Stochastic binary synapses having sigmoidal cumulative distribution functions for unsupervised learning with spike timing-dependent plasticity

Abstract

Spike timing-dependent plasticity (STDP), which is widely studied as a fundamental synaptic update rule for neuromorphic hardware, requires precise control of continuous weights. From the viewpoint of hardware implementation, a simplified update rule is desirable. Although simplified STDP with stochastic binary synapses was proposed previously, we find that it leads to degradation of memory maintenance during learning, which is unfavourable for unsupervised online learning. In this work, we propose a stochastic binary synaptic model where the cumulative probability of the weight change evolves in a sigmoidal fashion with potentiation or depression trials, which can be implemented using a pair of switching devices consisting of serially connected multiple binary memristors. As a benchmark test we perform simulations of unsupervised learning of MNIST images with a two-layer network and show that simplified STDP in combination with this model can outperform conventional rules with continuous weights not only in memory maintenance but also in recognition accuracy. Our method achieves 97.3% in recognition accuracy, which is higher than that reported with standard STDP in the same framework. We also show that the high performance of our learning rule is robust against device-to-device variability of the memristor's probabilistic behaviour.

Non-Bernoulli operations of stochastic binary synapses using resistive switching devices for unsupervised learning in a spiking neural network

“…The highest accuracy achieved in this work was 96.7%, which was obtained using a network having 6400 neurons per layer in the case of 2. b = For comparison, the figure also shows the classification accuracy of the following previously reported methods. 26,27,30) It can be seen that the accuracy achieved in this work is equivalent to that with stochastic S-STDP using a gamma distribution, 27) putting our result among the best reported to date for learning with binary synapses.…”

“…The weight w itself is stored in a set-reset (SR) latch whose output Q switches off and on the synaptic current in accordance with the stored value of w (0 or 1). 27) We omit the details of the synaptic current circuit 28) here because this is beyond the scope of this work.…”

“…It is realistic to use a simplified STDP (S-STDP) rule for hardware implementation. 22,27,29) In the S-STDP rule, a synapse is modified only when the post-synaptic neuron fires. The synapse is potentiated if…”

“…Details of the simulation have been described elsewhere. 27) Figure 2(a) shows the classification accuracy as a function of the number of training images used for learning. For comparison, the accuracy obtained with continuous weights using a deterministic S-STDP rule 29) is also shown.…”

Memristive devices based hardware for unlabeled data processing

“…Besides the standard STDP mechanism, other local training rules such as stochastic simplified STDP [124][125][126] and SRDP [127,128] can be implemented by memristor naturally. For example, WO 3−x devices (figure 9(j)) is second-order memristor which has delicate dynamics originating from the temporary effects such as the electric double-layer capacitance [129].…”