Deep Supervised, but Not Unsupervised, Models May Explain IT Cortical Representation

PLoS Computational Biology, Journal Year: 2014, Volume and Issue: 10(11), P. e1003915 - e1003915

Published: Nov. 6, 2014

Inferior temporal (IT) cortex in human and nonhuman primates serves visual object recognition. Computational object-vision models, although continually improving, do not yet reach performance. It is unclear to what extent the internal representations of computational models can explain IT representation. Here we investigate a wide range model (37 total), testing their categorization performance ability account for representational geometry. The include well-known neuroscientific object-recognition (e.g. HMAX, VisNet) along with several from computer vision SIFT, GIST, self-similarity features, deep convolutional neural network). We compared dissimilarity matrices (RDMs) RDMs obtained (measured fMRI) monkey cell recording) same set stimuli (not used training models). Better performing were more similar that they showed greater clustering patterns by category. In addition, better also strongly resembled terms within-category dissimilarities. Representational geometries significantly correlated between many models. However, categorical observed was largely unexplained unsupervised network, which trained supervision over million category-labeled images, reached highest best explained IT, it did fully data. Combining features this appropriate weights adding linear combinations maximize margin animate inanimate objects faces other yielded representation our Overall, results suggest explaining requires through supervised learning emphasize behaviorally important divisions prominently reflected IT.

Language: Английский

---

Deep Neural Networks: A New Framework for Modeling Biological Vision and Brain Information Processing

Annual Review of Vision Science, Journal Year: 2015, Volume and Issue: 1(1), P. 417 - 446

Published: Nov. 18, 2015

Recent advances in neural network modeling have enabled major strides computer vision and other artificial intelligence applications. Human-level visual recognition abilities are coming within reach of systems. Artificial networks inspired by the brain, their computations could be implemented biological neurons. Convolutional feedforward networks, which now dominate vision, take further inspiration from architecture primate hierarchy. However, current models designed with engineering goals, not to model brain computations. Nevertheless, initial studies comparing internal representations between these brains find surprisingly similar representational spaces. With human-level performance no longer out reach, we entering an exciting new era, will able build biologically faithful recurrent computational how perform high-level feats intelligence, including vision.

Language: Английский

---

Consolidation alters motor sequence-specific distributed representations

eLife, Journal Year: 2019, Volume and Issue: 8

Published: March 18, 2019

Functional magnetic resonance imaging (fMRI) studies investigating the acquisition of sequential motor skills in humans have revealed learning-related functional reorganizations cortico-striatal and cortico-cerebellar systems accompanied with an initial hippocampal contribution. Yet, significance these activity-level changes remains ambiguous as they convey evolution both sequence-specific knowledge unspecific task ability. Moreover, do not specifically assess occurrence plasticity. To address issues, we investigated local circuits tuning to information using multivariate distances between patterns evoked by consolidated or newly acquired sequences production. The results reveal that representations dorsolateral striatum, prefrontal secondary cortices are greater when executing than untrained ones. By contrast, sequence hippocampus dorsomedial striatum becomes less engaged. Our findings show, for first time humans, complementary evolve distinctively during critical phases skill consolidation.

Language: Английский

---

Building a Science of Individual Differences from fMRI

Trends in Cognitive Sciences, Journal Year: 2016, Volume and Issue: 20(6), P. 425 - 443

Published: May 1, 2016

Language: Английский

---

CoSMoMVPA: Multi-Modal Multivariate Pattern Analysis of Neuroimaging Data in Matlab/GNU Octave

Frontiers in Neuroinformatics, Journal Year: 2016, Volume and Issue: 10

Published: July 22, 2016

Recent years have seen an increase in the popularity of multivariate pattern (MVP) analysis functional magnetic resonance (fMRI) data, and, to a much lesser extent, magneto- and electro-encephalography (M/EEG) data. We present CoSMoMVPA, lightweight MVPA (MVP analysis) toolbox implemented intersection Matlab GNU Octave languages, that treats both fMRI M/EEG data as first-class citizens. CoSMoMVPA supports all state-of-the-art MVP techniques, including searchlight analyses, classification, correlations, representational similarity analysis, time generalization method. These can be used address data-driven hypothesis-driven questions about neural organization representations, within across: space, time, frequency bands, neuroimaging modalities, individuals, species. It uses uniform representation volume or on surface, at sensor source level. Through various external toolboxes, it directly reading writing variety formats, where applicable, convert between them. As result, integrated readily existing pipelines with preprocessed datasets. overloads traditional volumetric concept support neighborhoods for surface-based which localization effects interest across dimensions. also provides generalized approach multiple comparison correction these dimensions using Threshold-Free Cluster Enhancement clustering permutation techniques. is highly modular abstractions provide interface measures. Typical analyses require few lines code, making accessible beginner users. At same expert programmers easily extend its functionality. comes extensive documentation, runnable demonstration scripts exercises (with example solutions). best software engineering practices version control, distributed development, automated test suite, continuous integration testing. proprietary free software, complies open distribution platforms such NeuroDebian. Free/Open Source Software under permissive MIT license. Website: http://cosmomvpa.org code: https://github.com/CoSMoMVPA/CoSMoMVPA.

Language: Английский

---

Reliability of dissimilarity measures for multi-voxel pattern analysis

NeuroImage, Journal Year: 2015, Volume and Issue: 137, P. 188 - 200

Published: Dec. 18, 2015

Language: Английский

---

Cognitive computational neuroscience

Nature Neuroscience, Journal Year: 2018, Volume and Issue: 21(9), P. 1148 - 1160

Published: Aug. 10, 2018

Language: Английский

---

Generic decoding of seen and imagined objects using hierarchical visual features

Nature Communications, Journal Year: 2017, Volume and Issue: 8(1)

Published: May 22, 2017

Abstract Object recognition is a key function in both human and machine vision. While brain decoding of seen imagined objects has been achieved, the prediction limited to training examples. We present approach for arbitrary using vision principle that an object category represented by set features rendered invariant through hierarchical processing. show visual features, including those derived from deep convolutional neural network, can be predicted fMRI patterns, greater accuracy achieved low-/high-level with lower-/higher-level areas, respectively. Predicted are used identify seen/imagined categories (extending beyond decoder training) computed numerous images. Furthermore, reveals progressive recruitment higher-to-lower representations. Our results demonstrate homology between its utility brain-based information retrieval.

Language: Английский

---

A Primer on Pattern-Based Approaches to fMRI: Principles, Pitfalls, and Perspectives

Neuron, Journal Year: 2015, Volume and Issue: 87(2), P. 257 - 270

Published: July 1, 2015

Language: Английский

---

Neural network models and deep learning

Current Biology, Journal Year: 2019, Volume and Issue: 29(7), P. R231 - R236

Published: April 1, 2019

Language: Английский

---