Bio-Inspired Techniques in a Fully Digital Approach for Lifelong Learning

Abstract

Lifelong learning has deeply underpinned the resilience of biological organisms respect to a constantly changing environment. This flexibility has allowed the evolution of parallel-distributed systems able to merge past information with new stimulus for accurate and efficient brain-computation. Nowadays, there is a strong attempt to reproduce such intelligent systems in standard artificial neural networks (ANNs). However, despite some great results in specific tasks, ANNs still appear too rigid and static in real life respect to the biological systems. Thus, it is necessary to define a new neural paradigm capable of merging the lifelong resilience of biological organisms with the great accuracy of ANNs. Here, we present a digital implementation of a novel mixed supervised-unsupervised neural network capable of performing lifelong learning. The network uses a set of convolutional filters to extract features from the input images of the MNIST and the Fashion-MNIST training datasets. This information defines an original combination of responses of both trained classes and non-trained classes by transfer learning. The responses are then used in the subsequent unsupervised learning based on spike-timing dependent plasticity (STDP). This procedure allows the clustering of non-trained information thanks to bio-inspired algorithms such as neuronal redundancy and spike-frequency adaptation. We demonstrate the implementation of the neural network in a fully digital environment, such as the Xilinx Zynq-7000 System on Chip (SoC). We illustrate a user-friendly interface to test the network by choosing the number and the type of the non-trained classes, or drawing a custom pattern on a tablet. Finally, we propose a comparison of this work with networks based on memristive synaptic devices capable of continual learning, highlighting the main differences and capabilities respect to a fully digital approach.

A Brain-Inspired Homeostatic Neuron Based on Phase-Change Memories for Efficient Neuromorphic Computing

“…The output homeostatic neurons (POSTs) must specialize on different classes of images presented at the input of the WTA, Figure 1B , thus enabling the spike-frequency adaptive mechanism that limits the power consumption and enables efficient classification ( Figure 1C ; Pedretti et al, 2018 ). Classification is achieved by using both excitatory synapses, which evolve by increasing or decreasing the conductance accordingly to STDP, and inhibitory synapses, which prevent the same specialization on different patterns by discharging the integration at each POST firing activity (Bianchi et al, 2020a ). Synaptic excitatory dynamics are reproduced by using PCM devices switching from low resistive state (LRS) to high resistive state (HRS), and vice versa.…”

Combining Accuracy and Plasticity in Convolutional Neural Networks Based on Resistive Memory Arrays for Autonomous Learning

“…The principal idea is to merge the benefits introduced in terms of accuracy by the CNN and the resilience typical of unsupervised learning in spiking neural networks. The first part of the network relies on two types of convolutional filters whose principal aim is to extract information from the input dataset: these two types of filters are named "class" and "feature" [14], [18]. The class filter is the result of a software training procedure whose aim is to give as convolutional output a digital "1" if, and only if, a specific class of the training dataset appears at the input; the convolutional training algorithm is implemented on a single layer convolutional network with max-pooling, Fig.…”

“…Several solutions have been proposed for overcoming this problem in ANNs, such as: (i) task-specific synaptic consolidation [11]; (ii) replacement of the old redundant information, useless for achieving better accuracy, with new one [12]; (iii) allocation of additional neural resources [13]. However, all these techniques still appear too static towards the creation of real autonomous networks since they require too great complexity in order to be efficiently implemented in hardware [14].…”

“…More precisely, unsupervised bio-inspired Hebbian learning has recently known a rapid progress in the theoretical and applied research especially with spike-timing-dependent plasticity (STDP) [15]. The main advantage of Hebbian STDP is its intrinsic plasticity, at the expense of lower accuracy with respect to standard trained networks based on backpropagation [14]. Furthermore, optimal prediction and inference of con-tinuously transforming environments seems to rely on a sort of short-term STDP, thus highlighting the timing-dependent plasticity as a good way to implement resilient neuromorphic systems [16].…”

“…These works offered an artificial approach to the biological complementary learning of hippocampus and neocortex [3]. In particular, it has been demonstrated that the merging of supervised and unsupervised learning blocks using a static CMOS logic to properly modify the convolutional results provides resilience to achieve plasticity inside a CNN [14], [22].…”