Biologically inspired adaptive mechanisms for artificial neural networks

Relational reasoning is a crucial cognitive capacity of humans and other non-human animals and it can be viewed as an generalized concept of a task rule that can be applied whenever certain relations between objects hold true. Animals and humans readily learn statistical regularities in their environment, which is demonstrated by their ability to quickly apply the same task structure in different settings as shown by reversal learning, among other findings. It is therefore possible that relational reasoning and generalization are facilitated by similar architectural characteristics. It has been shown that varying levels of neuron selectivity and firing rates influence the ability to retain contextual information in animals, which enables making different actions depending on the task at hand. However, the optimal neural connections are formed not just by learning task related aspects and so strengthening connections between different units, but also by degrading connections that are unnecessary and may interfere with future learning. This is often an issue in artificial neural networks, where novel examples can interfere with previously learned rules. The aim of this project is to implement mechanisms that are analogous to modulatory systems that facilitate learning and forgetting in biological nervous systems and investigate the effect of such systems on lifelong learning in artificial neural networks.