Understanding the theoretical foundations of how memories are encoded and retrieved in neural populations is a central challenge in neuroscience. A popular theoretical scenario for modeling memory function is an attractor neural network with Hebbian learning (e.g. the Hopfield model). The model simplicity and the locality of the synaptic update rules come at the cost of a limited storage capacity, compared with the capacity achieved with supervised learning algorithms, whose biological plausibility is questionable. Here, we present an on-line learning rule for a recurrent neural network that achieves near-optimal performance without an explicit supervisory error signal and using only locally accessible information, and which is therefore biologically plausible. The fully connected network consists of excitatory units with plastic recurrent connections and non-plastic inhibitory feedback stabilizing the network dynamics; the patterns to be memorized are presented on-line as strong afferent currents, producing a bimodal distribution for the neuron synaptic inputs (’local fields’). Synapses corresponding to active inputs are modified as a function of the position of the local field with respect to three thresholds. Above the highest threshold, and below the lowest threshold, no plasticity occurs. In between these two thresholds, potentiation/depression occurs when the local field is above/below an intermediate threshold. An additional parameter of the model allows to trade storage capacity for robustness, i.e. increased size of the basins of attraction. We simulated a network of 1001 excitatory neurons implementing this rule and measured its storage capacity for different sizes of the basins of attraction: our results show that, for any given basin size, our network more than doubles the storage capacity, compared with a standard Hopfield network. Our learning rule is consistent with available experimental data documenting how plasticity depends on firing rate. It predicts that at high enough firing rates, no potentiation should occur.
Input-driven unsupervised learning in recurrent neural networks
BALDASSI, CARLO;Nicolas Brunel;ZECCHINA, RICCARDO
2014
Abstract
Understanding the theoretical foundations of how memories are encoded and retrieved in neural populations is a central challenge in neuroscience. A popular theoretical scenario for modeling memory function is an attractor neural network with Hebbian learning (e.g. the Hopfield model). The model simplicity and the locality of the synaptic update rules come at the cost of a limited storage capacity, compared with the capacity achieved with supervised learning algorithms, whose biological plausibility is questionable. Here, we present an on-line learning rule for a recurrent neural network that achieves near-optimal performance without an explicit supervisory error signal and using only locally accessible information, and which is therefore biologically plausible. The fully connected network consists of excitatory units with plastic recurrent connections and non-plastic inhibitory feedback stabilizing the network dynamics; the patterns to be memorized are presented on-line as strong afferent currents, producing a bimodal distribution for the neuron synaptic inputs (’local fields’). Synapses corresponding to active inputs are modified as a function of the position of the local field with respect to three thresholds. Above the highest threshold, and below the lowest threshold, no plasticity occurs. In between these two thresholds, potentiation/depression occurs when the local field is above/below an intermediate threshold. An additional parameter of the model allows to trade storage capacity for robustness, i.e. increased size of the basins of attraction. We simulated a network of 1001 excitatory neurons implementing this rule and measured its storage capacity for different sizes of the basins of attraction: our results show that, for any given basin size, our network more than doubles the storage capacity, compared with a standard Hopfield network. Our learning rule is consistent with available experimental data documenting how plasticity depends on firing rate. It predicts that at high enough firing rates, no potentiation should occur.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.