Learning and sleep in a thalamo-cortical multi-area model

Poster code: SP3-4; Title: Learning and sleep in a thalamo-cortical multi-area model Authors: Chiara De Luca, Cristiano Capone, Elena Pastorelli, Giulia De Bonis, Pier Stanislao Paolucci Poster abstract: Wakefulness and sleep are brain-states that are essential for cognitive performances. During wakefulness, our perceptual system is continuously subjected to sensory inputs from different sources and modalities. The involved brain areas process the input in a framework set by previous knowledge (acquired through individual and evolutionary experience) with a crucial role played by the exchange of signals with other brain areas. The ability of the brain to integrate and segregate this information by building a coherent and complete representation of the environment is impressive. Despite there is plenty of empirical evidence suggesting that the nervous system uses a statistically optimal approach in combining external information, little is known about how the brain implements these strategies. Moreover, recent studies have shown that sleep plays a central role in storing and reorganizing information gained while awake and in the optimization of the energetic post-sleeping rates. Our work focuses on two main issues: first, we aim to simulate the ability of the awake brain to combine different kinds of information, throughout multisensory perception, with contextual information starting from the case of the integration of the two visual hemicampi that in the brain are processed by areas placed in two different hemispheres. Second, we study the beneficial effects of a deep-sleep-like biologically plausible slow oscillation activity on the classification accuracy. In summary, in this work, we create a simplified thalamo-cortical multi-area simulation model trained to learn, sleep and perform a classification task on handwritten digits. Moreover, we compared the network behaviour and performances in a single area versus a multi-area model within a noisy environment showing a higher resilience to noise in the multi-area model. Also, we explored the formation of groups of neurons with a precise temporal order whenever dynamic examples are presented to the network during both the training and the testing phases demonstrating easier learning over moving inputs rather than static ones. We demonstrated better performances when learning moving inputs (either vibrating or sliding) rather than static inputs. Finally, we observed that the novel two-hemicampi model exhibits superior improvement in the classification accuracy induced by deep-sleep-like slow oscillation activity compared with that observed in the single area model.