Unsupervised Learning of Visual Features through Spike Timing Dependent Plasticity

By Timothée Masquelier, Simon J. Thorpe. 2007

581 citations recorded by the source corpus.

Spike timing dependent plasticity (STDP) is a learning rule that modifies synaptic strength as a function of the relative timing of pre- and postsynaptic spikes. When a neuron is repeatedly presented with similar inputs, STDP is known to have the effect of concentrating high synaptic weights on afferents that systematically fire early, while postsynaptic spike latencies decrease. Here we use this learning rule in an asynchronous feedforward spiking neural network that mimics the ventral visual pathway and shows that when the network is presented with natural images, selectivity to intermediate-complexity visual features emerges. Those features, which correspond to prototypical patterns that are both salient and consistently present in the images, are highly informative and enable robust object recognition, as demonstrated on various classification tasks. Taken together, these results show that temporal codes may be a key to understanding the phenomenal processing speed achieved by the visual system and that STDP can lead to fast and selective responses.

Subfields: Cognitive Neuroscience, Electrical and Electronic Engineering

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

In Electrical and Electronic Engineering

Citations in the corpus

HFirst: A Temporal Approach to Object Recognition (2015, background): “Adap-tations of frame-based CNN techniques for training SNNs and implementing them in FPGA have also been recently presented [42], including a recent PAMI paper [35] which presented a high speed card pip recognition task which we also tackle in this paper as a comparison to existing works.”
Spiking Neural Networks and Their Applications: A Review (2022, methodology): “A probabilistic variant of simpliﬁed STDP [58] that adjusts the synaptic weight for LTP according to an exponential function of the current weight magnitude was introduced by [59].”
CIFAR10-DVS: An Event-Stream Dataset for Object Classification (2017, methodology): “One way is to convert the frame-based images into event streams by simulation (Masquelier and Thorpe, 2007; O’Connor et al., 2013).”
Learning and disrupting invariance in visual recognition (2011, methodology): “Temporal association methods are attractive solutions to the problem of invariance development or learning [1] [7] [8] [9] [10] [11] [12] [13] [14].”
Unsupervised learning of visual invariance with temporal coherence (2011, background): “Another plausible principle specific to learning visual invariance is temporal coherenc e [4, 5, 6, 7].”
Converting Static Image Datasets to Spiking Neuromorphic Datasets Using Saccades (2015, background): “Masquelier and Thorpe (2007) assume spike times to be proportional to local image contrast for a static image, while (O’Connor et al., 2013) simulate image motion to create a spike sequence.”
Training Deep Spiking Neural Networks Using Backpropagation (2016, methodology): “The key idea of our approach is to generate a continuous and differentiable signal on which SGD can work, using low-pass ﬁltered spiking signals added onto the membrane potential and treating abrupt changes of the membrane potential as noise during error backpropagation.”
Interpreting individual classifications of hierarchical networks (2013, background): “Similarly, one can visualize the network topology in order to understand the overall strategies that might be used [6].”
Deep Learning in Spiking Neural Networks (2018, methodology): “Hand-crafted convolutional kernels have been used in the first layer of a number of spiking CNNs that have obtained high classification performance [59], [84], [60], [169], [170].”
A review of learning in biologically plausible spiking neural networks (2019, unknown): “…kernels and these SCNNs have been applied in a number of classification tasks (Kheradpisheh, Ganjtabesh, & Masquelier, 2016; Kheradpisheh, Ganjtabesh, Thorpe, & Masquelier, 2018; Masquelier & Thorpe, 2007; Tavanaei, Masquelier, & Maida, 2016; Zhao, Ding, Chen, Linares-Barranco, & Tang, 2015).”
The Limits of Feedforward Vision: Recurrent Processing Promotes Robust Object Recognition when Objects Are Degraded (2012, methodology): “Overall, the model can be viewed as an extension of a large class of hierarchical feedforward models of visual processing in the brain (eg, Masquelier & Thorpe, 2007 ; Serre, Kreiman, et al, 2007 ; Delorme & Thorpe, 2001 ; Riesenhuber & Poggio, 1999 ).”
Dendritic cortical microcircuits approximate the backpropagation algorithm (2018, background): “, 2011; Froemke, 2015) and competition (Masquelier and Thorpe, 2007; Nessler et al., 2013).”
Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification (2017, methodology): “Another suggestion is to use a temporal Winner-Take-All based on time-to-first-spike encoding, in which the first neuron to fire is considered the maximally firing one (Masquelier and Thorpe, 2007; Orchard et al., 2015b).”
Recurrent Processing during Object Recognition (2013, background): “Computational models of object recognition that implement a feedforward, hierarchical version of the ventral pathway have explained many aspects of the initial neural response properties across these different brain areas (Fukushima, 1980, 2003; Wallis and Rolls, 1997; Riesenhuber and Poggio, 1999; ”
Conditional modulation of spike-timing-dependent plasticity for olfactory learning (2012, background): “This in turn should curb bLN population activity (see also ref”
Unsupervised Natural Visual Experience Rapidly Reshapes Size Invariant Object Representation in Inferior Temporal Cortex (2010, unknown): “Several instantiated computational models show how this conceptual strategy can build tolerance (Foldiak, 1991; Masquelier et al., 2007; Masquelier and Thorpe, 2007; Wallis and Rolls, 1997; Wiskott and Sejnowski, 2002; Wyss et al., 2006), and such models can be implemented using variants of…”