Oral 2 - 2.1
1,2Lina M. Tran, 1,2,3Sheena A. Josselyn, 4Blake A. Richards, 1,2,3Paul W. Frankland
1 Dept. of Physiology, University of Toronto; 2 Neurosciences and Mental Health, The Hospital for Sick Children; 3 Dept. of Psychology, University of Toronto; 4 Biological Sciences, University of Toronto Scarborough
In neural networks, the stability of stored information may be compromised by a) changing the neural architecture, and b) adding new memories. We tested these ideas in a three layer feed-forward neural network. In this network the input, middle and output layers represent the entorhinal cortex, DG and CA3 regions, respectively. We presented the network with input patterns, drawn from two partially overlapping distributions A and B. The network was trained to transform these A and B input patterns into one of two discrete output patterns. The ability of the network to generalize was assessed by presenting novel input patterns that were either drawn from the A and B distributions, and asking whether it could correctly categorize them. Following initial training, we either a) added new neurons to the middle layer, or b) trained the network on two new distributions, C and D. Changes in neural architecture induced by either a) new neuron addition or b) new learning impaired AB categorization performance, consistent with previous modeling and experimental data. Moreover, increasing rates of neuron addition, excitability and connectivity (input and output) of the new neurons exacerbated these forgetting effects. To test how these effects may impact new learning, we trained the network on distributions A and B and then added new neurons, as above. We then retrained the network on the new distributions C and D. In this case, the addition of new neurons weakened memory for the original AB categorization as expected, but, at the same time, enhanced learning of the new CD categorization. Furthermore, there is a positive correlation between how much the original memory was impaired and the enhancement in learning of new patterns. By exploring how new neuron addition impacts stored memories, and new memory storage, our results begin to help us understand how adult neurogenesis in the hippocampus influences cognition.