertens and Lee proposed an evolvable neural unit (ENU) that can evolve individual somatic and synaptic compartment models of neurons in a scalable manner and try to solve a T-maze environment task [25].…”

Section: Introductionmentioning

confidence: 99%

Plasticity Neural Network Based on Astrocytic effects at Critical Period, Synaptic Competition and Strength Rebalance by Current and Mnemonic Brain Plasticity and Synapse Formation

Tao¹,

Sun²,

Zhu³

et al. 2022

Preprint

View full text Add to dashboard Cite

Beyond the consideration of the basic connection, this research focus on a new Neutral Network (NN) model, referred as PNN (Plasticity Neural Network) in which the current and mnemonic range of action between synapses and neurons plastic changes with iteration are both taken into account at the same time, specifically, the current and mnemonic synaptic range of action weights. In addition, the coding result on PNN mode is given that as one synapse strengthens, neighboring synapses weaken on their own to compensate it; this mechanism is consistent with the findings in the recent research on brain plasticity at the MIT Picower Institute. Regarding the importance of this mechanism, Dr. Luo's team at Stanford University has mentioned that the competition regarding synapse formation for dendritic morphogenesis is important. The influence of astrocyte impacts on brain plasticity and synapse formation is an important mechanism of our Neural Network at critical periods and the end of critical periods. We try to examine the mechanism of failure in brain plasticity by model at the end of critical periods in details by contrasting study before [17]. The new PNN model is not just modified on the frame of NN based on current gradient informational synapse formation and brain plasticity, but also the mnemonic gradient informational synapse formation and brain plasticity at critical periods. The mnemonic gradient information needs to consider forgotten memory-astrocytic synapse formation mnemonic factor. And mnemonic brain plasticity involves plus or minus disturbance-astrocytic brain plasticity disturbance factor. The influence of astrocyte made local synaptic range of action remains an appropriate length at critical periods. When one synapse transfers a signal from one neuron to other with a better stimulation signal, PNN will change the current synaptic range of action based on Mean Squared Error (MSE) of loss function, and vice versa for the same reason. For a given neuron, the synaptic plasticity of the connecting neurons from input to output units is enhanced and diminished with iteration. Considering the Recurrent Neural Network (RNN), each input variable corresponds to neurons sharing connection weights over a period of time interval; the weights are updated by the MSE loss of the activation function of the amount of output change within this synaptic range of action. The synaptic range of action is reflected in the time series, for which the real value of the simulation would lead to gradient-based change in the synaptic range of action by Back Propagation. The Back Propagation of PNN includes connection weights and synaptic range of action weights. In the simulation, the

show abstract

Network of evolvable neural units can learn synaptic learning rules and spiking dynamics

Cited by 11 publications

References 46 publications

A Bio‐Inspired Neuromorphic Sensory System

A Bio‐Inspired Neuromorphic Sensory System

Minimal Neural Network Models for Permutation Invariant Agents

Plasticity Neural Network Based on Astrocytic effects at Critical Period, Synaptic Competition and Strength Rebalance by Current and Mnemonic Brain Plasticity and Synapse Formation

Contact Info

Product

Resources

About