A fundamental issue in the study of human cognition is what computations are carried out by the brain to implement cognitive processes. The connectionist framework assumes that cognitive processes are implemented in terms of complex, nonlinear interactions among a large number of simple, neuron-like processing units that form a neural network. This approach has been used in cognitive psychology - often with some success – to develop functional models that clearly represent a great advance over previous verbal-diagrammatic models because they can produce highly detailed simulations of human skilled performance and its breakdown following brain damage. However, a crucial step for the computational modeling of cognition is to bridge the gap between function and structure. Much of the modeling work has been carried out using connectionist networks that have no biological plausibility beyond the metaphor of “neuron-like” processing. Most models have one, or more often a combination, of the following undesirable features: i) strictly feed-forward spread of activation (e.g., no feedback and/or lateral connections); ii) implausible learning procedures (e.g., error back-propagation); iii) implausible learning environment (e.g., supervised learning). Researchers have chosen to ignore these problems as it was seen as an essential compromise to achieve efficient learning of complex cognitive tasks. The aim of the present research program is to exploit the latest findings in neural network and machine learning research to develop generative connectionist models of cognition. Generative models are appealing because they represent plausible models of cortical learning that emphasize the mixing of bottom-up and top-down interactions in the brain. Moreover, generative models of cognition would offer a unified theoretical framework that encompasses classic connectionism and the emerging Bayesian approach to cognition, as well as a means to bridge the gap between neurons and behavior.