Summary
The goal of computational neuroscience is to find mechanistic explanations of how the nervous system processes information to support cognitive function and behaviour. At the heart of the field are its models, i.e. mathematical and computational descriptions of the system being studied. These models typically map sensory stimuli to neural responses and/or neural to behavioural responses and range for simple to complex. Recently, deep neural networks (DNNs), using either feedforward and recurrent architectures, have come to dominate several domains of artificial intelligence (AI). As the term “neural network” suggests, these models are inspired by biological brains. However, current DNN models abstract from many details of biological neural networks. Their abstractions contribute to their computational efficiency, enabling to perform complex feats of intelligence, ranging from perceptual tasks (e.g. visual object and auditory speech recognition) to cognitive tasks (e.g. machine translation), and on to motor control tasks (e.g. playing computer games or controlling a robot arm). In addition to their ability to model complex intelligent behaviours, DNNs have been shown to predict neural responses to novel sensory stimuli that cannot be predicted with any other currently available type of model. DNNs can have millions of parameters (connection strengths), which are required to capture the domain knowledge needed for task performance. These parameters are often set by task training using stochastic gradient descent. The computational properties of the units are the result of four directly manipulable elements: input statistics, network structure, functional objective, and learning algorithm. The advances with neural nets in engineering provide the technological basis for building task-performing models of varying degrees of biological realism that promise substantial insights for computational neuroscience.