SEARCH · PubMed Health

Results for “Network inference”

Explore indexed PubMed citations for clinical trials, systematic reviews and public health research. Read source abstracts and follow each citation to its original PubMed record.

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

At least 1,171 records · Page 65Linked to original sources

PubMed2001

Stochastic organization of output codes in multiclass learning problems.

The best-known decomposition schemes of multiclass learning problems are one per class coding (OPC) and error-correcting output coding (ECOC). Both methods perform a prior decomposition, that is, before training of the classifier takes place. The impact of output codes on the inferred decision rules can be experienced only after learning. Therefore, we present a novel algorithm for the code design of multiclass learning problems. This algorithm applies a maximum-likelihood objective function in conjunction with the expectation-maximization (EM) algorithm. Minimizing the augmented objective function yields the optimal decomposition of the multiclass learning problem in two-class problems. Experimental results show the potential gain of the optimized output codes over OPC or ECOC methods.

Algorithms↗

PubMed2006

Bayesian learning of sparse gene regulatory networks.

Differential equations (DEs) have been the most widespread formalism for gene regulatory network (GRN) modeling, as they offer natural interpretation of biological processes, easy elucidation of gene relationships, and the capability of using efficient parameter estimation methods. However, an important limitation of DEs is their requirement of O(d(2)) parameters where d is the number of genes modeled, which often causes over-parameterization for large d, leading to the over-fitting of data and dense parameter sets that are hard to interpret. This paper presents the first effort to address the over-parameterization problem by applying the sparse Bayesian learning (SBL) method to sparsify the GRN model of DEs. SBL operates on the parsimony principle, with the objective to reduce the number of effective parameters by driving the redundant parameters to zero. The resulting sparse parameter set offers three important advantages for GRN inference: first, the inferred GRNs are more plausible, since the biological counterparts are known to be sparse; second, gene relationships can be more easily elucidated from sparse sets than from dense sets; and third, the solutions become more optimal and consistent, due to the reduction in the volume of solution space. Experiments are conducted on the yeast Saccharomyces cerevisiae time-series gene expression data, in which known regulatory events related to the cell cycle G1/S phase are reliably reproduced.

Bayes Theorem↗

PubMed1988

A synergism of plagues: "planned shrinkage," contagious housing destruction, and AIDS in the Bronx.

Techniques adapted from population and community ecology, quantitative geography, and epidemiology are applied to ecosystem and environmental index data on the Bronx in an attempt to understand the origins and potential impacts of rampant spread of human immunodeficiency virus (HIV) and its sequelae of acquired immunodeficiency syndrome (AIDS) in that borough: Recent work by Drucker and Vermund (1987), ("Estimating Prevalence of Human Immunodeficiency Virus Infection in Urban Areas with High Rates of Intravenous Drug Abuse: A Model of the Bronx in 1987," Poster presented at the Third International Conference on AIDS, June 2, 1987) estimates HIV seroprevalence levels of from 8 to 21% among men of age 25-44 in the south Bronx, at this writing, comparable to the cities of Central Africa. It is found that the "South Bronx" process of fulminating, contagious urban decay which devastated the region in the 1970s, and its associated forced population migrations, spread intravenous drug abuse, the principal HIV vector in the Bronx, from a geographically contained center in the South-Central Bronx to a virtually borough-wide phenomenon. This has significantly complicated attempts to contain HIV infection, both by shredding the social networks which are the natural vehicles for education, and by vastly enlarging the area requiring intensive targeting. Since the "planned shrinkage" municipal service cuts which triggered the "South Bronx" burnout persist, and since levels of housing overcrowding now approach those of the early 1970s in the Bronx, it is expected that a new outbreak of contagious urban decay will occur, likely again dispersing population and seriously compromising any in-place HIV control strategies. If overt AIDS itself becomes a contributor to urban deterioration in overcrowded neighborhoods susceptible to "South Bronx" process, we could then see a nonlinear ecosystem coupling between AIDS, contagious urban decay, and population shift. Elementary mathematical models are provided. Thus, in striking contrast to the middle-class male homosexual community, successful control of HIV infection in the Bronx, and by inference in other devastated ghetto communities, seems predicated on quick reestablishment of demographic stability: The tools to make the tools for control must first be reconstructed. Necessary elements of any program toward this end are briefly outlined. AIDS in the Bronx and similar areas, like tuberculosis, seems increasingly a marker disease of extreme poverty, and again like tuberculosis, seems increasingly a marker disease of extreme poverty, and again like tuberculosis, may well form an important reservoir for further spread or resurg

Acquired Immunodeficiency Syndrome↗

PubMed2000

Cadaveric versus living donor kidney transplantation: a Medicare payment analysis.

BACKGROUND: We found previously that the clinical advantages of living donor (LD) renal transplantation lead to financial cost savings compared to either cadaveric donation (CAD) or dialysis. Here, we analyze the sources of the cost savings of LD versus CAD kidney transplantation. METHODS: We used United States Renal Data System data to merge United Network for Organ Sharing registry information with Medicare claims data for 1991-1996. Information was available for 42,868 CAD and 13,754 LD transplants. More than 5 million Medicare payment records were analyzed. We calculated the difference in average payments made by Medicare for CAD and LD for services provided during the first posttransplant year. RESULTS: Average total payments were $39,534 and $24,652 for CAD and LD, respectively (P<0.0001) during the first posttransplant year. The largest source of the difference in payments was in inpatient hospitals, representing $10,653.67 (P<0.0001). For patients who had Medicare as the primary payer, average transplant charges were significantly higher for CAD donation ($79,730 vs. $69,547, P<0.0001); average transplant payments demonstrated no statistical differences ($28,483 vs. $28,447, P = 0.858). Therefore, inferred profitability was significantly higher for LD. CONCLUSIONS: Medicare payments are remarkably lower for LD compared to CAD in every category. The single largest cost saving comes from inpatient hospital services. A portion of the savings from LD could be invested in programs to expand living kidney donation.

Cadaver↗

PubMed1997

Adaptive sampling in behavioral surveys.

Studies of populations such as drug users encounter difficulties because the members of the populations are rare, hidden, or hard to reach. Conventionally designed large-scale surveys detect relatively few members of the populations so that estimates of population characteristics have high uncertainty. Ethnographic studies, on the other hand, reach suitable numbers of individuals only through the use of link-tracing, chain referral, or snowball sampling procedures that often leave the investigators unable to make inferences from their sample to the hidden population as a whole. In adaptive sampling, the procedure for selecting people or other units to be in the sample depends on variables of interest observed during the survey, so the design adapts to the population as encountered. For example, when self-reported drug use is found among members of the sample, sampling effort may be increased in nearby areas. Types of adaptive sampling designs include ordinary sequential sampling, adaptive allocation in stratified sampling, adaptive cluster sampling, and optimal model-based designs. Graph sampling refers to situations with nodes (for example, people) connected by edges (such as social links or geographic proximity). An initial sample of nodes or edges is selected and edges are subsequently followed to bring other nodes into the sample. Graph sampling designs include network sampling, snowball sampling, link-tracing, chain referral, and adaptive cluster sampling. A graph sampling design is adaptive if the decision to include linked nodes depends on variables of interest observed on nodes already in the sample. Adjustment methods for nonsampling errors such as imperfect detection of drug users in the sample apply to adaptive as well as conventional designs.

Behavior↗

PubMed1995

A Bayesian framework for understanding texture segmentation in the primary visual cortex.

This paper presents a mathematical theory for understanding the computations involved in texture segmentation in the primary visual cortex. We propose that texture segmentation is a part of the early visual system's overall strategy to infer surfaces of objects in a visual scene. Based on this insight, we use the Bayesian inference paradigm to formulate the texture segmentation problem into a maximum a posteriori surface inference problem. The dynamical system for finding the optimal solution of this problem can be characterized by two concurrent and interactive processes: a gradual sharpening of the boundary signals and a simultaneous smoothing of the surface signals. The behavior of these dynamical processes was studied using both analytical and computational methods. We present some computational results and mathematical predictions. This theory suggests a novel framework for understanding the functional roles of the complex cells in the primary visual cortex.

Bayes Theorem↗

PubMed1990

Mitochondrial and allozyme genetics of incipient speciation in a landlocked population of Galaxias truttaceus (Pisces: Galaxiidae).

Galaxias truttaceus is found in coastal rivers and streams in south-eastern Australia. It spawns at the head of estuaries in autumn and the larvae spend 3 months of winter at sea before returning to fresh water. In Tasmania there are landlocked populations of G. truttaceus in a cluster of geologically young lakes on the recently glaciated Central Plateau. These populations have no marine larval stage and spawn in the lakes in spring. Speciation due to land locking is thought to be a frequent occurrence within Galaxias. To investigate the nature of the speciation event which may be occurring within lake populations of G. truttaceus we studied the mitochondrial DNA (mtDNA) and allozyme diversity of both lake and stream populations. Using the presence or absence of restriction sites recognized by 13 six-base restriction endonucleases, we found 58 mtDNA haplotypes among 150 fish collected from 13 Tasmanian and one south-east Australian mainland stream populations. The most parsimonious network relating the haplotypes by site loss or gain was starlike in shape. We argue that this arrangement is best explained by selection upon slightly beneficial mutations within the mitochondrial genome. Gene diversity analysis under Wright's island model showed that the populations in each drainage were not genetically subdivided. Only two of these stream haplotypes were found among the 66 fish analyzed from four lake populations. Despite the extreme lack of mtDNA diversity in lake populations, the observed nuclear DNA heterozygosity of 40 lake fish (0.10355) was only slightly less than that of 82 stream fish (0.11635). In the short time (3000-7000 years) that the lake fish have been landlocked, random genetic drift in a finite, stable-sized population was probably not responsible for the lack of mtDNA diversity in the lake populations. We infer the lake populations have probably experienced at least one, severe, but transitory bottleneck possibly induced by natural selection for life-history characters essential for survival in the lacustrine habitat. If speciation is occurring in the landlocked populations of G. truttaceus, then it may be driven by genetic transilience.

Animals↗

PubMed2006

Arl1p is involved in transport of the GPI-anchored protein Gas1p from the late Golgi to the plasma membrane.

The molecular mechanisms involved in the transport of GPI-anchored proteins from the trans-Golgi network (TGN) to the cell periphery have not been established. Arl1p is a member of the Arf-like protein (Arl) subfamily of small GTPases and is localized in the late Golgi. Although Arl1p is implicated in regulation of Golgi structure and function, no endogenous cargo protein that is regulated by Arl1p has been identified in yeast. In this study, we demonstrate that Arl1p is involved in the anterograde transport from the Golgi to the cell surface of the glycosylphosphatidylinositol (GPI)-anchored plasma-membrane-resident protein Gas1p, but not the cell-wall-localized GPI-anchored proteins Crh1p, Crh2p and Cwp1p, or non-GPI-anchored plasma membrane-protein Gap1p. We also show that regulators of Arl1p (Sys1p, Arl3p and Gcs1p) and an effector (Imh1p) all participate in the transport of Gas1p. Thus, we infer that the signaling cascade Sys1p-Arl3p-Arl1p-Imh1p specifically participates in the transport of a GPI-anchored protein from the late Golgi to the plasma membrane.

ADP-Ribosylation Factors↗

PubMed1993

Analysis of T-cell receptor alpha beta chains of CD8+ suppressor T cells induced by tolerogenic conjugates of antigen and monomethoxypolyethylene glycol. Involvement of TCR alpha-CDR3 domain in immunosuppression.

A number of investigators have demonstrated that there exists a relationship between Ag receptors of Ts cells (TCR) and their soluble suppressor factors. With a view to elucidating this relationship, the primary structures of receptors of Ts cells, induced in mice by tolerogenic conjugates of Ag and monomethoxypolyethylene glycol, were characterized in this study. The cDNA encoding the alpha and beta chains of TCR of cloned Ts cells specific for (i) OVA and (ii) human monoclonal (myeloma) IgG (HIgG) were produced by polymerase chain reaction. From the analysis of the V alpha genes of TCR of Ts cells it was deduced that these receptors utilized a new member of the V alpha 15 gene family, which was productively joined to the J alpha genes that differed for each of the Ts cells of the two distinct specificities. Similarly, sequence analysis of the beta chain cDNA of the two Ts cell clones revealed that both clones utilized the V beta 8.2 gene, and that their J beta gene differed from each other. It is inferred that the Ts cells generated in response to the different tolerogenic Ag(mPEG)n conjugates belonged to a subset of T cells utilizing similar TCR alpha beta chains and differed only in their J alpha/J beta regions. Most importantly, pretreatment of mice with a mixture of pentadecapeptides comprising the TCR alpha chain of the OVA-Ts cells, down-regulated the immune response specific to OVA, but not to HIgG. Moreover, injection of mice with a pentadecapeptide corresponding to the CDR3 region of the TCR alpha chain of either OVA-Ts or HIgG-Ts suppressed specifically the Ab response to OVA or HIgG, respectively. On the basis of all these results, it is concluded that the CDR3 of the TCR-alpha chain of Ts cells plays a pivotal role in the Ts network underlying the specific down-regulation of the immune responses induced by tolerogenic Ag(mPEG)n conjugates.

Amino Acid Sequence↗

PubMed2004

Industrial application of fuzzy control in bioprocesses.

In a bioprocess, for example a fermentation process, many biological reactions are always working in intracellular space and the control of such a process is very complicated. Bioprocesses have therefore been controlled by the judgment of the experts who are the skilled operators and have much experience in the control of such processes. Such experience is normally described in terms of linguistic IF-THEN rules. Fuzzy inference is a powerful tool for incorporating linguistic rules into computer control of such processes. Fuzzy control is divided into two types--direct fuzzy control of process variables, for example sugar feed rate and fermentation temperature, and indirect control via phase recognition. In bioprocess control the experts decide the value of controllable process variables such as sugar feed rate or temperature as output data from several state variables as input data. Fuzzy control is regarded as a computational algorithm in which the causal relationship between input and output data are incorporated. In Japan fuzzy control has already been applied to practical industrial processes such as production of pravastatin precursor and vitamin B2 and to the Japanese sake mashing process; these examples are reviewed. In addition, an advanced control tool developed from a study on fuzzy control, fuzzy neural networks (FNN), are introduced. FNN can involve complicated causality between input and output data in a network model. FNN have been proven to be applicable to a research in biomedicine, for example modeling of the complicated causality between electroencephalogram or gene expression profiling data and prognostic prediction. Successful results on this research will be also explained.

Algorithms↗

PubMed1980

Kinetics of oxidation and reduction of high-potential iron-sulfur proteins with nonphysiological reactants.

We have investigated the kinetic and equilibrium oxidation-reduction properties of HIPIP's from four sources: Chromatium vinosum, Thiocapsa pfennigii, Rhodopseudomonas gelatinosa, and a paracoccus sp. These studies focused on the interaction of the various HIPIP's and the iron hexacyanides as a function of ionic strength, pH, and temperature. On the basis of the kinetic and equilibrium data obtained and the known structrual properties of the various HIPIP's a number of conclusions concerning the mechanism of electron transfer by HIPIP can be inferred. (1) The oxidation and reduction of HIPIP by the iron hexacyanides show no kinetic complexity due to heterogeneity or changes in rate-limiting step. (2) The kinetic and equilibrium studies (oxidation-reduction potential) are in good agreement, although a rapid binding equilibrium between HIPIP and the iron hexacyanides is possible. (3) The kinetics of oxidation and reduction of the various HIPIP's studied do not correlate with their oxidation-reduction potentials. (4) The electrostatic interactions between a particular HIPIP and the iron hexacyanides appear to be influenced by the charge on the iron-sulfur cluster with specific amino acid side chains playing a significant but limited role in the interactions leading to electron transfer. (5) The iron-sulfur cluster charge is apparently distributed on the surface of the HIPIP molecule through a network of hydrogen bonds. (6) On the basis of an analysis of the known primary and tertiary structure of the HIPIP's studied, it is concluded that peptide backbone carbonyls and amino acid side chains for the amino acid sequence positions 46-49, 63-65, 78-82, 31-34, and 16-17 define the interaction site of HIPIP and the iron hexacyanides irrespective of the HIPIP source. This is a structural region at which the iron-sulfur cluster is near to the protein surface and in this respect resembles the active site in mitochondrial cytochrome c.

Amino Acid Sequence↗

PubMed1994

Molecular pathways to parallel evolution: I. Gene nexuses and their morphological correlates.

Aspects of the regulatory interactions among genes are probably as old as most genes are themselves. Correspondingly, similar predispositions to changes in such interactions must have existed for long evolutionary periods. Features of the structure and the evolution of the system of gene regulation furnish the background necessary for a molecular understanding of parallel evolution. Patently "unrelated" organs, such as the fat body of a fly and the liver of a mammal, can exhibit fractional homology, a fraction expected to become subject to quantitation. This also seems to hold for different organs in the same organism, such as wings and legs of a fly. In informational macromolecules, on the other hand, homology is indeed all or none. In the quite different case of organs, analogy is expected usually to represent attenuated homology. Many instances of putative convergence are likely to turn out to be predominantly parallel evolution, presumably including the case of the vertebrate and cephalopod eyes. Homology in morphological features reflects a similarity in networks of active genes. Similar nexuses of active genes can be established in cells of different embryological origins. Thus, parallel development can be considered a counterpart to parallel evolution. Specific macromolecular interactions leading to the regulation of the c-fos gene are given as an example of a "controller node" defined as a regulatory unit. Quantitative changes in gene control are distinguished from relational changes, and frequent parallelism in quantitative changes is noted in Drosophila enzymes. Evolutionary reversions in quantitative gene expression are also expected. The evolution of relational patterns is attributed to several distinct mechanisms, notably the shuffling of protein domains. The growth of such patterns may in part be brought about by a particular process of compensation for "controller gene diseases," a process that would spontaneously tend to lead to increased regulatory and organismal complexity. Despite the inferred increase in gene interaction complexity, whose course over evolutionary time is unknown, the number of homology groups for the functional and structural protein units designated as domains has probably remained rather constant, even as, in some of its branches, evolution moved toward "higher" organisms. In connection with this process, the question is raised of parallel evolution within the purview of activating and repressing master switches and in regard to the number of levels into which the hierarchies of genic master switches will eventually be resolved.

Animals↗

PubMed2006

Functional activation studies of word processing in the recovery from aphasia.

Some reviews on theories of recovery in aphasia put an emphasis on neural network models based on empirical data from evoked-potentials in aphasia as an approach to mapping recovery of cognitive function to neural structure. We will focus here on what we call an "anatomical" approach to look at recovery in aphasia. "Anatomical" theories of recovery stated by classical aphasiologists have contributed to the understanding of language representations in the human brain. But many aspects of these theories can only be investigated by using modern techniques of lesion analysis, psychometric assessment and functional imaging. Whereas structure-function relations have been primarily established by looking for the association of deficit symptoms with certain lesions, functional activation methods offer a means to study more directly the functional anatomy of recovered or retained functions in neuropsychological patients. To falsify or build up anatomical theories of recovery we will propose a stepwise approach of inference. The methodological pitfalls of this approach will be discussed by focussing on anatomical hypotheses of semantic word comprehension and its impairment and recovery in aphasia.

Aphasia↗

PubMed2004

Phylogeography, genetic structure and diversity in the endangered bearded vulture (Gypaetus barbatus, L) as revealed by mitochondrial DNA.

Bearded vulture populations in the Western Palearctic have experienced a severe decline during the last two centuries that has led to the near extinction of the species in Europe. In this study we analyse the sequence variation at the mitochondrial control region throughout the species range to infer its recent evolutionary history and to evaluate the current genetic status of the species. This study became possible through the extensive use of museum specimens to study populations now extinct. Phylogenetic analysis revealed the existence of two divergent mitochondrial lineages, lineage A occurring mainly in Western European populations and lineage B in African, Eastern European and Central Asian populations. The relative frequencies of haplotypes belonging to each lineage in the different populations show a steep East-West clinal distribution with maximal mixture of the two lineages in the Alps and Greece populations. A genealogical signature for population growth was found for lineage B, but not for lineage A; futhermore the Clade B haplotypes in western populations and clade A haplo-types in eastern populations are recently derived, as revealed by their peripheral location in median-joining haplotype networks. This phylogeographical pattern suggests allopatric differentiation of the two lineages in separate Mediterranean and African or Asian glacial refugia, followed by range expansion from the latter leading to two secondary contact suture zones in Central Europe and North Africa. High levels of among-population differentiation were observed, although these were not correlated with geographical distance. Due to the marked genetic structure, extinction of Central European populations in the last century re-sulted in the loss of a major portion of the genetic diversity of the species. We also found direct evidence for the effect of drift altering the genetic composition of the remnant Pyrenean population after the demographic bottleneck of the last century. Our results argue for the management of the species as a single population, given the apparent ecological exchangeability of extant stocks, and support the ongoing reintroduction of mixed ancestry birds in the Alps and planned reintroductions in Southern Spain.

Africa↗

PubMed2006

A network-based analysis of allergen-challenged CD4+ T cells from patients with allergic rhinitis.

We performed a network-based analysis of DNA microarray data from allergen-challenged CD4(+) T cells from patients with seasonal allergic rhinitis. Differentially expressed genes were organized into a functionally annotated network using the Ingenuity Knowledge Database, which is based on manual review of more than 200,000 publications. The main function of this network is the regulation of lymphocyte apoptosis, a role associated with several genes of the tuber necrosis factor superfamily. The expression of TNFRSF4, one of the genes in this family, was found to be 48 times higher in allergen-challenged cells than in diluent-challenged cells. TNFRSF4 is known to inhibit apoptosis and to enhance Th2 proliferation. Examination of a different material of allergen-stimulated peripheral blood mononuclear cells showed a higher number of interleukin-4(+) type 2 CD4(+) T (Th2) cells in patients than in controls (P<0.01), as well as a higher number of non-apoptotic Th2 cells in patients (P<0.01). The number of Th2 cells expressing TNFRSF4, TNFSF7 and TNFRSF1B was also significantly higher in patients. Treatment with anti-TNFSF4 resulted in a significantly decreased number of Th2 cells (P<0.05). A logical inference from all this is that the proliferation of allergen-challenged Th2 cells is associated with a decreased apoptosis of Th2 cells and an increase in TNFRSF4 signalling.

Adolescent↗

PubMed1999

Social intelligence in the normal and autistic brain: an fMRI study.

There is increasing support for the existence of 'social intelligence' [Humphrey (1984) Consciousness Regained], independent of general intelligence. Brothers et al. 1990) J. Cog. Neurosci., 4, 107-118] proposed a network of neural regions that comprise the 'social brain': the orbito-frontal cortex (OFC), superior temporal gyrus (STG) and amygdala. We tested Brothers' theory by examining both normal subjects as well as patients with high-functioning autism or Asperger syndrome (AS), who are well known to have deficits in social intelligence, and perhaps deficits in amygdala function [Bauman & Kemper (1988) J. Neuropath. Exp. Neurol., 47, 369]. We used a test of judging from the expressions of another person's eyes what that other person might be thinking or feeling. Using functional magnetic resonance imaging (fMRI) we confirmed Brothers' prediction that the STG and amygdala show increased activation when using social intelligence. Some areas of the prefrontal cortex also showed activation. In contrast, patients with autism or AS activated the fronto-temporal regions but not the amygdala when making mentalistic inferences from the eyes. These results provide support for the social brain theory of normal function, and the amygdala theory of autism.

Adult↗

PubMed2006

Hybridization interactions between probesets in short oligo microarrays lead to spurious correlations.

BACKGROUND: Microarrays measure the binding of nucleotide sequences to a set of sequence specific probes. This information is combined with annotation specifying the relationship between probes and targets and used to make inferences about transcript- and, ultimately, gene expression. In some situations, a probe is capable of hybridizing to more than one transcript, in others, multiple probes can target a single sequence. These 'multiply targeted' probes can result in non-independence between measured expression levels. RESULTS: An analysis of these relationships for Affymetrix arrays considered both the extent and influence of exact matches between probe and transcript sequences. For the popular HGU133A array, approximately half of the probesets were found to interact in this way. Both real and simulated expression datasets were used to examine how these effects influenced the expression signal. It was found not only to lead to increased signal strength for the affected probesets, but the major effect is to significantly increase their correlation, even in situations when only a single probe from a probeset was involved. By building a network of probe-probeset-transcript relationships, it is possible to identify families of interacting probesets. More than 10% of the families contain members annotated to different genes or even different Unigene clusters. Within a family, a mixture of genuine biological and artefactual correlations can occur. CONCLUSION: Multiple targeting is not only prevalent, but also significant. The ability of probesets to hybridize to more than one gene product can lead to false positives when analysing gene expression. Comprehensive annotation describing multiple targeting is required when interpreting array data.

Artifacts↗

PubMed2006

Soft mixer assignment in a hierarchical generative model of natural scene statistics.

Gaussian scale mixture models offer a top-down description of signal generation that captures key bottom-up statistical characteristics of filter responses to images. However, the pattern of dependence among the filters for this class of models is prespecified. We propose a novel extension to the gaussian scale mixture model that learns the pattern of dependence from observed inputs and thereby induces a hierarchical representation of these inputs. Specifically, we propose that inputs are generated by gaussian variables (modeling local filter structure), multiplied by a mixer variable that is assigned probabilistically to each input from a set of possible mixers. We demonstrate inference of both components of the generative model, for synthesized data and for different classes of natural images, such as a generic ensemble and faces. For natural images, the mixer variable assignments show invariances resembling those of complex cells in visual cortex; the statistics of the gaussian components of the model are in accord with the outputs of divisive normalization models. We also show how our model helps interrelate a wide range of models of image statistics and cortical processing.

Animals↗