Brain-Inspired Approaches to Deep Convolutional Networks for Minimalist Datasets

Deep Neural Networks have achieved impressive classification rates in vision tasks, however, the current methods are fundamentally different from biological vision. Biological systems can identify new objects after only a single exposure to the new object. In contrast, the best performing Deep Networks require thousands of examples of the object at different orientations and scales to achieve close to human performance on identify the same object. This hints at a fundamental difference in the two vision systems. The standard counterpoint to this observation has been mammalian systems simply have more neurons, however, larger and more complex models have failed to overcome the training requirement. Yet, biological research indicates that convolution and Gabor like filters are critical to the first steps of mammalian vision. This implies that some part of the current state of the art is on the right track toward successful compute vision. We propose to research alternative Neural Network architectures that so far have received comparatively little attention. Specifically, Tau will investigate combing the lower layers of Deep Neural Networks (i.e. the layers that look like Gabor filters) with Spiking Networks, Capsule Networks, and Neocognitron based networks.  These three networks are fundamentally different from the standard feed forward networks used by the most successful Deep Neural Networks. Of paramount interest is that these networks require only a few examples of a given object to be able to classify that object with high accuracy. Yet, they have many problems which have not been adequately solved to make them as effective as the state of the art Deep Neural Networks such as YOLO, ResNet or Faster RCNN. In this proposal we will outline the issues and the research direction Tau will take to attempt to overcome the issues with these novel networks.