Speaker
Description
Recurrent connectivity is widely thought to support cognitive functions that require activity to persist beyond the timescale of sensory input, including working memory and delayed-response behavior. Classical theoretical models of recurrent neural networks, from the Hopfield model to more biologically detailed variants, have shown how recurrent architectures can implement distributed memory and persistent neural activity. However, many theoretical treatments separate learning and retrieval phases, by assuming recurrent connectivity does not interfere with external inputs during learning. Here, we make a step beyond this unrealistic assumption, asking how recurrent connectivity shapes the learning of novel stimuli, and whether additional plasticity mechanisms are required for stable memory storage and recall when recurrent connections are not neglected during learning.
We investigate a rate-based recurrent network equipped with a biologically constrained Hebbian learning rule. The rule is motivated by in vivo measurements of how response distributions in inferotemporal cortex change as initially novel stimuli become familiar. As a first step, we analyze the learning and retrieval of two activity patterns using mean-field theory and network simulations. The network can store and retrieve the first pattern. However, when a second pattern is presented for learning, strong recurrent input biases the network state toward the previously stored representation. As a result, the synaptic update reinforces the first pattern together with the second, leading to representational collapse and a single attractor correlated with both patterns. Simulations confirm this mean-field prediction.
These results suggest that recurrent connections, while essential for memory retrieval, can interfere with the storage of new memories unless additional mechanisms decorrelate sensory responses from previously learned representations. We thus extend the mean-field theory to multiple scenarios, including cholinergic suppression of recurrent input during learning and a BCM-like mechanism, and investigate in which parameter regions stable learning occurs in these scenarios.
| Preferred Presentation | Poster Presentation |
|---|