Search PubMed⌕ Search

SEARCH · Search PubMed

Results for “representation learning”

Search indexed PubMed citations on genomics, clinical trials, systematic reviews and public health. Explore titles, authors and supplied subject terms, then open the PubMed record.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,027 records · Page 57Linked to original sources
PubMed2002

Factorial hidden Markov models and the generalized backfitting algorithm.

Previous researchers developed new learning architectures for sequential data by extending conventional hidden Markov models through the use of distributed state representations. Although exact inference and parameter estimation in these architectures is computationally intractable, Ghahramani and Jordan (1997) showed that approximate inference and parameter estimation in one such architecture, factorial hidden Markov models (FHMMs), is feasible in certain circumstances. However, the learning algorithm proposed by these investigators, based on variational techniques, is difficult to understand and implement and is limited to the study of real-valued data sets. This chapter proposes an alternative method for approximate inference and parameter estimation in FHMMs based on the perspective that FHMMs are a generalization of a well-known class of statistical models known as generalized additive models (GAMs; Hastie & Tibshirani, 1990). Using existing statistical techniques for GAMs as a guide, we have developed the generalized backfitting algorithm. This algorithm computes customized error signals for each hidden Markov chain of an FHMM and then trains each chain one at a time using conventional techniques from the hidden Markov models literature. Relative to previous perspectives on FHMMs, we believe that the viewpoint taken here has a number of advantages. First, it places FHMMs on firm statistical foundations by relating them to a class of models that are well studied in the statistics community, yet it generalizes this class of models in an interesting way. Second, it leads to an understanding of how FHMMs can be applied to many different types of time-series data, including Bernoulli and multinomial data, not just data that are real valued. Finally, it leads to an effective learning procedure for FHMMs that is easier to understand and easier to implement than existing learning procedures. Simulation results suggest that FHMMs trained with the generalized backfitting algorithm are a practical and powerful tool for analyzing sequential data.

Algorithms↗
PubMed1990

Six children with lead poisoning.

Although most commonly seen in children from lower socioeconomic backgrounds, all children are at risk for lead poisoning. Management is a potential problem for all primary care providers. Because few individuals in the primary care practice of pediatrics have many patients with lead poisoning, it may be difficult to understand the nuances of management. We describe six patients, each of whom reflects different aspects of lead poisoning in children, and discuss the lessons we have learned in the course of their treatment. We have found that graphic representation of the blood lead and erythrocyte protoporphyrin values is helpful in the longitudinal follow-up of these youngsters.

Child, Preschool↗
PubMed2004

Acquisition and generalization of visuomotor transformations by nonhuman primates.

The kinematics of straight reaching movements can be specified vectorially by the direction of the movement and its extent. To explore the representation in the brain of these two properties, psychophysical studies have examined learning of visuomotor transformations of either rotation or gain and their generalization. However, the neuronal substrates of such complex learning are only beginning to be addressed. As an initial step in ensuring the validity of such investigations, it must be shown that monkeys indeed learn and generalize visuomotor transformations in the same manner as humans. Here, we analyze trajectories and velocities of movements as monkeys adapt to either rotational or gain transformations. We used rotations with different signs and magnitudes, and gains with different signs, and analyzed transfer of learning to untrained movements. The results show that monkeys can adapt to both types of transformation with a time course that resembles human learning. Analysis of the aftereffects reveals that rotation is learned locally and generalizes poorly to untrained directions, whereas gain is learned more globally and can be transferred to other amplitudes. The results lend additional support to the hypothesis that reaching movements are learned locally but can be easily rescaled to other magnitudes by scaling the peak velocity. The findings also indicate that reaching movements in monkeys are planned and executed very similarly to those in humans. This validates the underlying presumption that neuronal recordings in primates can help elucidate the mechanisms of motor learning in particular and motor planning in general.

Animals↗
PubMed2005

The role of long-term and short-term familiarity in visual and haptic face recognition.

Recent studies have suggested that the familiarity of a face leads to more robust recognition, at least within the visual domain. The aim of our study was to investigate whether face familiarity resulted in a representation of faces that was easily shared across the sensory modalities. In Experiment 1, we tested whether haptic recognition of a highly familiar face (one's own face) was as efficient as visual recognition. Our observers were unable to recognise their own face models from tactile memory alone but were able to recognise their faces visually. However, haptic recognition improved when participants were primed by their own live face. In Experiment 2, we found that short-term familiarisation with a set of previously unfamiliar face stimuli improved crossmodal recognition relative to the recognition of unfamiliar faces. Our findings suggest that familiarisation provides a strong representation of faces but that the nature of the information encoded during learning is critical for efficient crossmodal recognition.

Adult↗
PubMedDate not supplied

Roles of visual experience and intrinsic mechanism in the activity-dependent self-organization of orientation maps: theory and experiment.

It is widely accepted that functional maps in the mammalian visual cortex such as ocular dominance columns and orientation columns are formed depending on neural activity. There is still, however, controversy on how much visual experience contributes to the map formation during development. In the present study, we address this issue from mathematical modeling and experimental investigation. Using a model of activity-dependent self-organization of geniculo-cortical afferent inputs, we showed that spontaneous activity in the LGN can produce orientation maps, while the exposure to drifting gratings results in sharply segregated orientation maps as observed in cat visual cortex. The restricted exposure to a single orientation of the grating led to the over-representation of the exposed orientation, which was moderated by the contribution of learning based on the spontaneous activity. These theoretical results were confirmed by intrinsic optical recordings from area 18 of kittens reared under various visual conditions.

Animals↗
PubMed1999

A working memory "theory of relativity": elasticity in temporal, spatial, and modality dimensions conserves item capacity in radial maze, verbal tasks, and other cognition.

It is remarkable that working memory (WM) capacity for numbers of items remains modest, at approximately 7+/-2 (the so-called "magical number"), across a wide variety of kinds of material. Indeed, consideration of radial maze studies together with more traditional memory research shows that WM capacity remains fairly constant whether the items are verbal or visuospatial, and that this same capacity is true of other species as of humans. In contrast to their limited numerousness, WM items are extremely flexible in ways that are here brought under the heading of "dimensionality." Therefore, the physical items represented in WM, can vary widely in any quantitative characteristic and in the temporal pace at which they are encountered. Combinatorial considerations suggest that WM numerousness results from evolution of a middle ground between a sterile parsimony and an overwhelming excess, for organizing neurocognitive associations. Such natural selection seems likely to have worked opportunistically to yield diverse characteristics of neuronal tissue, from subcellular components to properties of ensembles, which converge on the required cognitive properties of WM. Priming and implicit memory may play supporting roles with WM. These intermediate-term memory phenomena allow certain kinds of background information to be accumulated at higher volume than seems possible from the textbook, "modal model" of memory. By expediting attentional focus on subsets of information already in long-term memory, priming may help WM chunks to emerge in limited number as appropriately scaled "figures" from the primed "ground." The larger neuronal dynamic patterns that embody these cognitive phenomena must regulate their microscopic component systems, automatically selecting those having parameters of temporal persistence, rhythm, and connectivity patterns that are pertinent to the current task. Relevant neural phenomena may include "Hebbian" associativity and persistence of firing patterns in prefrontal or hippocampal neurons. A conceivable basis for scaling and normalizing WM representations, along arbitrarily long or short ranges of any cognitive dimension, involves harmonic multiplier relationships among brain electrical rhythms and/or among topographical spatial periodic representations.

Animals↗
PubMed1991

Composites, compromises, and CHARM: what is the evidence for blend memory representations?

Metcalfe's (1990) distributed memory model simulates many misinformation effects by assuming representations that superimpose information from multiple sources. In the present article, two types of evidence are reviewed for such "blend" representations: composite recollections, including items from both the original and postevent sources (e.g., a previously seen intersection is remembered with a subsequently suggested stop sign), and compromise recollections, including features that cannot be exclusively associated with either source (e.g., a green car that was later suggested to be blue is remembered as bluish green). The considerable evidence for composite recollections provides little support for blend representations. Compromise recollections, though seemingly more persuasive, are both rare and interpretable without postulating blend representations. Speculation is made about potential findings that would support blend representations.

Association Learning↗
PubMed1991

Morphine's effects on differential serial compound conditioning and reflex modification of the rabbit's (Oryctolagus cuniculus) nictitating membrane response.

This study sought to determine the effects of morphine (0, 2, and 5 mg/kg) on (a) differential classical conditioning of the rabbit's (Oryctolagus cuniculus) nictitating membrane response (NMR) to the serial compounds A-X-unconditioned stimulus (US) and B-X-US (Experiment 1) and (b) the reflex modification effects of the compounds and their components (Experiment 2). These experiments determined specifically morphine's effects on the distinctiveness and time course of stimulus representations by examining morphine's dose-response effect on (a) differential responding to A and B and their conditional control over responding to X within the compounds and (b) the unconditioned excitatory effects of the compounds and their components as assessed by their ability to modify the amplitude of the unconditioned NMR. The results of these experiments indicate that morphine, in a dose-dependent manner, can operate to profoundly attenuate the distinctiveness and persistence (short-term memory) of stimulus representations.

Animals↗
PubMed1998

Medieval theories of mental representation.

Throughout most of the Middle ages, it was generally held that stored mental representations of perceived objects or events preserved the forms or species of such objects. This belief was consistent with a metaphor used by Plato. It was also consistent with the medieval belief that a number of cognitive processes took place in the ventricles of the brain and with the phenomenology of afterimages and imagination itself. In the 14th century, William of Ockham challenged this belief by claiming that mental representations are not stored but instead constructed in the basis of past learned experiences.

Brain↗
PubMed2003

Examining the interactivity of lexical orthographic and phonological processing.

The number and type of connections involving different levels of orthographic and phonological representations differentiate between several models of spoken and visual word recognition. At the sublexical level of processing, Borowsky, Owen, and Fonos (1999) demonstrated evidence for direct processing connections from grapheme representations to phoneme representations (i.e., a sensitivity effect) over and above any bias effects, but not in the reverse direction. Neural network models of visual word recognition implement an orthography to phonology processing route that involves the same connections for processing sublexical and lexical information, and thus a similar pattern of cross-modal effects for lexical stimuli are expected by models that implement this single type of connection (i.e., orthographic lexical processing should directly affect phonological lexical processing, but not in the reverse direction). Furthermore, several models of spoken word perception predict that there should be no direct connections between orthographic representations and phonological representations, regardless of whether the connections are sublexical or lexical. The present experiments examined these predictions by measuring the influence of a cross-modal word context on word target discrimination. The results provide constraints on the types of connections that can exist between orthographic lexical representations and phonological lexical representations.

Attention↗
PubMed2000

Long-lasting cortical plasticity in the object naming system.

A single exposure to an object can produce long-lasting behavioral change. Here, using event-related functional magnetic resonance imaging (fMRI), we provide evidence for long-lasting changes in cortical activity associated with perceiving and naming objects. In posterior regions, we observed an immediate (30-second) and long-lasting (3-day) decrease in neural activity after brief (200-ms) exposure to nameable and nonsense objects. In addition, slower-developing decreases in left inferior frontal activity were observed concurrently with increases in left insula activity, only for nameable objects. These time-dependent cortical changes may reflect two distinct learning mechanisms: the formation of sparser, yet more object-form-specific, representations in posterior regions, and experience-induced reorganization of the brain circuitry underlying lexical retrieval in anterior regions.

Cerebral Cortex↗
PubMed2003

Modulation of neural activity during object naming: effects of time and practice.

Repeated exposure to objects improves our ability to identify and name them, even after a long delay. Previous brain imaging studies have demonstrated that this experience-related facilitation of object naming is associated with neural changes in distinct brain regions. We used event-related functional magnetic resonance imaging (fMRI) to examine the modulation of neural activity in the object naming system as a function of experience and time. Pictures of common objects were presented repeatedly for naming at different time intervals (1 h, 6 h and 3 days) before scanning, or at 30 s intervals during scanning. The results revealed that as objects became more familiar with experience, activity in occipitotemporal and left inferior frontal regions decreased while activity in the left insula and basal ganglia increased. In posterior regions, reductions in activity as a result of multiple repetitions did not interact with time, whereas in left inferior frontal cortex larger decreases were observed when repetitions were spaced out over time. This differential modulation of activity in distinct brain regions provides support for the idea that long-lasting object priming is mediated by two neural mechanisms. The first mechanism may involve changes in object-specific representations in occipitotemporal cortices, the second may be a form of procedural learning involving a reorganization in brain circuitry that leads to more efficient name retrieval.

Adult↗
PubMed2001

Annotation: the cognitive neuroscience of face recognition: implications for developmental disorders.

Face recognition is often considered to be a modular (encapsulated) function. This annotation supports the proposal that faces are special, but suggests that their identification makes use of general-purpose cortical systems that are implicated in high-level vision and also in memory and learning more generally. These systems can be considered to function within two distinct cortical streams: a medial stream (for learning and salience of faces encountered) and a lateral stream (for distributed representations of visual properties and identities of faces). Function in the lateral stream, especially, may be critically dependent on the normal development of magnocellular vision. The relevance of face recognition anomalies in three developmental syndromes (Autism, Williams syndrome, and Turner syndrome) and the two-route model sketched above is considered.

Autistic Disorder↗
PubMed1990

Sensorimotor and action development in autistic children from infancy to early childhood.

Home movies of eight normal children and eight children later diagnosed as autistic were analysed according to development in sensorimotor and action competencies. The age range covered was from 4 to 42 months of age. From the beginning of the second year of life, the timing and sequence of developmental gains in normal and autistic children show striking differences. One year later the development of the autistic children seems to be not only delayed, but qualitatively different with respect to the more challenging kinds of action outcomes. This is particularly noticeable for action outcomes involving mental representations of the intended outcome.

Association Learning↗
PubMed2003

Muscarinic blockade slows and degrades the location-specific firing of hippocampal pyramidal cells.

The firing of rat hippocampal pyramidal cells is determined both by the animal's location and by the state of the hippocampal EEG. Because cholinergic transmission plays a role in EEG activity, we expected that its modification would alter place cell activity. We therefore investigated the effects on place cell activity of blocking muscarinic transmission with intracerebroventricular injections of scopolamine. Scopolamine reduced both the rate of place cell discharge inside firing fields and the spatial coherence of the fields; discharge outside of the fields also showed small increases. After injections, fields were shifted farther from their previous location than for saline controls, indicating reduced reproducibility after muscarinic blockade. Scopolamine increased the time rats were stationary, but changes in place cell activity persisted even after analysis was restricted to periods of walking, suggesting that the behavioral changes cannot account for the cell discharge changes. The scopolamine effects were dose dependent to an extent that varied between different measures. The firing rates of interneurons showed only a minor trend to decrease after scopolamine. Nevertheless, the spatial coherence of interneuron firing patterns was reduced, consistent with the recent demonstration that their positional firing is mediated by the location-specific firing of pyramids (Marshall et al., 2002). These results demonstrate that acetylcholine enhances positional firing patterns in the hippocampus. Muscarinic blockade weakens the positional firing of most place cells and therefore renders them less useful for precise representation of the environment. This effect may underlie the difficulties in spatial learning and problem solving caused by abnormalities of cholinergic transmission.

Acetylcholine↗
PubMed1981

The representation of pictures in memory.

The two experiments presented in this article examined the memory representation of pictorial information. The technique used to investigate structure was priming in item recognition. Subjects studied a list of pictures and then were tested for recognition of parts of pictures. In Experiment 1, the time to recognize a target part of a picture was speeded (primed) if the immediately preceding part in the test list was from the same picture. This priming effect was larger if the two parts were interacting with each other in the picture than if they were not interacting. Experiment 2 showed more priming between the interacting, foreground parts of a picture than between one of the interacting parts and a background part. For noninteracting parts, priming between foreground parts was equal to priming between foreground and background parts. It is suggested that priming may prove a useful technique for investigating other aspects of the representation of pictorial information.

Adult↗
PubMed1981

[Voluntary control of the activity of the individual motor units of different muscles].

16 subjects learned in feedback experiments to control activity of separate motor units (MU) in muscles of face, neck, hand, forearm, arm, thigh and crus. The test consisted in modulation of frequency and duration of the MU firing. The time total of the test, number of MU acting correctly, ratio of the "controllable" units to the number of units studied in a particular muscle, were recorded. MUs of the same muscle were found to be controllable to a different extent. All the muscles hand unit resisting learning for any period of time. No correlation between the muscle cortical representation and the percentage of the "controllable" units in it, was found.

Adolescent↗