Seminars

Recent research has seen many behavioral comparisons between humans and deep neural networks (DNNs) in the domain of image classification. Often, comparison studies focus on the end-result of the learning process by measuring and comparing the similarities in the representations of object categories once they have been formed. However, the process of how these representations emerge—that is, the behavioral changes and intermediate stages observed during the acquisition—is less often directly and empirically compared. In this talk, I'm going to report a detailed investigation of the learning dynamics in human observers and various classic and state-of-the-art DNNs. We develop a constrained supervised learning environment to align learning-relevant conditions such as starting point, input modality, available input data and the feedback provided. Across the whole learning process we evaluate and compare how well learned representations can be generalized to previously unseen test data. Comparisons across the entire learning process indicate that DNNs demonstrate a level of data efficiency comparable to human learners, challenging some prevailing assumptions in the field. However, our results also reveal representational differences: while DNNs' learning is characterized by a pronounced generalisation lag, humans appear to immediately acquire generalizable representations without a preliminary phase of learning training set-specific information that is only later transferred to novel data.

Within the last decade, Deep Artificial Neural Networks (DNNs) have emerged as powerful computer vision systems that match or exceed human performance on many benchmark tasks such as image classification. But whether current DNNs are suitable computational models of the human visual system remains an open question: While DNNs have proven to be capable of predicting neural activations in primate visual cortex, psychophysical experiments have shown behavioral differences between DNNs and human subjects, as quantified by error consistency. Error consistency is typically measured by briefly presenting natural or corrupted images to human subjects and asking them to perform an n-way classification task under time pressure. But for how long should stimuli ideally be presented to guarantee a fair comparison with DNNs? Here we investigate the influence of presentation time on error consistency, to test the hypothesis that higher-level processing drives behavioral differences. We systematically vary presentation times of backward-masked stimuli from 8.3ms to 266ms and measure human performance and reaction times on natural, lowpass-filtered and noisy images. Our experiment constitutes a fine-grained analysis of human image classification under both image corruptions and time pressure, showing that even drastically time-constrained humans who are exposed to the stimuli for only two frames, i.e. 16.6ms, can still solve our 8-way classification task with success rates way above chance. We also find that human-to-human error consistency is already stable at 16.6ms.

Understanding Human ability to learn novel concepts from just a few sensory experiences is a fundamental problem in cognitive neuroscience. I will describe a recent work with Ben Sorcher and Surya Ganguli (PNAS, October 2022) in which we propose a simple, biologically plausible, and mathematically tractable neural mechanism for few-shot learning of naturalistic concepts. We posit that the concepts that can be learned from few examples are defined by tightly circumscribed manifolds in the neural firing-rate space of higher-order sensory areas. Discrimination between novel concepts is performed by downstream neurons implementing ‘prototype’ decision rule, in which a test example is classified according to the nearest prototype constructed from the few training examples. We show that prototype few-shot learning achieves high few-shot learning accuracy on natural visual concepts using both macaque inferotemporal cortex representations and deep neural network (DNN) models of these representations. We develop a mathematical theory that links few-shot learning to the geometric properties of the neural concept manifolds and demonstrate its agreement with our numerical simulations across different DNNs as well as different layers. Intriguingly, we observe striking mismatches between the geometry of manifolds in intermediate stages of the primate visual pathway and in trained DNNs. Finally, we show that linguistic descriptors of visual concepts can be used to discriminate images belonging to novel concepts, without any prior visual experience of these concepts (a task known as ‘zero-shot’ learning), indicated a remarkable alignment of manifold representations of concepts in visual and language modalities. I will discuss ongoing effort to extend this work to other high level cognitive tasks.

Behavioral studies have demonstrated that certain task features reliably enhance classification performance for challenging visual stimuli. These include extended image presentation time and the valid cueing of attention. Here, I will show how convolutional neural networks can be used as a model of the visual system that connects neural activity changes with such performance changes. Specifically, I will discuss how different anatomical forms of recurrence can account for better classification of noisy and degraded images with extended processing time. I will then show how experimentally-observed neural activity changes associated with feature attention lead to observed performance changes on detection tasks. I will also discuss the implications these results have for how we identify the neural mechanisms and architectures important for behavior.

Research in the brain and cognitive sciences attempts to uncover the neural mechanisms underlying intelligent behavior in domains such as vision. Due to the complexities of brain processing, studies necessarily had to start with a narrow scope of experimental investigation and computational modeling. I argue that it is time for our field to take the next step: build system models that capture a range of visual intelligence behaviors along with the underlying neural mechanisms. To make progress on system models, we propose integrative benchmarking – integrating experimental results from many laboratories into suites of benchmarks that guide and constrain those models at multiple stages and scales. We show-case this approach by developing Brain-Score benchmark suites for neural (spike rates) and behavioral experiments in the primate visual ventral stream. By systematically evaluating a wide variety of model candidates, we not only identify models beginning to match a range of brain data (~50% explained variance), but also discover that models’ brain scores are predicted by their object categorization performance (up to 70% ImageNet accuracy). Using the integrative benchmarks, we develop improved state-of-the-art system models that more closely match shallow recurrent neuroanatomy and early visual processing to predict primate temporal processing and become more robust, and require fewer supervised synaptic updates. Taken together, these integrative benchmarks and system models are first steps to modeling the complexities of brain processing in an entire domain of intelligence.

The brain represents the external world through the bottleneck of sensory organs. The network of hierarchically organized neurons is thought to recover the causes of sensory inputs to reconstruct the reality in the brain in idiosyncratic ways depending on individuals and their internal states. How can we understand the world model represented in an individual’s brain, or the neuroverse? My lab has been working on brain decoding of visual perception and subjective experiences such as imagery and dreaming using machine learning and deep neural network representations. In this talk, I will outline the progress of brain decoding methods and present how subjective experiences are externalized as images and how they could be shared across individuals via neural code conversion. The prospects of these approaches in basic science and neurotechnology will be discussed.

To better understand human scene understanding, we extracted features from images using CLIP, a neural network model of visual concept trained with supervision from natural language. We then constructed voxelwise encoding models to explain whole brain responses arising from viewing natural images from the Natural Scenes Dataset (NSD) - a large-scale fMRI dataset collected at 7T. Our results reveal that CLIP, as compared to convolution based image classification models such as ResNet or AlexNet, as well as language models such as BERT, gives rise to representations that enable better prediction performance - up to a 0.86 correlation with test data and an r-square of 0.75 - in higher-level visual cortex in humans. Moreover, CLIP representations explain distinctly unique variance in these higher-level visual areas as compared to models trained with only images or text. Control experiments show that the improvement in prediction observed with CLIP is not due to architectural differences (transformer vs. convolution) or to the encoding of image captions per se (vs. single object labels). Together our results indicate that CLIP and, more generally, multimodal models trained jointly on images and text, may serve as better candidate models of representation in human higher-level visual cortex. The bridge between language and vision provided by jointly trained models such as CLIP also opens up new and more semantically-rich ways of interpreting the visual brain.

How do humans perceive daily objects of various features and categorize these seemingly intuitive and effortless mental representations? Prior literature focusing on the role of the inferotemporal region (IT) has revealed object category clustering that is consistent with the semantic predefined structure (superordinate, ordinate, subordinate). It has however been debated whether the neural signals in the IT regions are a reflection of such categorical hierarchy [Wen et al.,2018; Bracci et al., 2017]. Visual attributes of images that correlated with semantic and category dimensions may have confounded these prior results. Our study aimed to address this debate by building and comparing models using the DNN AlexNet, to explain the variance in representational dissimilarity matrix (RDM) of neural signals in the IT region. We found that mid and high level perceptual attributes of the DNN model contribute the most to neural RDMs in the IT region. Semantic categories, as in predefined structure, were moderately correlated with mid to high DNN layers (r = [0.24 - 0.36]). Variance partitioning analysis also showed that the IT neural representations were mostly explained by DNN layers, while semantic categorical RDMs brought little additional information. In light of these results, we propose future works should focus more on the specific role IT plays in facilitating the extraction and coding of visual features that lead to the emergence of categorical conceptualizations.

It is an open question whether preferences for visual art can be lawfully predicted from the basic constituent elements of a visual image. Here, we developed and tested a computational framework to investigate how aesthetic values are formed. We show that it is possible to explain human preferences for a visual art piece based on a mixture of low- and high-level features of the image. Subjective value ratings could be predicted not only within but also across individuals, using a regression model with a common set of interpretable features. We also show that the features predicting aesthetic preference can emerge hierarchically within a deep convolutional neural network trained only for object recognition. Our findings suggest that human preferences for art can be explained at least in part as a systematic integration over the underlying visual features of an image.

Materials with complex appearances, like textiles and foodstuffs, pose challenges for conventional theories of vision. How does the brain learn to see properties of the world—like the glossiness of a surface—that cannot be measured by any other senses? Recent advances in unsupervised deep learning may help shed light on material perception. I will show how an unsupervised deep neural network trained on an artificial environment of surfaces that have different shapes, materials and lighting, spontaneously comes to encode those factors in its internal representations. Most strikingly, the model makes patterns of errors in its perception of material that follow, on an image-by-image basis, the patterns of errors made by human observers. Unsupervised deep learning may provide a coherent framework for how many perceptual dimensions form, in material perception and beyond.

Not all individuals are equally competent at recognizing the faces they interact with. Revealing how the brains of different individuals support variations in this ability is a crucial step to develop an understanding of real-world human visual behaviour. In this talk, I will present findings from a large high-density EEG dataset (>100k trials of participants processing various stimulus categories) and computational approaches which aimed to characterise the brain representations behind real-world proficiency of “super-recognizers”—individuals at the top of face recognition ability spectrum. Using decoding analysis of time-resolved EEG patterns, we predicted with high precision the trial-by-trial activity of super-recognizers participants, and showed that evidence for face recognition ability variations is disseminated along early, intermediate and late brain processing steps. Computational modeling of the underlying brain activity uncovered two representational signatures supporting higher face recognition ability—i) mid-level visual & ii) semantic computations. Both components were dissociable in brain processing-time (the first around the N170, the last around the P600) and levels of computations (the first emerging from mid-level layers of visual Convolutional Neural Networks, the last from a semantic model characterising sentence descriptions of images). I will conclude by presenting ongoing analyses from a well-known case of acquired prosopagnosia (PS) using similar computational modeling of high-density EEG activity.

Visual knowledge obtained from our lifelong experience of the world plays a critical role in our ability to build short-term memories. We propose a mechanistic explanation of how working memory (WM) representations are built from the latent representations of visual knowledge and can then be reconstructed. The proposed model, Memory for Latent Representations (MLR), features a variational autoencoder with an architecture that corresponds broadly to the human visual system and an activation-based binding pool of neurons that binds items’ attributes to tokenized representations. The simulation results revealed that shape information for stimuli that the model was trained on, can be encoded and retrieved efficiently from latents in higher levels of the visual hierarchy. On the other hand, novel patterns that are completely outside the training set can be stored from a single exposure using only latents from early layers of the visual system. Moreover, the representation of a given stimulus can have multiple codes, representing specific visual features such as shape or color, in addition to categorical information. Finally, we validated our model by testing a series of predictions against behavioral results acquired from WM tasks. The model provides a compelling demonstration of visual knowledge yielding the formation of compact visual representation for efficient memory encoding.

Classically, vision is seen as a cascade of local, feedforward computations. This framework has been tremendously successful, inspiring a wide range of ground-breaking findings in neuroscience and computer vision. Recently, feedforward Convolutional Neural Networks (ffCNNs), inspired by this classic framework, have revolutionized computer vision and been adopted as tools in neuroscience. However, despite these successes, there is much more to vision. I will present our work using visual crowding and related psychophysical effects as probes into visual processes that go beyond the classic framework. In crowding, perception of a target deteriorates in clutter. We focus on global aspects of crowding, in which perception of a small target is strongly modulated by the global configuration of elements across the visual field. We show that models based on the classic framework, including ffCNNs, cannot explain these effects for principled reasons and identify recurrent grouping and segmentation as a key missing ingredient. Then, we show that capsule networks, a recent kind of deep learning architecture combining the power of ffCNNs with recurrent grouping and segmentation, naturally explain these effects. We provide psychophysical evidence that humans indeed use a similar recurrent grouping and segmentation strategy in global crowding effects. In crowding, visual elements interfere across space. To study how elements interfere over time, we use the Sequential Metacontrast psychophysical paradigm, in which perception of visual elements depends on elements presented hundreds of milliseconds later. We psychophysically characterize the temporal structure of this interference and propose a simple computational model. Our results support the idea that perception is a discrete process. Together, the results presented here provide stepping-stones towards a fuller understanding of the visual system by suggesting architectural changes needed for more human-like neural computations.

Irina is a research scientist at DeepMind, where she works in the Froniers team. Her work aims to bring together insights from the fields of neuroscience and physics to advance general artificial intelligence through improved representation learning. Before joining DeepMind, Irina was a British Psychological Society Undergraduate Award winner for her achievements as an undergraduate student in Experimental Psychology at Westminster University, followed by a DPhil at the Oxford Centre for Computational Neuroscience and Artificial Intelligence, where she focused on understanding the computational principles underlying speech processing in the auditory brain. During her DPhil, Irina also worked on developing poker AI, applying machine learning in the finance sector, and working on speech recognition at Google Research."" https://arxiv.org/pdf/2006.14304.pdf