Hippocampus, cortex, and basal ganglia: Insights from computational models of complementary learning systems Hisham E. Atallah, Michael J. Frank, and Randall C. O’Reilly * Department of Psychology, Center for Neuroscience, University of Colorado at Boulder, 345 UCB, Boulder, CO 80309, USA Received 16 April 2004; revised 4 June 2004; accepted 8 June 2004 Available online 20 July 2004 Abstract We present a framework for understanding how the hippocampus, neocortex, and basal ganglia work together to support cognitive and behavioral function in the mammalian brain. This framework is based on computational tradeoffs that arise in neural network models, where achieving one type of learning function requires very different parameters from those necessary to achieve another form of learning. For example, we dissociate the hippocampus from cortex with respect to general levels of activity, learning rate, and level of overlap between activation patterns. Similarly, the frontal cortex and associated basal ganglia system have im- portant neural specializations not required of the posterior cortex system. Taken together, this overall cognitive architecture, which has been implemented in functioning computational models, provides a rich and often subtle means of explaining a wide range of behavioral and cognitive neuroscience data. Here, we summarize recent results in the domains of recognition memory, contextual fear conditioning, effects of basal ganglia lesions on stimulus–response and place learning, and flexible responding. Ó 2004 Elsevier Inc. All rights reserved. Keywords: Computational models; Hippocampus; Basal ganglia; Neocortex 1. Introduction The brain is not a homogenous organ: different brain areas clearly have some degree of specialized function. There have been many attempts to specify what these functions are, based on a variety of theoretical ap- proaches and data. In this paper, we summarize our approach to this problem, which is based on the logic of computational tradeoffs in neural network models of brain areas. The core idea behind this approach is that different brain areas are specialized to satisfy funda- mental tradeoffs in the way that neural systems perform different kinds of learning and memory tasks. This way of characterizing the specializations of brain areas is in many ways consistent with ideas from other frame- works, but we argue that it offers a level of precision and subtlety that may prove beneficial in understanding complex interactions between different brain areas. This paper reviews a number of illustrations of this point, through applications of computational models to a range of data in both the human and animal literatures, including: recognition memory, contextual fear condi- tioning, effects of basal ganglia lesions on stimulus–re- sponse and place learning, and flexible responding. One of the central tradeoffs behind our approach involves the process of learning novel information rap- idly without interfering catastrophically with prior knowledge. This form of learning requires a neural network with very sparse levels of overall activity (leading to highly separated representations), and a relatively high learning rate. These features are incom- patible with the kind of network that is required to acquire general statistical information about the envi- ronment, which needs highly overlapping, distributed representations with relatively higher levels of activity, and a slow rate of learning. The conclusion we have drawn from this mutual incompatibility is that the brain must have two different learning systems to perform these different functions, and this fits quite well with a * Corresponding author. Fax: 1-303-492-2967. E-mail address: oreilly@psych.colorado.edu (R.C. O’Reilly). 1074-7427/$ - see front matter Ó 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.nlm.2004.06.004 Neurobiology of Learning and Memory 82 (2004) 253–267 www.elsevier.com/locate/ynlme