P303 Sleep-like Homeostatic and Associative Intra- and Inter-Areal Plasticity Enhances Cognitive and Energetic Performance in Hierarchical Spiking Network
Leonardo Tonielli1, Cosimo Lupo1, Elena Pastorelli1, Giulia De Bonis1, Francesco Simula1, Alessandro Lonardo1, Pier Stanislao Paolucci1
1Istituto Nazionale di Fisica Nucleare, Sezione di Roma Introduction
Can hierarchical bio-inspired AI spiking networksand biological brainsengaged in incremental learning benefit from unsupervised plasticity during an offline deep-sleep-like period? We show that simultaneous intra- and inter-areal plasticity enhances the cognitive and energetic benefitsofdeepsleep-likeinathalamo-cortical modelinspired bythecorticalorganizing principle[1]andthehomeostatic-associative sleephypothesisas in[2, 3],that learns,retrievesand classifies handwritten digits from few examples. This outperformsresultspresented in[4], where deep sleep is limited to cortico-cortical plasticity. Methods The network is a two-areaspiking model (Fig1A) using Integrate-and-fire neurons with spike-frequency adaptation. Each layer is composed of excitatory-inhibitory populations. The input consists of MNIST images preprocessed with a HOG filter[5](30 training, 250 test). The perceptual stream is released from the thalamusand propagates through plastic feedforward connections to cortex, which encodes memories within neural assemblies elicited by specific contextual stimuli. Sleep-like dynamics is stimulated by non-specific cortical noise generating slow-oscillation activity that promotes memoryreplayand thusconsolidateslearning through homeostatic and associative processes within cortical synapses and thethalamo-cortical loop. Results We assessed the cognitive and energetic performance of the network by measuring the most-active neuron’sclassification accuracy (Fig1B),thenetwork’s mean firing rate (C) and the synaptic change (C, D) over 2000 seconds of sleep. We comparedthalamo-cortical plastic sleep with cortico-cortical plasticityonly. Our findingsindicatethat fullthalamo-cortical plasticity strongly enhances classification performance (B) and firing rate downscaling (C) while preserving the same associative-homeostaticbehaviourat the cortico-cortical synaptic level (D). Specifically, weobserveda significant 5% improvement in classification accuracy and a 25% reduction in firing rate, enabling the network to classify better by consuming less energy. Discussion We proposed a minimalthalamo-cortical model that classifies images drawn from the MNIST set of handwritten digits,capable of improving cognitive performance by homeostatic-associative cortical plastic deep-sleep-like activity. While cortical sleep is important to normalize high level representations and to develop new synapses, our new results suggest thatthalamo-cortical sleep is fundamental to coordinate cortical activation and to regulateits waking activity. This effect might be beneficial also to deep neural networkalgorithms which lack this generalizationfeature,andit’salsorelevant for cerebral neural networks.
Figure 1. Solid lines: full plasticity, dotted: cortico-cortical only. Deep-sleep after training with 3 examples / digit class (A) Network’s structure. (B) Classification from most active neuron (C) Mean firing rate during classification and overall synaptic change. (D) cortico-cortical synaptic change: synapses encoding assemblies (blue), same class (yellow) different class (red). 100 configurations. Acknowledgements Workcofundedbythe European Next Generation EUgrants,ItaliangrantsCUP I53C22001400006 (FAIR PE0000013 PNRR) and CUP B51E22000150006 (EBRAINS-Italy IR00011 PNRR).APE parallel/distributed lab at INFN Roma,BRAINSTAIN. LeonardoTonielliis a PhD studentofthe National PhD program in Artificial IntelligenceXLcycle Healthand life sciences, organized by Università Campus Bio-Medico di Roma.
