| Summary: | <p>One of the hallmarks of intelligent behaviour is the ability to generalise, i.e. to make use of information from specific experiences to solve unencountered scenarios. Here I focus on the generalisation task of transitive inference, which tests the ability to infer that stimulus A is preferable to C, based solely on previous training that A is preferred over B and B is preferred over C. Such inference tasks are thought to be performed by hippocampal-entorhinal circuits and rely heavily on recurrent and conjunctive connectivity to bind associated stimuli. Notably, a 3-layer recurrent neural network model called REMERGE (Kumaran & McClelland, 2012) can solve the transitive inference task through recall dynamics in a circuit that also relies on recurrent and conjunctive connectivity. However, it is not clear how the architecture could be constructed by a learning rule. Here, I develop a biologically plausible non-linear Hebbian plasticity rule to build REMERGE’s connectivity starting from an all-to-all architecture. I separate the problem of learning the complete architecture into two parts: 1 - how to learn input connections and 2 - how to learn output connections. For the first part, I design a learning rule that forms connections from coincidentally stimulated inputs to separated conjunctive units in the hidden layer. Learning the second set of weights requires a more complex rule that takes into account a reward signal if the choice was correct, allowing only connections that lead to reward to be strengthened. Ultimately, the results predict a reorganisation of recurrent activity flow across training by potentiating feedforward and feedback connections in hippocampal-entorhinal stimulus pathways, and the need of reward-dependent plasticity to form connections to response neurons. My model bridges the gap between conceptual models of generalisation and how to learn them through biologically plausible plasticity rules.</p>
|
|---|