The X? Project Recent advances we have made in parallel computing architecture and natural language theory enhance the capability to process speech through the combined use of powerful concurrent processing and AI parsing technology. We propose to investigate the use of novel computing architechtures together with advanced phonological processing techniques to recognize ordinary spoken utterances in real-time, to learn to recognize novel inputs, and more generally, to gain a deeper understanding of parallel design methodologies. Background Information processing in biological systems exhibit two fundamental properties that concisely characterize the problem of recognizing speech with computer systems. One is fault tolerant behavior both with respect to internal failures and to input data errors, allowing the system to operate with distorted or incomplete inputs. Another is plasticity, or the ability to adapt to changes in the environment, with the associated mechanism of selection out of degenerate initial configurations. Yet a third property is characteristic only of intelligent systems. This is representational encoding, where input data from the environment is subject to further cognitive processing that may be rooted in either evolutionary behavior or arbitrary convention. Parallel computation is a promising development in areas ranging from numerical methods to robotics. Advances in VLSI technology have made it possible to construct concurrent machines based on a regular array of locally connected elements. We have recently shown that besides their useful applications, such machines have the potential of exhibiting behavior characterized by self-organization, learning, and recognition. If coupled with a system that provides of formal representation of knowledge, such machines prove capable of coping with the registration problem in cognitive systems, where recognition must be assumed to operate over an arbitrarily large set of inputs and exhibit fault-tolerant behavior and plasticity. Knowledge-based systems get around the uncertainty of information in various ways, typically through a process of "compartmentalization" which may involve cascade-like use of knowledge in a general domain. Using such a modular approach, we develop a representation of physical signals whereby we can use both non-linguistic situation-independent knowledge plus higher-level language-specific constraints in the information extraction process. We then consider arrays of simple local units that operate on integer data received locally from its neighbors. Input to the machine takes place only along the edges and the computation is systolically advanced from row to row in step with a global clock. Each processor has an internal state, represented by an integer, which can only take a small set of values depending on a given adaptive rule. The unit determines its local output based on its inputs and its internal state. At each time step, every element receives data values from the adjoining units in one direction and computes its output, which is then passed along to its next neighbors. We have recently shown that there is class of such architectures that can be made to compute in a distributed, deterministic , self-repairing fashion. In the language of dynamical systems, this corresponds to the appearance of fixed points in the phase space of the system. Arrays of such simple units achieve a certain amount of segmentation of the speech waveform, and perform a discrete recoding of the continuous input. The causal relation between language-independent representations and the physical signal is established in this procedure, which amounts to a qualitative labelling of the inputs. The signal is analyzed in a qualitative way, but, interestingly, linguistic constraints do not mediate this process. Instead, they are used in a logically subsequent one. Language-particular information processors evaluate the output of the arrays. Some outputs of the arrays may be discarded as non-meaningful in a given language, or accepted as meaningful only in a fixed context. The acoustic labels are then parsed by the segmental parser into linguistic objects. Using an active chart parser, the initial labelling over each pair of vertices vi and vi' is parsed into higher-level objects if appropriate constraints are satisfied. We analyze a sound without first knowing if it is part of a linguistic object by assigning categories based on the physical signal itself. Then, since we can formulate a grammar of acoustic events, we can use natural language techniques to go from the represented signal to higher linguistic levels. The research problem may be viewed as one of determining the classification of events in the physical signal, and then deciding on the appropriate parser that can map first from these to phonetic categories, then to phonological categories and finally to words. We seek to answer two questions: 1) How can we arrive at an appropriate representation of physical signals? 2) How do we compute with these representations to extract linguistic meaning? Plan While at present we already have components that provide solutions to some of the difficulties associated with partially obscured or incomplete information, we wish to systematically investigate how to improve them both from the point of view of the scientific model and from an efficiency perspective. In addition, we want to implement our software model in a special-purpose chip which would enable far greater processing speed. Some of the scientific questions which must be addressed include how to develop better adaptive structures for time-varying inputs, whether a two-tape finite-state automata will suffice to map the array outputs to the language-specific parser, how to isolate language-specific knowledge, and to switch natural language targets within the same machine. Furthermore, we would like to determine the smallest set of variables needed to characterize systems with high complexity. We envision a three-year effort in order to make substantial progress on these issues. Even if partly successful, we anticipate obtaining answers to a number of related problems such as pattern recognition and a deeper understanding of the complexity issues underlying parallel computation. ENCLOSURES B. A. Huberman and T. Hogg, "Adaptation and Self-Repair in Parallel Computing Structures," Phys. Rev. Lett. 52, 1048-1051 (1984). T. Hogg and B. A. Huberman, "Understanding Biological Computation: Reliable Learning and Recognition," Proc. Natl. Acad. Sci. USA 81, 6871-6875 (1984). WIthgott (International Joint Conference on Artificial Intelligence) F HELVETICA Fz¸