Probabilistic Population Coding in Convolutional Neural Networks

Abstract

Computers are very practical in our everyday life and sometimes also fundamental for our work: they can store all our holidays pictures, solve quickly complex systems of equations and simulate the outcomes of experiments that cannot be executed in the real world. Computers, however, are not quite as good as we are when it comes to tell a joke, recognize a person or playing computer games. In fact, the ability to perform cognitive tasks and to process complex sensory stimuli is still a human brain exclusive. The surprising skills that we owe to our brain have inspired fields like Computational Neuroscience, Machine Learning and AI. Computational neuroscientists are trying to gain a better understanding of the brain via mathematical modelling of its functions. Machine Learning and AI specialists aim to create machines able to perform accurately on those tasks where only humans excel. A common denominator in these fields are Artificial Neural Networks (ANNs), computational models (loosely) inspired by the biology of the brain. Especially in visual tasks, like object classification, ANNs became very useful and popular: on one hand because they are able to (broadly) predict and match neuronal patterns in the visual cortex [4], [5], [13], on the other because they reach human accuracy in performance [25]. Unfortunately, each of these two properties does not imply the other. In fact, there can be ANNs resembling our brain firing patterns without solving the task and, conversely, neural networks exceeding human performance but not very informative about the brain. In this work we want to examine the population code of a state of the art ANN for computer vision. At first we test the robustness of the network against Poisson noise. Successively, we test the Efficient Coding hypothesis on the inner activations of the network by finding sparse representations and analyzing their characteristics