Resilient machines through continuous self-modeling

Animals sustain the ability to operate after injury by creating qualitatively different compensatory behaviors. Although such robustness would be desirable in engineered systems, most machines fail in the face of unexpected damage. We describe a robot that can recover from such change autonomously, through continuous self-modeling. A four-legged machine uses actuation-sensation relationships to indirectly infer its own structure, and it then uses this self-model to generate forward locomotion. When a leg part is removed, it adapts the self-models, leading to the generation of alternative gaits. This concept may help develop more robust machines and shed light on self-modeling in animals.     

Pattern classification by adaptive machines

Man's intelligent behavior is due in part to his ability to select, classify, and abstract significant information reaching him from his environment by way of his senses. This function, pattern recognition, has become a major focus of research by scientists working in the field of artificial intelligence. At the lowest level, pattern recognition reduces to pattern classification, which consists of the separation, into desired classes, of groups of objects, sounds, odors, events, properties, and the like; the separations are based on sets of measurements made on the entities being classified. The pattern classifier is composed of a data filter and a categorizer. The data filter selects the distinguishing features and represents them as sets of real numbers; each set is termed a pattern. The categorizer assigns each pattern to one of several desired classes. Patterns can be represented geometrically as points in an n-dimensional space; the n coordinates of each point are the numerical values of the features selected to represent the pattern. A pattern classification system separates an n-dimensional space into regions, each of which ideally contains points of only one class. One method to effect this separation is by means of ldquo;trainablerdquo; categorizers-major components of adaptive machines. They consist of networks whose internal parameters are varied according to a set of fixed rules during a training cycle. A statistically large sample of known patterns are presented, one at a time, to the networks; internal corrections are made each time a pattern is erroneously classified. Classifica-tion performance tends to improve as the set of known patterns is cycled repetitively through the machine. Finally, the adequacy of adaptation is tested by a separate set of similar patterns which have not been used in the training process. A number of different machine organizations and training rules have been developed and are being applied successfully to numerous classification problems. More difficult recognition problems requiring the aid of logioal tests and analysis, search and association, use the digital computer programmed to supplement the functions of the adaptive classifier.     

Signal-processing machines at the postsynaptic density

Dendrites of individual neurons in the vertebrate central nervous system are contacted by thousands of synaptic terminals relaying information about the environment. The postsynaptic membrane at each synaptic terminal is the first place where information is processed as it converges on the dendrite. At the postsynaptic membrane of excitatory synapses, neurotransmitter receptors are attached to large protein "signaling machines" that delicately regulate the strength of synaptic transmission. These machines are visible in the electron microscope and are called the postsynaptic density. By changing synaptic strength in response to neural activity, the postsynaptic density contributes to information processing and the formation of memories.