as everyone knows , The plasticity circuit in the brain is affected by the distribution of synaptic weight through the local regulation mechanism of synaptic integration and synaptic strength . However , Most of the artificial neural network training algorithms designed so far have ignored the complex interaction between stimulus dependent plasticity and local learning signals .
ad locum ,IBM Researchers at the Institute , A new biological inspired artificial and spike neural network optimizer is proposed , The optimizer combines the key principles of synaptic plasticity observed in cortical dendrites ：GRAPES（Group Responsibility for Adjusting the Propagation of Error Signals）.GRAPES The error signal modulation related to the weight distribution is realized on each node of the network .
Studies have shown that , This biologically inspired mechanism can be significantly improved with feedforward 、 Performance of artificial and spike networks with convolutional and cyclic architectures , It can reduce catastrophic forgetting , And it is most suitable for special hardware implementation . To make a long story short , Coordinating neurophysiological insights with machine intelligence is the key to improving the performance of Neural Networks .
The research 「 Introducing principles of synaptic integration in the optimization of deep neural networks」 entitled , On 2022 year 4 month 7 Published on 《 Nature Communications》.
Data address ： https://github.com/IBM/GRAPES
Thesis link ： https://www.nature.com/articles/s41467-022-29491-2