The brain combines information from multiple sensory modalities to build a consistent representation of the world. The principles by which multimodal stimuli are integrated in cortical hierarchies are well studied, but it is less clear whether and how unimodal inputs shape the processing of signals carried by a different modality. In rodents, for instance, direct connections from primary auditory cortex reach visual cortex, but studies disagree on the impact of these projections on visual cortical processing. Both enhancement and suppression of visually evoked responses by auditory inputs have been reported, as well as sharpening of orientation tuning and improvement in the coding of visual information. Little is known, however, about the functional impact of auditory signals on rodent visual perception. Here we trained a group of rats in a visual temporal frequency (TF) classification task, where the visual stimuli to categorize were paired with simultaneous but task-irrelevant auditory stimuli, to prevent high-level multisensory integration and investigate instead the spontaneous, direct impact of auditory signals on the perception of visual stimuli. Rat classification of visual TF was strongly and systematically altered by the presence of sounds, in a way that was determined by sound intensity but not by its temporal modulation. To investigate the mechanisms underlying this phenomenon, we developed a Bayesian ideal observer model, combined with a neural coding scheme where neurons linearly encode visual TF but are inhibited by concomitant sounds by a measure that depends on their intensity. This model captured very precisely the full spectrum of rat perceptual choices we observed, supporting the hypothesis that auditory inputs induce an effective compression of the visual perceptual space. This suggests an important role for inhibition as the key mediator of auditory-visual interactions and provides clear, mechanistic hypotheses to be tested by future work on visual cortical codes.

Visual object recognition has been extensively studied in both neuroscience and computer vision. Recently, the most popular class of artificial systems for this task, deep convolutional neural networks (CNNs), has been shown to provide excellent models for its functional analogue in the brain, the ventral stream in visual cortex. This has prompted questions on what, if any, are the common principles underlying the reformatting of visual information as it flows through a CNN or the ventral stream. Here we consider some prominent statistical patterns that are known to exist in the internal representations of either CNNs or the visual cortex and look for them in the other system. We show that intrinsic dimensionality (ID) of object representations along the rat homologue of the ventral stream presents two distinct expansion-contraction phases, as previously shown for CNNs. Conversely, in CNNs, we show that training results in both distillation and active pruning (mirroring the increase in ID) of low- to middle-level image information in single units, as representations gain the ability to support invariant discrimination, in agreement with previous observations in rat visual cortex. Taken together, our findings suggest that CNNs and visual cortex share a similarly tight relationship between dimensionality expansion/reduction of object representations and reformatting of image information.

Computing global motion direction of extended visual objects is a hallmark of primate high-level vision. Although neurons selective for global motion have also been found in mouse visual cortex, it remains unknown whether rodents can combine multiple motion signals into global, integrated percepts. To address this question, we trained two groups of rats to discriminate either gratings (G group) or plaids (i.e., superpositions of gratings with different orientations; P group) drifting horizontally along opposite directions. After the animals learned the task, we applied a visual priming paradigm, where presentation of the target stimulus was preceded by the brief presentation of either a grating or a plaid. The extent to which rat responses to the targets were biased by such prime stimuli provided a measure of the spontaneous, perceived similarity between primes and targets. We found that gratings and plaids, when used as primes, were equally effective at biasing the perception of plaid direction for the rats of the P group. Conversely, for the G group, only the gratings acted as effective prime stimuli, while the plaids failed to alter the perception of grating direction. To interpret these observations, we simulated a decision neuron reading out the representations of gratings and plaids, as conveyed by populations of either component or pattern cells (i.e., local or global motion detectors). We concluded that the findings for the P group are highly consistent with the existence of a population of pattern cells, playing a functional role similar to that demonstrated in primates. We also explored different scenarios that could explain the failure of the plaid stimuli to elicit a sizable priming magnitude for the G group. These simulations yielded testable predictions about the properties of motion representations in rodent visual cortex at the single-cell and circuitry level, thus paving the way to future neurophysiology experiments.

Deep neural networks progressively transform their inputs across multiple processing layers. What are the geometrical properties of the representations learned by these networks? Here we study the intrinsic dimensionality (ID) of data representations, i.e. the minimal number of parameters needed to describe a representation. We find that, in a trained network, the ID is orders of magnitude smaller than the number of units in each layer. Across layers, the ID first increases and then progressively decreases in the final layers. Remarkably, the ID of the last hidden layer predicts classification accuracy on the test set. These results can neither be found by linear dimensionality estimates (e.g., with principal component analysis), nor in representations that had been artificially linearized. They are neither found in untrained networks, nor in networks that are trained on randomized labels. This suggests that neural networks that can generalize are those that transform the data into low-dimensional, but not necessarily flat manifolds.

The anterior inferotemporal cortex (IT) is the highest stage along the hierarchy of visual areas that, in primates, processes visual objects. Although several lines of evidence suggest that IT primarily represents visual shape information, some recent studies have argued that neuronal ensembles in IT code the semantic membership of visual objects (i.e., represent conceptual classes such as animate and inanimate objects). In this study, we investigated to what extent semantic, rather than purely visual information, is represented in IT by performing a multivariate analysis of IT responses to a set of visual objects. By relying on a variety of machine-learning approaches (including a cutting-edge clustering algorithm that has been recently developed in the domain of statistical physics), we found that, in most instances, IT representation of visual objects is accounted for by their similarity at the level of shape or, more surprisingly, low-level visual properties. Only in a few cases we observed IT representations of semantic classes that were not explainable by the visual similarity of their members. Overall, these findings reassert the primary function of IT as a conveyor of explicit visual shape information, and reveal that low-level visual properties are represented in IT to a greater extent than previously appreciated. In addition, our work demonstrates how combining a variety of state-of-the-art multivariate approaches, and carefully estimating the contribution of shape similarity to the representation of object categories, can substantially advance our understanding of neuronal coding of visual objects in cortex.