Joshua Arnold,
Peter Stratton,
Janet WilesZoom meeting id: 98578299540 Meeting link:
https://uqz.zoom.us/j/98578299540Spiking neurons inherently represent time due to their momentary discrete action potentials; as such, they are well poised to process spatiotemporal data. Despite their temporal nature, most computational learning rules focus on modulating synaptic efficacy (weight), which only indirectly influences a neuron's temporal dynamics. Weight-based rules are well suited to solving synchronous spatial learning tasks, as demonstrated by the surge of interest in rate-coded neurons performing frame-based image classification using backpropagation. For temporal tasks, however, weight based learning rules often implicitly rely on the temporal dynamics of membrane equations or synaptic transfer functions to discriminate between spatially identical, but temporally distinct, inputs. Allowing spiking neurons to perform some aspect of explicit temporal learning offers significant advantages for learning asynchronous spatiotemporal patterns compared to weight-based rules alone. With improvements in imaging techniques, there is accumulating evidence for action-potential conduction velocity plasticity over long and short timescales [1, 2]. The biological mechanisms implementing Conduction Delay Plasticity (CDP) could include myelination, changes in axon diameter, changes to nodes of Ranvier length, bouton movement, or likely some combination of these mechanisms and others not listed. While the precise nature and interaction of the biological mechanisms underlying CDP remain elusive, computational models provide a framework in which theories can be tested. Several CDP learning rules have been suggested with greatly varying levels of biological fidelity and computational efficiency; in particular, we focus on one rule called Synaptic Delay Variance Learning [3]. Here we demonstrate the ability of a Leaky Integrate and Fire spiking model using only CDP (no weight learning) to learn a repeating spatiotemporal pattern in a continuous time input stream with no training signal; that is, the delays self-organise to represent the temporal structure of the input. A neuron receives 2000 afferents firing with Poisson distributions of 10Hz, while the embedded pattern is presented with a Poisson distribution of 5Hz and consists of 500 afferents firing once within a 50ms period. The input is normalised such that the patterns cause no change in overall activity during presentations and all afferents involved in the pattern are adjusted to maintain a 10Hz firing rate. After 250 seconds of training, the neuron is tested for 50 seconds and successfully responds to 99.7% of pattern presentations with 3.1% false positives, averaged over 100 trials. These results provide a demonstration of CDP as a functional computational learning rule enabling spiking neurons to perform unsupervised learning of spatiotemporal data.
[1] Fields RD. A new mechanism of nervous system plasticity: activity- dependent myelination. Nat Rev Neurosci. 2015 Dec;16(12):756-67.
[2] Arancibia-Carcamo IL, Ford MC, Cossell L, Ishida K, Tohyama K, Attwell D. Node of Ranvier length as a potential regulator of myelinated axon conduction speed. Elife. 2017 Jan 28;6:e23329.
[3] Wright PW, Wiles J. Learning transmission delays in spiking neural networks: A novel approach to sequence learning based on spike delay variance. In The 2012 International Joint Conference on Neural Networks (IJCNN) 2012 Jun 10 (pp. 1-8). IEEE.
