P286 Learning in Visual Cortex: Sparseness, Balance, Decorrelation, and the Parameters of the Leaky Integrate-and-Fire Model.
Martin J. Spencer1*, Marko A. Ruslim1*, Hinze Hogendoorn2, Hamish Meffin1, Yanbo Lian1, Anthony N. Burkitt1,3
1Department of Biomedical Engineering, University of Melbourne, Melbourne, Victoria 3010, Australia 2School of Psychology and Counselling, Queensland University of Technology, Kelvin Grove, Queensland 4059, Australia 3Graeme Clark Institute for Biomedical Engineering, University of Melbourne, Melbourne, Victoria 3010, Australia
*Equal first authors. Email:martin.spencer@unimelb.edu.au InIntroduction:Sparseness is a known property of information representation in the cortex [1]. A sparse neural code represents the underlying causes of sensory stimuli and is resource efficient [2]. Computational models of sparse coding in the visual cortex typically use an objective function with an information maximization term and a neural activity minimization term, a top-down approach [3]. In contrast, this study trained a spiking neural network using Spike-Timing-Dependent Plasticity (STDP) learning rules [4]. The resulting sparseness, decorrelation, and balance in the network was then quantified; a bottom-up approach [5]. To confirm the mechanisms of sparseness, results were replicated across 3 models of increasing complexity. Methods:A biologically grounded V1 model was made up of separate populations of excitatory and inhibitory Leaky Integrate and Fire (LIF) neurons with all-to-all connectivity via delta-current synapses. Input was provided by Poisson neurons with spike rates representing the output of separate ON and OFF neurons calculated using a centre-surround whitening filter applied to natural images. The V1 LIF neuron spike rates were maintained at a target rate using a homeostatic threshold adjustment. Synaptic weights were adjusted using a triplet STDP rule [4] for the excitatory-excitatory neuron synapses and a symmetric STDP rule for other connections. Learning was normalised using subtractive normalisation and multiplicative normalisation. Results:Training was performed using 1200 batches of 100 natural image patches for 400 ms each (~11 hours). There were 512 LGN neurons (256 ON, 256 OFF) and 500 V1 neurons (400 Excitatory, and 100 Inhibitory). The network was found to achieve a sparse representation. The level of sparseness was found to depend on the parameters of the LIF model. These mechanisms were additionally explored in a simple single neuron model and computationally efficient smaller model (Figure 1). Decorrelation was observed to result from the weights chosen by STDP. ‘Loose’ and ‘tight’ balance was confirmed using comparison of the relative strength of excitatory and inhibitory input. Discussion:In the biologically grounded V1 model the results showed that the balance was maintained across long (~1 s) and short (~10 ms) times scales. Where pairs of neurons had receptive fields with high correlations it was found that there was correlated high mutual inhibition leading to diversity and information maximization in the network. In all 3 models, higher sparseness (ς) was caused by lower output spike rates in the LIF neurons (Figure 1 A and C, efficient model). In the efficient and biologically grounded models this was associated more Gabor-like receptive fields (Figure 1 B and D). Other parameters of the LIF model were also examined, including membrane time constant, input spike rate and number of inputs.
Figure 1. Figure 1: (A) Sparseness (ς) measured in the computationally efficient V1 neuron model of 64 neurons with a 5 Hz target mean spike rate. (B) Associated normalised synaptic weight weights to 9 V1 neurons from the ON (red) and OFF (blue) input neurons. (C-D) 30 Hz target mean spike rate. Acknowledgements Acknowledgements
This work was supported by an Australian Research Council Discovery Grant (DP220101166). References References [1] - https://doi.org/10.1038/s41467-020-14645-x [2] - https://doi.org/10.1038/srep17531 [3] - https://doi.org/10.1038/381607a0 [4] -https://doi.org/10.1523/JNEUROSCI.1425-06.2006 [5] - https://doi.org/10.1101/2024.12.05.627100
