How does the brain solve visual object recognition?
By James J. DiCarlo, Davide Zoccolan, Nicole C. Rust. 2012
1,895 citations recorded by the source corpus.
Mounting evidence suggests that “core object recognition,” the ability to rapidly recognize objects despite substantial appearance variation, is solved in the brain via a cascade of reflexive, largely feedforward computations that culminate in a powerful neuronal representation in the inferior temporal cortex. However, the algorithm that produces this solution remains little-understood. Here we review evidence ranging from individual neurons, to neuronal populations, to behavior, to computational models. We propose that understanding this algorithm will require using neuronal and psychophysical data to sift through many computational models, each based on building blocks of small, canonical sub-networks with a common functional goal.
Subfields: Cognitive Neuroscience, Neuroscience
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Cognitive Neuroscience
- James J. DiCarlo
- Nikolaus Kriegeskorte
- Matthias Bethge
- Tomaso A. Poggio
- Daniel L. K. Yamins
- Xiao-Jing Wang
- Yoshua Bengio
- Grace W. Lindsay
In Neuroscience
- Nikolaus Kriegeskorte
- Matthias Bethge
- James J. DiCarlo
- Daniel L. K. Yamins
- Yoshua Bengio
- Alexander S. Ecker
- Leon A. Gatys
- Thomas Serre
Citations in the corpus
- Learning Visual Features from Large Weakly Supervised Data (2015, background): “…rather inefficient compared to how humans learn to recognize objects: unsupervised and weakly supervised learning plays an important role in human vision [9], as a result of which humans do not need to see thousands of images of, say, aardvarks to obtain a good grasp of what an aardvark looks like.”
- Representation Learning: A Review and New Perspectives (2012, background): “This confers the resulting pooled feature layer some degree of invariance to input translations, and this style of architecture (alternating selective feature extraction and invariance-creating pooling) has been the basis of convolutional networks, the Neocognitron (Fukushima and Miyake, 1982) and H”
- Building high-level features using large scale unsupervised learning (2011, background): “(1) Here, λ is a tradeoff parameter between sparsity and reconstruction; m, k are the number of examples and pooling units in a layer respectively; Hj is the vector of weights of the j-th pooling unit.”
- Deep learning in neural networks: An overview (2014, neutral): “…decades (e.g., Bichot, Rossi, & Desimone, 2005; Connor, Brincat, & Pasupathy, 2007; Desimone, Albright, Gross, & Bruce, 1984; DiCarlo, Zoccolan, & Rust, 2012; Felleman & Van Essen, 1991; Hubel & Wiesel, 1968; Hung, Kreiman, Poggio, & DiCarlo, 2005; Kobatake & Tanaka, 1994; Kriegeskorte et al.,…”
- View-tolerant face recognition and Hebbian learning imply mirror-symmetric neural tuning to head orientation (2016, positive): “Abstract: The primate brain contains a hierarchy of visual areas, dubbed the ventral stream, which rapidly computes object representations that are both specific for object identity and relatively robust against identity-preserving transformations like depth-rotations [33, 32, 23, 13].”
- Theory of Mind: A Neural Prediction Problem (2013, positive): “Also like vision, theory of mind is a complex cognitive process that depends on many different brain regions with likely distinct computational roles (DiCarlo et al., 2012).”
- This Looks Like That, Because ... Explaining Prototypes for Interpretable Image Recognition (2020, background): “Moreover, where humans are tolerant of moderate changes in object position, size, pose, contrast, illumination and clutter [14], deep image classification models might not be robust for such transformations [40, 44].”
- Re-evaluating circuit mechanisms underlying pattern separation (2019, positive): “This expansion and sparsening of visual information is consistent with the theory that the successive stages of the visual pathway increase the separability of cortical representations of object identity (DiCarlo et al., 2012).”
- Generalisation in humans and deep neural networks (2018, background): “We know, for example, that it is possible to reliably identify objects in the central visual field within a single fixation in less than 200 ms when viewing “standard” images [2–4].”
- Texture synthesis and the controlled generation of natural stimuli using convolutional neural networks (2015, background): “It is thought that a main computation in the ventral visual pathway is to make high-level visual information increasingly explicit [3].”
- Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence (2016, background): “This showed that the object DNN had representations in a hierarchy of increasing complexity, akin to representations in the primate visual brain hierarchy 5,7 .”
- Dynamic Scene Classification: Learning Motion Descriptors with Slow Features Analysis (2013, background): “Figure 1 illustrates how SFA learning can significantly improve the untangling problem objective.”
- Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures (2013, background): “There is a long tradition of basing computer vision systems on models of biological vision (Fukushima, 1980; LeCun et al., 1989; Riesenhuber & Poggio, 1999; Lowe, 1999; Hinton et al., 2006; DiCarlo et al., 2012).”