Conference on Cognitive Computational Neuroscience

Brain-inspired machine learning is gaining increasing consideration, particularly in computer vision. Several studies investigated the inclusion of top-down feedback connections in convolutional networks; however, their functional role remains unclear. Here, we consider deep convolutional networks and implement Predictive Coding (PC) dynamics through feedback connections to perform object recognition under noisy conditions. Then, we interpret the functional role of the predictive feedback by letting the network optimize the corresponding hyper-parameters in various experimental situations. Across different model depths and architectures (3-layer CNN, ResNet18, and EfficientNetB0) and against various types of noise (CIFAR100-C), we find that the network increasingly relies on top-down predictions as the noise level increases. This effect is most prominent at lower layers in deeper networks. Moreover, the accuracy of the network implementing PC dynamics significantly increases over time-steps, compared to its equivalent feed-forward network. Our results provide novel insights relevant to Neuroscience by confirming the computational role of feedback connections in sensory systems, and to Machine Learning by revealing the benefits of PC dynamics in robust computer vision. An extended version of this work is available as preprint at: https://arxiv.org/abs/2106.04225