TY - JOUR T1 - Deep neural networks: a new framework for modelling biological vision and brain information processing JF - bioRxiv DO - 10.1101/029876 SP - 029876 AU - Nikolaus Kriegeskorte Y1 - 2015/01/01 UR - http://biorxiv.org/content/early/2015/10/26/029876.abstract N2 - Recent advances in neural network modelling have enabled major strides in computer vision and other artificial intelligence applications. Human-level visual recognition abilities are coming within reach of artificial systems. Artificial neural networks are inspired by the brain and their computations could be implemented in biological neurons. Convolutional feedforward networks, which now dominate computer vision, take further inspiration from the architecture of the primate visual hierarchy. However, the current models are designed with engineering goals and not to model brain computations. Nevertheless, initial studies comparing internal representations between these models and primate brains find surprisingly similar representational spaces. With human-level performance no longer out of reach, we are entering an exciting new era, in which we will be able to build neurobiologically faithful feedforward and recurrent computational models of how biological brains perform high-level feats of intelligence, including vision.Unitmodel abstraction of a neuron, typically computing a weighted sum of incoming signals, followed by a static nonlinear transformation.Feedforward networknetwork whose connections form a directed acyclic graph, precluding recurrent information flow.Recurrent networknetwork with recurrent information flow, which produces dynamics and lends itself naturally to the perception and generation of spatiotemporal patterns.Convolutional networknetwork where a layer’s preactivation (before the nonlinearity) implements convolutions of the previous layer with a number of weight templates.Deep neural networknetwork with more than one hidden layer between input and output layers; more loosely, network with many hidden layers.Deep learningmachine learning of complex representations in a deep neural network, typically using stochastic gradient descent by error backpropagation.Universal function approximatormodel family that can approximate any function mapping input patterns to output patterns (with arbitrary precision when allowed enough parameters)Universal approximator of dynamical systemsa model family generating dynamics that can approximate any dynamical system (with arbitrary precision when allowed enough parameters)Maxpoolingsummary operation implementing invariances by retaining only the maxima of sets of detectors differing in irrelevant properties (e.g. local position).Normalisationoperation (e.g. division) applied to a set of activations so as to hold fixed a summary statistic (e.g. the sum).Dropoutregularisation method for neural network training with each unit omitted from the architecture with probability 0.5 on each training trial.Graphics processing unit (GPU)specialised computer hardware developed for graphics computations that greatly accelerates matrix-matrix multiplications and is essential for efficient deep learning.Supervised learninglearning process requiring input patterns along with additional information about the desired representation or the outputs (e.g. category labels).Unsupervised learninglearning process that requires only a set of input patterns and captures aspects of the probability distribution of the inputs.Backpropagation (and backpropagation through time)supervised neural-net learning algorithm that backpropagates error derivatives with respect to the weights through the connectivity to iteratively minimise errors.Generative modelmodel of the process that generated the data (e.g. the image), to be inverted in data analysis (e.g. visual recognition).Discriminative modelmodel extracting information of interest from the data (e.g. the image) without explicitly representing the process that generated the dataReceptive field modellingpredictive modelling of the response to arbitrary sensory inputs of neurons (or measured channels of brain activity).Representational similarity analysismethod for testing computational models of brain information processing through statistical comparisons of representational distance matrices that characterise population-code representations.Synthetic neurophysiologycomputational analysis of responses and dynamics of artificial neural nets aimed to gain a higher-level understanding of their computational mechanisms. ER -