STDP-based spiking deep convolutional neural networks for object recognition

Abstract

Previous studies have shown that spike-timing-dependent plasticity (STDP) can be used in spiking neural networks (SNN) to extract visual features of low or intermediate complexity in an unsupervised manner. These studies, however, used relatively shallow architectures, and only one layer was trainable. Another line of research has demonstrated - using rate-based neural networks trained with back-propagation - that having many layers increases the recognition robustness, an approach known as deep learning. We thus designed a deep SNN, comprising several convolutional (trainable with STDP) and pooling layers. We used a temporal coding scheme where the most strongly activated neurons fire first, and less activated neurons fire later or not at all. The network was exposed to natural images. Thanks to STDP, neurons progressively learned features corresponding to prototypical patterns that were both salient and frequent. Only a few tens of examples per category were required and no label was needed. After learning, the complexity of the extracted features increased along the hierarchy, from edge detectors in the first layer to object prototypes in the last layer. Coding was very sparse, with only a few thousands spikes per image, and in some cases the object category could be reasonably well inferred from the activity of a single higher-order neuron. More generally, the activity of a few hundreds of such neurons contained robust category information, as demonstrated using a classifier on Caltech 101, ETH-80, and MNIST databases. We also demonstrate the superiority of STDP over other unsupervised techniques such as random crops (HMAX) or auto-encoders. Taken together, our results suggest that the combination of STDP with latency coding may be a key to understanding the way that the primate visual system learns, its remarkable processing speed and its low energy consumption. These mechanisms are also interesting for artificial vision systems, particularly for hardware solutions.

The covariance perceptron: A new paradigm for classification and processing of time series in recurrent neuronal networks

“…In fact, the biologically-inspired model of spike-timing-dependent plasticity (STDP) can be expressed in terms of covariances between spike trains [22,16], which was an inspiration of the present study. STDPlike learning rules were used for object recognition [23] and related to the expectation-maximization algorithm [30]. Although genuine STDP relates to unsupervised learning, extensions were developed to implement supervised learning for spike patterns [20,35,19,14,41].…”

Researchers Keep Rejecting Grandmother Cells after Running the Wrong Experiments: The Issue Is How Familiar Stimuli Are Identified

“…We understand this intuition, but it is worth emphasizing again that there are existing models of visual object recognition that are designed to be biologically plausible that have grandmother cells . Indeed, there are biologically plausible models of visual object identification that learn grandmother cells . Even artificial neural network models that are claimed to learn distributed codes in fact learn grandmother representations when trained to coactivate multiple words (or objects or faces) at the same time in short‐term memory .…”

“…[31] Indeed, there are biologically plausible models of visual object identification that learn grandmother cells. [70] Even artificial neural network models that are claimed to learn distributed codes in fact learn grandmother representations when trained to coactivate multiple words (or objects or faces) at the same time in short-term memory. [50,71] These findings suggest that there are computational advantages of grandmother cells in the context of shortterm memory, much like there are computational advantages of learning highly selective representations in the hippocampus for the sake of episodic memory.…”

Computational Foundations of Natural Intelligence

“…For example, in order for neural networks to act as accurate models of neural information processing it may be imperative to use spike-based rather than rate-based formulations (Brette, 2015). Efforts are underway to effectively train spiking neural networks (Gerstner et al, 2014;Gerstner and Kistler, 2002;O'Connor and Welling, 2016;Huh and Sejnowski, 2017) and endow them with the same cognitive capabilities as their rate-based cousins Zambrano and Bohte, 2016;Kheradpisheh et al, 2016;Lee et al, 2016;Thalmeier et al, 2015). In the same vein, researchers are exploring how probabilistic computations can be performed in neural networks (Pouget et al, 2013;Nessler et al, 2013;Orhan and Ma, 2016;Heeger, 2017) and deriving new biologically plausible synaptic plasticity rules (Schiess et al, 2016;Brea and Gerstner, 2016).…”