Conference on Cognitive Computational Neuroscience

Recurrent neural networks (RNNs) have yielded promising results for both recognizing objects in challenging conditions and modeling aspects of primate vision. However, the representational dynamics of recurrent computations remain poorly understood. We performed such an analysis based on RNNs trained for object classification on MiniEcoset, a structured subset of Ecoset. We report multiple observations. First, the networks' readout vectors aligned with the hierarchical structure of the dataset. Second, upon inference, representations continued to evolve after correct classification, suggesting a lack of the notion of being "done with classification". Third, focusing on "readout zones" as a way to characterize the activation trajectories, we observed that misclassified representations are located lower and peripherally in the readout zones. Such arrangements help the misclassified representations move into the correct zones. Our findings generalized to lateral and top-down connections, as well as how they interact with the bottom-up sweep (additive and multiplicative), and therefore contribute to a better understanding of RNN dynamics in naturalistic tasks. Our analysis framework will aid future investigations of other types of RNNs (e.g. LSTMs) and our understanding of representational dynamics in primate vision.