Associative memory: an embodied spiking neural network robotic implementation
This article proposes a minimalist bio-inspired associative memory (AM) mechanism that can be implemented in a spiking neural network acting as a controller for virtual and real robots. As such, several main features of a general AM concept were reproduced. Using the strength of temporal coding at the single spike resolution level, this study approach the AM phenomenon with basic examples in the visual modality. Specifically, these embedded neural architectures include varying time delays in synaptic links and asymmetry in the spike-timing dependent plasticity learning rules to resolve visual tasks of pattern-matching, pattern-completion and noise-tolerance for autoassociative and heteroassociative memories. This preliminary work could serve as a step toward future comparative analysis with traditional artificial neural network paradigm.
Note : The following is all supplementary materials available for the article.
SNN architecture
AM process based on STDP
This figure shows the detail of the SNN AM process in a visual learning patterns task, based on asymmetrical input delays and asymmetry in the STDP rule
One could see in the figure, graphics A-B-C (IBL=input-bottom-left, IML=input-middle-left, IUL=input-upper-left) that the image caption of the three black dots on the left column generates single spikes of their respective neurons. The randomized spike delays (graphics D-E-F: 0-4 cycles of algorithm scale) between the input and the associative neural layer are produced (graphics G-H-I). These small delays are enough to be used from the STDP rule to adjust the modulating synaptic factor (graphics J to O: 0-500 percentage scale). Since the synaptic delays are randomized, anti-hebb happens half of the time, but because there was a positive bias favoring the hebb side, this parameter eventually results in an increased weight and affects the spiking response of the associative neural units. The graphics P-Q-R simply represent the output spikes to the LCD device. Balance of the parameters were made between the synaptic weight percentage change per step, the bias factor from the STDP rule and the number of input patterns. This was done in order to reach the critical phase where all elements composing the input patterns are effectively learned after a few repetitions.
Equations
