A unifying framework for functional organization in early and higher ventral visual cortex.

A key feature of cortical systems is functional organization: the arrangement of functionally distinct neurons in characteristic spatial patterns. However, the principles underlying the emergence of functional organization in the cortex are poorly understood. Here, we develop the topographic deep artificial neural network (TDANN), the first model to predict several aspects of the functional organization of multiple cortical areas in the primate visual system. We analyze the factors driving the TDANN's success and find that it balances two objectives: learning a task-general sensory representation and maximizing the spatial smoothness of responses according to a metric that scales with cortical surface area. In turn, the representations learned by the TDANN are more brain-like than in spatially unconstrained models. Finally, we provide evidence that the TDANN's functional organization balances performance with between-area connection length. Our results offer a unified principle for understanding the functional organization of the primate ventral visual system.

The Limiting Dynamics of SGD: Modified Loss, Phase-Space Oscillations, and Anomalous Diffusion.

In this work, we explore the limiting dynamics of deep neural networks trained with stochastic gradient descent (SGD). As observed previously, long after performance has converged, networks continue to move through parameter space by a process of anomalous diffusion in which distance traveled grows as a power law in the number of gradient updates with a nontrivial exponent. We reveal an intricate interaction among the hyperparameters of optimization, the structure in the gradient noise, and the Hessian matrix at the end of training that explains this anomalous diffusion. To build this understanding, we first derive a continuous-time model for SGD with finite learning rates and batch sizes as an underdamped Langevin equation. We study this equation in the setting of linear regression, where we can derive exact, analytic expressions for the phase-space dynamics of the parameters and their instantaneous velocities from initialization to stationarity. Using the Fokker-Planck equation, we show that the key ingredient driving these dynamics is not the original training loss but rather the combination of a modified loss, which implicitly regularizes the velocity, and probability currents that cause oscillations in phase space. We identify qualitative and quantitative predictions of this theory in the dynamics of a ResNet-18 model trained on ImageNet. Through the lens of statistical physics, we uncover a mechanistic origin for the anomalous limiting dynamics of deep neural networks trained with SGD. Understanding the limiting dynamics of SGD, and its dependence on various important hyperparameters like batch size, learning rate, and momentum, can serve as a basis for future work that can turn these insights into algorithmic gains.

A Unifying Principle for the Functional Organization of Visual Cortex.

A key feature of many cortical systems is functional organization: the arrangement of neurons with specific functional properties in characteristic spatial patterns across the cortical surface. However, the principles underlying the emergence and utility of functional organization are poorly understood. Here we develop the Topographic Deep Artificial Neural Network (TDANN), the first unified model to accurately predict the functional organization of multiple cortical areas in the primate visual system. We analyze the key factors responsible for the TDANN's success and find that it strikes a balance between two specific objectives: achieving a task-general sensory representation that is self-supervised, and maximizing the smoothness of responses across the cortical sheet according to a metric that scales relative to cortical surface area. In turn, the representations learned by the TDANN are lower dimensional and more brain-like than those in models that lack a spatial smoothness constraint. Finally, we provide evidence that the TDANN's functional organization balances performance with inter-area connection length, and use the resulting models for a proof-of-principle optimization of cortical prosthetic design. Our results thus offer a unified principle for understanding functional organization and a novel view of the functional role of the visual system in particular.

Increasing neural network robustness improves match to macaque V1 eigenspectrum, spatial frequency preference and predictivity.

Task-optimized convolutional neural networks (CNNs) show striking similarities to the ventral visual stream. However, human-imperceptible image perturbations can cause a CNN to make incorrect predictions. Here we provide insight into this brittleness by investigating the representations of models that are either robust or not robust to image perturbations. Theory suggests that the robustness of a system to these perturbations could be related to the power law exponent of the eigenspectrum of its set of neural responses, where power law exponents closer to and larger than one would indicate a system that is less susceptible to input perturbations. We show that neural responses in mouse and macaque primary visual cortex (V1) obey the predictions of this theory, where their eigenspectra have power law exponents of at least one. We also find that the eigenspectra of model representations decay slowly relative to those observed in neurophysiology and that robust models have eigenspectra that decay slightly faster and have higher power law exponents than those of non-robust models. The slow decay of the eigenspectra suggests that substantial variance in the model responses is related to the encoding of fine stimulus features. We therefore investigated the spatial frequency tuning of artificial neurons and found that a large proportion of them preferred high spatial frequencies and that robust models had preferred spatial frequency distributions more aligned with the measured spatial frequency distribution of macaque V1 cells. Furthermore, robust models were quantitatively better models of V1 than non-robust models. Our results are consistent with other findings that there is a misalignment between human and machine perception. They also suggest that it may be useful to penalize slow-decaying eigenspectra or to bias models to extract features of lower spatial frequencies during task-optimization in order to improve robustness and V1 neural response predictivity.

Ultra-high-resolution fMRI of Human Ventral Temporal Cortex Reveals Differential Representation of Categories and Domains.

Human ventral temporal cortex (VTC) is critical for visual recognition. It is thought that this ability is supported by large-scale patterns of activity across VTC that contain information about visual categories. However, it is unknown how category representations in VTC are organized at the submillimeter scale and across cortical depths. To fill this gap in knowledge, we measured BOLD responses in medial and lateral VTC to images spanning 10 categories from five domains (written characters, bodies, faces, places, and objects) at an ultra-high spatial resolution of 0.8 mm using 7 Tesla fMRI in both male and female participants. Representations in lateral VTC were organized most strongly at the general level of domains (e.g., places), whereas medial VTC was also organized at the level of specific categories (e.g., corridors and houses within the domain of places). In both lateral and medial VTC, domain-level and category-level structure decreased with cortical depth, and downsampling our data to standard resolution (2.4 mm) did not reverse differences in representations between lateral and medial VTC. The functional diversity of representations across VTC partitions may allow downstream regions to read out information in a flexible manner according to task demands. These results bridge an important gap between electrophysiological recordings in single neurons at the micron scale in nonhuman primates and standard-resolution fMRI in humans by elucidating distributed responses at the submillimeter scale with ultra-high-resolution fMRI in humans.SIGNIFICANCE STATEMENT Visual recognition is a fundamental ability supported by human ventral temporal cortex (VTC). However, the nature of fine-scale, submillimeter distributed representations in VTC is unknown. Using ultra-high-resolution fMRI of human VTC, we found differential distributed visual representations across lateral and medial VTC. Domain representations (e.g., faces, bodies, places, characters) were most salient in lateral VTC, whereas category representations (e.g., corridors/houses within the domain of places) were equally salient in medial VTC. These results bridge an important gap between electrophysiological recordings in single neurons at a micron scale and fMRI measurements at a millimeter scale.

Visual noise consisting of X-junctions has only a minimal adverse effect on object recognition.

In 1968, Guzman showed that the myriad of surfaces composing a highly complex and novel assemblage of volumes can readily be assigned to their appropriate volumes in terms of the constraints offered by the vertices of coterminating edges. Of particular importance was the L-vertex, produced by the cotermination of two contours, which provides strong evidence for the termination of a 2-D surface. An X-junction, formed by the crossing of two contours without a change of direction at the crossing, played no role in the segmentation of a scene. If the potency of noise elements to affect recognition performance reflects their relevancy to the segmentation of scenes, as was suggested by Guzman, gaps in an object's contours bounded by irrelevant X-junctions would be expected to have little or no adverse effect on shape-based object recognition, whereas gaps bounded by L-junctions would be expected to have a strong deleterious effect when they disrupt the smooth continuation of contours. Guzman's roles for the various vertices and junctions have never been put to systematic test with respect to human object recognition. By adding identical noise contours to line drawings of objects that produced either L-vertices or X-junctions, these shape features could be compared with respect to their disruption of object recognition. Guzman's insights that irrelevant L-vertices should be highly disruptive and irrelevant X-vertices would have only a minimal deleterious effect were confirmed.

A critical assessment of data quality and venous effects in sub-millimeter fMRI.

Advances in hardware, pulse sequences, and reconstruction techniques have made it possible to perform functional magnetic resonance imaging (fMRI) at sub-millimeter resolution while maintaining high spatial coverage and acceptable signal-to-noise ratio. Here, we examine whether sub-millimeter fMRI can be used as a routine method for obtaining accurate measurements of fine-scale local neural activity. We conducted fMRI in human visual cortex during a simple event-related visual experiment (7 T, gradient-echo EPI, 0.8-mm isotropic voxels, 2.2-s sampling rate, 84 slices), and developed analysis and visualization tools to assess the quality of the data. Our results fall along three lines of inquiry. First, we find that the acquired fMRI images, combined with appropriate surface-based processing, provide reliable and accurate measurements of fine-scale blood oxygenation level dependent (BOLD) activity patterns. Second, we show that the highly folded structure of cortex causes substantial biases on spatial resolution and data visualization. Third, we examine the well-recognized issue of venous contributions to fMRI signals. In a systematic assessment of large sections of cortex measured at a fine scale, we show that time-averaged T2*-weighted EPI intensity is a simple, robust marker of venous effects. These venous effects are unevenly distributed across cortex, are more pronounced in gyri and outer cortical depths, and are, to a certain degree, in consistent locations across subjects relative to cortical folding. Furthermore, we show that these venous effects are strongly correlated with BOLD responses evoked by the experiment. We conclude that sub-millimeter fMRI can provide robust information about fine-scale BOLD activity patterns, but special care must be exercised in visualizing and interpreting these patterns, especially with regards to the confounding influence of the brain's vasculature. To help translate these methodological findings to neuroscience research, we provide practical suggestions for both high-resolution and standard-resolution fMRI studies.

The cognitive neuroscience of person identification.

We compare and contrast five differences between person identification by voice and face. 1. There is little or no cost when a familiar face is to be recognized from an unrestricted set of possible faces, even at Rapid Serial Visual Presentation (RSVP) rates, but the accuracy of familiar voice recognition declines precipitously when the set of possible speakers is increased from one to a mere handful. 2. Whereas deficits in face recognition are typically perceptual in origin, those with normal perception of voices can manifest severe deficits in their identification. 3. Congenital prosopagnosics (CPros) and congenital phonagnosics (CPhon) are generally unable to imagine familiar faces and voices, respectively. Only in CPros, however, is this deficit a manifestation of a general inability to form visual images of any kind. CPhons report no deficit in imaging non-voice sounds. 4. The prevalence of CPhons of 3.2% is somewhat higher than the reported prevalence of approximately 2.0% for CPros in the population. There is evidence that CPhon represents a distinct condition statistically and not just normal variation. 5. Face and voice recognition proficiency are uncorrelated rather than reflecting limitations of a general capacity for person individuation.

What Is Actually Affected by the Scrambling of Objects When Localizing the Lateral Occipital Complex?

The lateral occipital complex (LOC), the cortical region critical for shape perception, is localized with fMRI by its greater BOLD activity when viewing intact objects compared with their scrambled versions (resembling texture). Despite hundreds of studies investigating LOC, what the LOC localizer accomplishes-beyond distinguishing shape from texture-has never been resolved. By independently scattering the intact parts of objects, the axis structure defining the relations between parts was no longer defined. This led to a diminished BOLD response, despite the increase in the number of independent entities (the parts) produced by the scattering, thus indicating that LOC specifies interpart relations, in addition to specifying the shape of the parts themselves. LOC's sensitivity to relations is not confined to those between parts but is also readily apparent between objects, rendering it-and not subsequent "place" areas-as the critical region for the representation of scenes. Moreover, that these effects are witnessed with novel as well as familiar intact objects and scenes suggests that the relations are computed on the fly, rather than being retrieved from memory.

The Lateral Occipital Complex shows no net response to object familiarity.

In 1995, Malach et al. discovered an area whose fMRI BOLD response was greater when viewing intact, familiar objects than when viewing their scrambled versions (resembling texture). Since then hundreds of studies have explored this late visual region termed the Lateral Occipital Complex (LOC), which is now known to be critical for shape perception (James, Culham, Humphrey, Milner, & Goodale, 2003). Malach et al. (1995) discounted a role of familiarity by showing that "abstract" Henry Moore sculptures, unfamiliar to the subjects, also activated this region. This characterization of LOC as a region that responds to shape independently of familiarity has been accepted but never tested with control of the same low-level features. We assessed LOC's response to objects that had identical parts in two different arrangements, one familiar and the other novel. Malach was correct: There is no net effect of familiarity in LOC. However, a multivoxel correlation analysis showed that LOC does distinguish familiar from novel objects.

An applet for the Gabor similarity scaling of the differences between complex stimuli.

It is widely accepted that after the first cortical visual area, V1, a series of stages achieves a representation of complex shapes, such as faces and objects, so that they can be understood and recognized. A major challenge for the study of complex shape perception has been the lack of a principled basis for scaling of the physical differences between stimuli so that their similarity can be specified, unconfounded by early-stage differences. Without the specification of such similarities, it is difficult to make sound inferences about the contributions of later stages to neural activity or psychophysical performance. A Web-based app is described that is based on the Malsburg Gabor-jet model (Lades et al., 1993), which allows easy specification of the V1 similarity of pairs of stimuli, no matter how intricate. The model predicts the psycho physical discriminability of metrically varying faces and complex blobs almost perfectly (Yue, Biederman, Mangini, von der Malsburg, & Amir, 2012), and serves as the input stage of a large family of contemporary neurocomputational models of vision.

Neuromelanin marks the spot: identifying a locus coeruleus biomarker of cognitive reserve in healthy aging.

Leading a mentally stimulating life may build up a reserve of neural and mental resources that preserve cognitive abilities in late life. Recent autopsy evidence links neuronal density in the locus coeruleus (LC), the brain's main source of norepinephrine, to slower cognitive decline before death, inspiring the idea that the noradrenergic system is a key component of reserve (Robertson, I. H. 2013. A noradrenergic theory of cognitive reserve: implications for Alzheimer's disease. Neurobiol. Aging. 34, 298-308). Here, we tested this hypothesis using neuromelanin-sensitive magnetic resonance imaging to visualize and measure LC signal intensity in healthy younger and older adults. Established proxies of reserve, including education, occupational attainment, and verbal intelligence, were linearly correlated with LC signal intensity in both age groups. Results indicated that LC signal intensity was significantly higher in older than younger adults and significantly lower in women than in men. Consistent with the LC-reserve hypothesis, both verbal intelligence and a composite reserve score were positively associated with LC signal intensity in older adults. LC signal intensity was also more strongly associated with attentional shifting ability in older adults with lower cognitive reserve. Together these findings link in vivo estimates of LC neuromelanin signal intensity to cognitive reserve in normal aging.