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.

Spatial and Temporal Context Jointly Modulate the Sensory Response within the Ventral Visual Stream

Both spatial and temporal context play an important role in visual perception and behavior. Humans can extract statistical regularities from both forms of context to help process the present and to construct expectations about the future. Numerous studies have found reduced neural responses to expected stimuli compared with unexpected stimuli, for both spatial and temporal regularities. However, it is largely unclear whether and how these forms of context interact. In the current fMRI study, 33 human volunteers were exposed to pairs of object stimuli that could be expected or surprising in terms of their spatial and temporal context. We found reliable independent contributions of both spatial and temporal context in modulating the neural response. Specifically, neural responses to stimuli in expected compared with unexpected contexts were suppressed throughout the ventral visual stream. These results suggest that both spatial and temporal context may aid sensory processing in a similar fashion, providing evidence on how different types of context jointly modulate perceptual processing.

What Does a Language-And-Vision Transformer See: The Impact of Semantic Information on Visual Representations

Neural networks have proven to be very successful in automatically capturing the composition of language and different structures across a range of multi-modal tasks. Thus, an important question to investigate is how neural networks learn and organise such structures. Numerous studies have examined the knowledge captured by language models (LSTMs, transformers) and vision architectures (CNNs, vision transformers) for respective uni-modal tasks. However, very few have explored what structures are acquired by multi-modal transformers where linguistic and visual features are combined. It is critical to understand the representations learned by each modality, their respective interplay, and the task’s effect on these representations in large-scale architectures. In this paper, we take a multi-modal transformer trained for image captioning and examine the structure of the self-attention patterns extracted from the visual stream. Our results indicate that the information about different relations between objects in the visual stream is hierarchical and varies from local to a global object-level understanding of the image. In particular, while visual representations in the first layers encode the knowledge of relations between semantically similar object detections, often constituting neighbouring objects, deeper layers expand their attention across more distant objects and learn global relations between them. We also show that globally attended objects in deeper layers can be linked with entities described in image descriptions, indicating a critical finding - the indirect effect of language on visual representations. In addition, we highlight how object-based input representations affect the structure of learned visual knowledge and guide the model towards more accurate image descriptions. A parallel question that we investigate is whether the insights from cognitive science echo the structure of representations that the current neural architecture learns. The proposed analysis of the inner workings of multi-modal transformers can be used to better understand and improve on such problems as pre-training of large-scale multi-modal architectures, multi-modal information fusion and probing of attention weights. In general, we contribute to the explainable multi-modal natural language processing and currently shallow understanding of how the input representations and the structure of the multi-modal transformer affect visual representations.

Face detection in untrained deep neural networks

Face-selective neurons are observed in the primate visual pathway and are considered as the basis of face detection in the brain. However, it has been debated as to whether this neuronal selectivity can arise innately or whether it requires training from visual experience. Here, using a hierarchical deep neural network model of the ventral visual stream, we suggest a mechanism in which face-selectivity arises in the complete absence of training. We found that units selective to faces emerge robustly in randomly initialized networks and that these units reproduce many characteristics observed in monkeys. This innate selectivity also enables the untrained network to perform face-detection tasks. Intriguingly, we observed that units selective to various non-face objects can also arise innately in untrained networks. Our results imply that the random feedforward connections in early, untrained deep neural networks may be sufficient for initializing primitive visual selectivity.

Unsupervised deep learning identifies semantic disentanglement in single inferotemporal face patch neurons

In order to better understand how the brain perceives faces, it is important to know what objective drives learning in the ventral visual stream. To answer this question, we model neural responses to faces in the macaque inferotemporal (IT) cortex with a deep self-supervised generative model, β-VAE, which disentangles sensory data into interpretable latent factors, such as gender or age. Our results demonstrate a strong correspondence between the generative factors discovered by β-VAE and those coded by single IT neurons, beyond that found for the baselines, including the handcrafted state-of-the-art model of face perception, the Active Appearance Model, and deep classifiers. Moreover, β-VAE is able to reconstruct novel face images using signals from just a handful of cells. Together our results imply that optimising the disentangling objective leads to representations that closely resemble those in the IT at the single unit level. This points at disentangling as a plausible learning objective for the visual brain.

Tuning in scene-preferring cortex for mid-level visual features gives rise to selectivity across multiple levels of stimulus complexity

The scene-preferring portion of the human ventral visual stream, known as the parahippocampal place area (PPA), responds to scenes and landmark objects, which tend to be large in real-world size, fixed in location, and inanimate. However, the PPA also exhibits preferences for low-level contour statistics, including rectilinearity and cardinal orientations, that are not directly predicted by theories of scene- and landmark-selectivity. It is unknown whether these divergent findings of both low- and high-level selectivity in the PPA can be explained by a unified computational theory. To address this issue, we fit hierarchical computational models of mid-level tuning to the image-evoked fMRI responses of the PPA, and we performed a series of high-throughput experiments on these models. Our findings show that hierarchical encoding models of the PPA exhibit emergent selectivity across multiple levels of complexity, giving rise to high-level preferences along dimensions of real-world size, fixedness, and naturalness/animacy as well as low-level preferences for rectilinear shapes and cardinal orientations. These results reconcile disparate theories of PPA function in a unified model of mid-level visual representation, and they demonstrate how multifaceted selectivity profiles naturally emerge from the hierarchical computations of visual cortex and the natural statistics of images.