ABSTRACT
Neural network hardware is sometimes claimed to be inspired by the design of the brain, that is, it is neuromorphic. However, the operations of the resulting systems only occasionally act psychomorphic, that is, working like the mind. In this article we point out some places where neural network technology must be significantly extended if it is to act more like minds. (1) There is little understanding of intermediate level organization above the level of single units and below the level of the entire system. (2) The theoretical formulation of neural network learning needs to advance beyond 1920s behaviorism. (3) Flexibility of operation and control of the direction of a computation are probably more important to behavior than retrieval accuracy. (4) Neural networks are almost always special purpose devices. Successful system performance lies in the details of the architecture and the data representation.
