Richard Granger

Incremental Learning from Noisy Data

Induction of a concept description given noisy instances is difficult and is further exacerbated when the concepts may change over time. This paper presents a solution which has been guided by psychological and mathematical results. The method is based on a distributed concept description which is composed of a set of weighted, symbolic characterizations. Two learning processes incrementally modify this description. One adjusts the characterization weights and another creates new characterizations. The latter process is described in terms of a search through the space of possibilities and is shown to require linear space with respect to the number of attribute-value pairs in the description language. The method utilizes previously acquired concept definitions in subsequent learning by adding an attribute for each learned concept to instance descriptions. A program called STAGGER fully embodies this method, and this paper reports on a number of empirical analyses of its performance. Since understanding the relationships between a new learning method and existing ones can be difficult, this paper first reviews a framework for discussing machine learning systems and then describes STAGGER in that framework.

Enhancement by an Ampakine of Memory Encoding in Humans

Acentrally active drug that enhances AMPA receptor-mediated currents was tested for its effects on memory in humans. Evidence for a positive influence on encoding was obtained in four tests: (i) visual associations, (ii) recognition of odors, (iii) acquisition of a visuospatial maze, and (iv) location and identity of playing cards. The drug did not improve scores in a task requiring cued recall of verbal information. The selectivity of drug effects on memory was confirmed using tests of visual recognition, motor performance, and general intellectual functioning. These results suggest that positive modulators of AMPA receptors selectively improve at least some aspects of memory.

Evidence That a Positive Modulator of AMPA-Type Glutamate Receptors Improves Delayed Recall in Aged Humans

Elderly subjects (65-76 years) were tested for recall of nonsense syllables prior to and after oral administration of 1-(quinoxalin-6 ylcarbonyl)piperidine (CX516), a centrally active drug that enhances currents mediated by AMPA-type glutamate receptors. A significant and positive drug effect was found for delayed (5 min) recall at 75 min posttreatment; average scores for the highest dose group were more than twofold greater than for the placebo group. The drug had no evident influence on heart rate or self-assessment of several psychological variables.

Facilitation of glutamate receptors reverses an age‐associated memory impairment in rats

The accuracy of memory for recent events is reported to decay between young adulthood and middle age in humans (Crook et al., 1990; Crook and West, 1990; Thomas et al., 1977) due to impairments in acquisition and/or retention (Craik, 1977; Huppert and Kopelman, 1989). Effects of this kind are also found in comparisons of middle‐aged (12–18 months) vs. young adult (3 months) rats in tests requiring retention of recently sampled spatial cues (Kadar et al., 1990a; Kadar et al., 1990b; Goudsmit et al., 1990; Weiss and Thompson, 1991). The causes of such changes in memory processing are unknown but might be expected to involve age‐related losses in forebrain glutamate receptors (Bahr et al., 1992; Magnusson and Cotman, 1993; Wenk et al., 1991); these receptors mediate fast excitatory transmission in many brain regions and play an essential role in the production of long‐term potentiation (LTP), a form of synaptic plasticity that has been implicated in memory encoding (Landfield and Lynch, 1977; Moore et al., 1993). In the present communication we report results indicating that a drug that enhances AMPA‐type glutamate receptors acts centrally to selectively increase hippocampal spatial cell firing and improves both acquisition performance and memory retention in middle‐aged rats to levels equivalent to those found in young adult animals.

Dissociation of hippocampal and entorhinal function in associative learning: A computational approach

Unsupervised stimulus-stimulus redundancy compression, one component of Gluck and Myers’s (1993) representational theory of the hippocampal-region function, could emerge from the anatomy and physiology of the entorhinal cortex. This hypothesis is suggested by a physiologically and anatomically realistic model of the entorhinal cortex derived from a similar model of the olfactory cortex previously proposed by Ambros-Ingerson, Granger, and Lynch (1990). To the extent that entorhinal function can survive damage strictly limited to the hippocampal formation (the H lesion), this has implications for interpreting the behavioral consequences of lesions which either do or do not spare overlying cortical areas. In particular, we expect that the H lesion should not interrupt stimulus-stimulus redundancy compression, thereby sparing conditioning behaviors, such as latent inhibition, which are eliminated by broader (H++) lesions to the hippocampal region. However, such other behaviors as the context sensitivity of latent inhibition and of learned responses are expected to be affected by the H lesion. These predictions are consistent with empirical data. The theory also leads to several novel predictions for behavioral comparisons of intact, H-lesioned, and H++-lesioned animals on tasks such as sensory preconditioning, compound preconditioning, and easy-hard transfer. A major theme of this paper is to illustrate how a bottom-up model of cortical processing can be integrated with a top-down model of hippocampal-region function to yield a more complete mapping from physiology to behavior.

Derivation of encoding characteristics of layer ii cerebral cortex

Computer simulations of layers I and II of pirifonn (olfactory) cortex indicate that this biological network can generate a series of distinct output responses to individual stimuli, such that different responses encode different levels of information about a stimulus. In particular, after learning a set of stimuli modeled after distinct groups of odors, the simulated networks initial response to a cue indicates only its group or category, whereas subsequent responses to the same stimulus successively subdivide the group into increasingly specific encoding of the individual cue. These sequences of responses amount to an automated organization of perceptual memories according to both their similarities and differences, facilitating transfer of learned information to novel stimuli without loss of specific information about exceptions. Human recognition performance robustly exhibits such multiple levels: a given object can be identified as a vehicle, as an automobile, or as a Mustang. The findings reported here suggest that a function as apparently complex as hierarchical recognition memory, which seems suggestive of higher cognitive processes, may be a fundamental intrinsic property of the operation of this single cortical cell layer in response to naturally-occurring inputs to the structure. We offer the hypothesis that the network function of superficial cerebral conical layers may simultaneously acquire and hierarchically organize information about the similarities and differences among perceived stimuli. Experimental manipulation of the simulation has generated hypotheses of direct links between the values of specific biological features and particular attributes of behavior, generating testable physiological and behavioral predictions.

Short-latency single unit processing in olfactory cortex

Single-unit recording of layer IIIII cells in olfactory (piriform) cortex was performed on awake, unrestrained rats actively engaged in learning novel odors in an olfactory discrimination task. Five of the 67 cells tested had very brief monophasic action potentials and high spontaneous firing rates (3080 Hz); it is suggested that these units were interneurons. The remainder of the neurons had broader spikes and did not discharge for prolonged periods. Thirty-nine percent of the broad spike cells responded to at least one and usually more of the odors presented to the rats during either of the first two trials on which that odor was present, but, in most cases, these responses occurred only very infrequently over the course of subsequent trials. Six percent of the broad-spike group, how ever, continued firing robustly to a single odor but not to others. From these results it appears that most cells in piriform cortex do not respond to most odors, i.e., coding is exceedingly sparse. A subgroup of the predominant broad-spike cell type does react to several odors but this response drops out with repeated exposure, perhaps because of training. However, a few members of this class (a small fraction of the total cell population) do go on responding to a particular odor, thus exhibiting a form of odor specificity. The results are discussed with regard to predictions from recently developed models of the olfactory cortex.

Derivation and Analysis of Basic Computational Operations of Thalamocortical Circuits

Shared anatomical and physiological features of primary, secondary, tertiary, polysensory, and associational neocortical areas are used to formulate a novel extended hypothesis of thalamocortical circuit operation. A simplified anatomically based model of topographically and nontopographically projecting (core and matrix) thalamic nuclei, and their differential connections with superficial, middle, and deep neocortical laminae, is described. Synapses in the model are activated and potentiated according to physiologically based rules. Features incorporated into the models include differential time courses of excitatory versus inhibitory postsynaptic potentials, differential axonal arborization of pyramidal cells versus interneurons, and different laminar afferent and projection patterns. Observation of the models responses to static and time-varying inputs indicates that topographic core circuits operate to organize stored memories into natural similarity-based hierarchies, whereas diffuse matrix circuits give rise to efficient storage of time-varying input into retrievable sequence chains. Examination of these operations shows their relationships with well-studied algorithms for related functions, including categorization via hierarchical clustering, and sequential storage via hash or scatter-storage. Analysis demonstrates that the derived thalamocortical algorithms exhibit desirable efficiency, scaling, and space and time cost characteristics. Implications of the hypotheses for central issues of perceptual reaction times and memory capacity are discussed. It is conjectured that the derived functions are fundamental building blocks recurrent throughout the neo cortex, which, through combination, gives rise to powerful perceptual, motor, and cognitive mechanisms.

Psychological effects of a drug that facilitates brain Ampa receptors

The effects of l-(quinovalin-6-ylcarbonyl)piperidine (CX516), a centrally active compound that facilitates AMPA receptor-mediated synaptic responses, were tested in human subjects. Separate tests of delayed recall were given prior to and nearly 3 h after administration of placebo (n = 12) or drug (n = 36). Control subjects exhibited poorer performance in the second session than in the first while subjects given 600–1200 mg of the drug did not. There were no pre- vs post-treatment differences in immediate recall in either group. The drug did not reliably affect self-assessment scores for any of several psychological variables but did disrupt the normally present correlations for within-subject changes in the variables. These results suggest that AMPA receptor modulators may (1) improve memory under some circumstances and (2) produce psychological effects that are subtle or not related to specific mood states.

Variations in synaptic plasticity and types of memory in corticohippocampal networks

If Synaptic long-term potentiation (LTP) represents a memory storage mechanism, its induction and expression characteristics may constitute rules governing encoding and read-out of memory in cortical circuitry, The presence of variants of the LTP effect in different anatomical networks provides grounds for predictions about the types of memory operations to which potentiation contributes. Computer modeling studies incorporating the complex rules for LTP induction and the characteristics of expressed potentiation can be used to make such predictions specific. We review ttie types of synaptic plasticity found in the successive stages of the corticohippocampal pathway, and present results indicating that LTP does participate in definably different forms of memory, suggesting a classification of memory types differing somewhat from categories deduced from behavioral studies. Specifically, the results suggest that subtypes of memory operate serially, in an assembly line of specialized functions, each of which adds a unique aspect to the processing of memories. The effects of lesions on the encoding versus expression of memory can be interpreted from the perspective of this hypothesis.