Networks of this type have a single layer of units with a selective response for some range of the input variables.

Experiments were performed to reveal some of the computational properties of the human motor memory system. We show that as humans practice reaching movements while interacting with a novel mechanical environment, they learn an internal model of the inverse dynamics of that environment. The representation of the internal model in memory is such that there is interference when there is an attempt to learn a new inverse dynamics map immediately after an anticorrelated mapping was learned. We suggest that this interference is an indication that the same computational elements used to encode the first inverse dynamics map are being used to learn the second mapping. We predict that this leads to a forgetting of the initially learned skill. 1 Introduction In tasks where we use our hands to interact with a tool, our motor system develops a model of the dynamics of that tool and uses this model to control the coupled dynamics of our arm and the tool (Shadmehr and Mussa-Ivaldi 1994).

A number of studies have described neurons in the dorsal part of the medial superior temporal (MSTd) monkey cortex that respond best to large expanding/contracting, rotating, or shifting patterns (Tanaka et al., 1986; Duffy & Wurtz, 1991a). Recently Graziano et al. (1994) found that MSTd cell responses correspond to a point in a multidimensional space of spiral motions, where the dimensions are these motion types. Combinationsof these motions are generated as an animal moves through its environment, whichsuggests that area MSTd could playa role in optical flow analysis. When an observer moves through a static environment, a singularity in the flow field known as the focus of expansion may be used to determine the direction of heading (Gibson, 1950; Warren & Hannon, 1988). Previous computational models of MSTd (Lappe & Rauschecker, 1993; Perrone & Stone, 1994) have shown how navigational information related to heading may be encoded by these cells.

In financial market applications, it is typical to have limited amount of relevant training data, with high noise levels in the data. The information content of such data is modest, and while the learning process can try to make the most of what it has, it cannot create new information on its own. This poses a fundamental limitation on the 412 YaserS.

When training neural networks by the classical backpropagation algorithm thewhole problem to learn must be expressed by a set of inputs and desired outputs. However, we often have high-level knowledge about the learning problem. In optical character recognition (OCR), for instance, weknow that the classification should be invariant under a set of transformations like rotation or translation. We propose a new modular classification system based on several autoassociative multilayer perceptrons whichallows the efficient incorporation of such knowledge. Results are reported on the NIST database of upper case handwritten letters and compared to other approaches to the invariance problem. 1 INCORPORATION OF EXPLICIT KNOWLEDGE The aim of supervised learning is to learn a mapping between the input and the output space from a set of example pairs (input, desired output). The classical implementation in the domain of neural networks is the backpropagation algorithm. If this learning set is sufficiently representative of the underlying data distributions, one hopes that after learning, the system is able to generalize correctly to other inputs of the same distribution.

We prove the convergence of an actor/critic algorithm that is equivalent toQ-Iearning by construction. Its equivalence is achieved by encoding Q-values within the policy and value function of the actor andcritic. The resultant actor/critic algorithm is novel in two ways: it updates the critic only when the most probable action is executed from any given state, and it rewards the actor using criteria thatdepend on the relative probability of the action that was executed.

Many real world learning problems are best characterized by an interaction of multiple independent causes or factors. Discovering suchcausal structure from the data is the focus of this paper. Based on Zemel and Hinton's cooperative vector quantizer (CVQ) architecture, an unsupervised learning algorithm is derived from the Expectation-Maximization (EM) framework. Due to the combinatorial natureof the data generation process, the exact E-step is computationally intractable. Two alternative methods for computing theE-step are proposed: Gibbs sampling and mean-field approximation, and some promising empirical results are presented.

Existing recurrent net learning algorithms are inadequate. We introduce theconceptual framework of viewing recurrent training as matching vector fields of dynamical systems in phase space. Phasespace reconstructiontechniques make the hidden states explicit, reducing temporal learning to a feed-forward problem. In short, we propose viewing iterated prediction [LF88] as the best way of training recurrent networks on deterministic signals. Using this framework, we can train multiple trajectories, insure their stability, anddesign arbitrary dynamical systems. 1 INTRODUCTION Existing general-purpose recurrent algorithms are capable of rich dynamical behavior. Unfortunately,straightforward applications of these algorithms to training fully-recurrent networks on complex temporal tasks have had much less success than their feedforward counterparts. For example, to train a recurrent network to oscillate like a sine wave (the "hydrogen atom" of recurrent learning), existing techniques such as Real Time Recurrent Learning (RTRL) [WZ89] perform suboptimally. Williams& Zipser trained a two-unit network with RTRL, with one teacher signal. One unit of the resulting network showed a distorted waveform, the other only half the desired amplitude.

Reinforcement learning-the problem of getting an agent to learn to act from sparse, delayed rewards-has been advanced by techniques based on dynamic programming (DP). These algorithms compute a value function which gives, for each state, the minimum possiblelong-term cost commencing in that state. For the high-dimensional and continuous state spaces characteristic of real-world control tasks, a discrete representation ofthe value function is intractable; some form of generalization is required. A natural way to incorporate generalization into DP is to use a function approximator, rather than a lookup table, to represent the value function. This approach, which dates back to uses of Legendre polynomials in DP [Bellman et al., 19631, has recently worked well on several dynamic control problems [Mahadevan and Connell, 1990, Lin, 1993] and succeeded spectacularly on the game of backgammon [Tesauro, 1992, Boyan, 1992].

Each year people spend a huge amount oftime typing. The text people type typically contains a tremendous amount of redundancy due to predictable word usage patterns and the text's structure. This paper describes a neural network system call AutoTypist that monitors a person's typing and predicts what will be entered next. AutoTypist displays the most likely subsequent word to the typist, who can accept it with a single keystroke, instead of typing it in its entirety. The multi-layer perceptron at the heart of Auto'JYpist adapts its predictions of likely subsequent text to the user's word usage pattern, and to the characteristics of the text currently being typed. Increases in typing speed of 2-3% when typing English prose and 10-20% when typing C code have been demonstrated using the system, suggesting a potential time savings of more than 20 hours per user per year. In addition to increasing typing speed, AutoTypist reduces the number of keystrokes a user must type by a similar amount (2-3% for English, 10-20% for computer programs). This keystroke savings has the potential to significantly reduce the frequency and severity of repeated stress injuries caused by typing, which are the most common injury suffered in today's office environment.