Statistical physics of neural systems

The ability of processing and storing information is considered a characteristic trait of intelligent systems. In biological neural networks, learning is strongly believed to take place at the synaptic level, in terms of modulation of synaptic efficacy. It can be thus interpreted as the expression of a collective phenomena, emerging when neurons connect each other in constituting a complex network of interactions. In this work, we represent learning as an optimization problem, actually implementing a local search, in the synaptic space, of specific configurations, known as solutions and making a neural network able to accomplish a series of different tasks. For instance, we would like the network to adapt the strength of its synaptic connections, in order to be capable of classifying a series of objects, by assigning to each object its corresponding class-label. Supported by a series of experiments, it has been suggested that synapses may exploit a very few number of synaptic states for encoding information. It is known that this feature makes learning in neural networks a challenging task. Extending the large deviation analysis performed in the extreme case of binary synaptic couplings, in this work, we prove the existence of regions of the phase space, where solutions are organized in extremely dense clusters. This picture turns out to be invariant to the tuning of all the parameters of the model. Solutions within the clusters are more robust to noise, thus enhancing the learning performances. This has inspired the design of new learning algorithms, as well as it has clarified the effectiveness of the previously proposed ones. We further provide quantitative evidence that the gain achievable when considering a greater number of available synaptic states for encoding information, is consistent only up to a very few number of bits. This is in line with the above mentioned experimental results. Besides the challenging aspect of low precision synaptic connections, it is also known that the neuronal environment is extremely noisy. Whether stochasticity can enhance or worsen the learning performances is currently matter of debate. In this work, we consider a neural network model where the synaptic connections are random variables, sampled according to a parametrized probability distribution. We prove that, this source of stochasticity naturally drives towards regions of the phase space at high densities of solutions. These regions are directly accessible by means of gradient descent strategies, over the parameters of the synaptic couplings distribution. We further set up a statistical physics analysis, through which we show that solutions in the dense regions are characterized by robustness and good generalization performances. Stochastic neural networks are also capable of building abstract representations of input stimuli and then generating new input samples, according to the inferred statistics of the input signal. In this regard, we propose a new learning rule, called Delayed Correlation Matching (DCM), that relying on the matching between time-delayed activity correlations, makes a neural network able to store patterns of neuronal activity. When considering hidden neuronal states, the DCM learning rule is also able to train Restricted Boltzmann Machines as generative models. In this work, we further require the DCM learning rule to fulfil some biological constraints, such as locality, sparseness of the neural coding and the Dale’s principle. While retaining all these biological requirements, the DCM learning rule has shown to be effective for different network topologies, and in both on-line learning regimes and presence of correlated patterns. We further show that it is also able to prevent the creation of spurious attractor states.