Gabriel Kreiman

Christopher Wang, Vighnesh Subramaniam, Adam Uri Yaari et al. · harvard, mit

We create a reusable Transformer, BrainBERT, for intracranial recordings bringing modern representation learning approaches to neuroscience. Much like in NLP and speech recognition, this Transformer enables classifying complex concepts, i.e., decoding neural data, with higher accuracy and with much less data by being pretrained in an unsupervised manner on a large corpus of unannotated neural recordings. Our approach generalizes to new subjects with electrodes in new positions and to unrelated tasks showing that the representations robustly disentangle the neural signal. Just like in NLP where one can study language by investigating what a language model learns, this approach opens the door to investigating the brain by what a model of the brain learns. As a first step along this path, we demonstrate a new analysis of the intrinsic dimensionality of the computations in different areas of the brain. To construct these representations, we combine a technique for producing super-resolution spectrograms of neural data with an approach designed for generating contextual representations of audio by masking. In the future, far more concepts will be decodable from neural recordings by using representation learning, potentially unlocking the brain like language models unlocked language.

Spandan Madan, You Li, Mengmi Zhang et al.

We present a new perspective on bridging the generalization gap between biological and computer vision -- mimicking the human visual diet. While computer vision models rely on internet-scraped datasets, humans learn from limited 3D scenes under diverse real-world transformations with objects in natural context. Our results demonstrate that incorporating variations and contextual cues ubiquitous in the human visual training data (visual diet) significantly improves generalization to real-world transformations such as lighting, viewpoint, and material changes. This improvement also extends to generalizing from synthetic to real-world data -- all models trained with a human-like visual diet outperform specialized architectures by large margins when tested on natural image data. These experiments are enabled by our two key contributions: a novel dataset capturing scene context and diverse real-world transformations to mimic the human visual diet, and a transformer model tailored to leverage these aspects of the human visual diet. All data and source code can be accessed at https://github.com/Spandan-Madan/human_visual_diet.

Stephen Casper, Taylor Killian, Gabriel Kreiman et al.

Adversarial examples can be useful for identifying vulnerabilities in AI systems before they are deployed. In reinforcement learning (RL), adversarial policies can be developed by training an adversarial agent to minimize a target agent's rewards. Prior work has studied black-box versions of these attacks where the adversary only observes the world state and treats the target agent as any other part of the environment. However, this does not take into account additional structure in the problem. In this work, we study white-box adversarial policies and show that having access to a target agent's internal state can be useful for identifying its vulnerabilities. We make two contributions. (1) We introduce white-box adversarial policies where an attacker observes both a target's internal state and the world state at each timestep. We formulate ways of using these policies to attack agents in 2-player games and text-generating language models. (2) We demonstrate that these policies can achieve higher initial and asymptotic performance against a target agent than black-box controls. Code is available at https://github.com/thestephencasper/lm_white_box_attacks

Zhiwei Ding, Xuezhe Ren, Erwan David et al.

Visual search is a ubiquitous challenge in natural vision, including daily tasks such as finding a friend in a crowd or searching for a car in a parking lot. Human rely heavily on relevant target features to perform goal-directed visual search. Meanwhile, context is of critical importance for locating a target object in complex scenes as it helps narrow down the search area and makes the search process more efficient. However, few works have combined both target and context information in visual search computational models. Here we propose a zero-shot deep learning architecture, TCT (Target and Context-aware Transformer), that modulates self attention in the Vision Transformer with target and contextual relevant information to enable human-like zero-shot visual search performance. Target modulation is computed as patch-wise local relevance between the target and search images, whereas contextual modulation is applied in a global fashion. We conduct visual search experiments on TCT and other competitive visual search models on three natural scene datasets with varying levels of difficulty. TCT demonstrates human-like performance in terms of search efficiency and beats the SOTA models in challenging visual search tasks. Importantly, TCT generalizes well across datasets with novel objects without retraining or fine-tuning. Furthermore, we also introduce a new dataset to benchmark models for invariant visual search under incongruent contexts. TCT manages to search flexibly via target and context modulation, even under incongruent contexts.

Parantak Singh, You Li, Ankur Sikarwar et al.

Curriculum design is a fundamental component of education. For example, when we learn mathematics at school, we build upon our knowledge of addition to learn multiplication. These and other concepts must be mastered before our first algebra lesson, which also reinforces our addition and multiplication skills. Designing a curriculum for teaching either a human or a machine shares the underlying goal of maximizing knowledge transfer from earlier to later tasks, while also minimizing forgetting of learned tasks. Prior research on curriculum design for image classification focuses on the ordering of training examples during a single offline task. Here, we investigate the effect of the order in which multiple distinct tasks are learned in a sequence. We focus on the online class-incremental continual learning setting, where algorithms or humans must learn image classes one at a time during a single pass through a dataset. We find that curriculum consistently influences learning outcomes for humans and for multiple continual machine learning algorithms across several benchmark datasets. We introduce a novel-object recognition dataset for human curriculum learning experiments and observe that curricula that are effective for humans are highly correlated with those that are effective for machines. As an initial step towards automated curriculum design for online class-incremental learning, we propose a novel algorithm, dubbed Curriculum Designer (CD), that designs and ranks curricula based on inter-class feature similarities. We find significant overlap between curricula that are empirically highly effective and those that are highly ranked by our CD. Our study establishes a framework for further research on teaching humans and machines to learn continuously using optimized curricula.

Trenton Bricken, Xander Davies, Deepak Singh et al.

Continual learning is a problem for artificial neural networks that their biological counterparts are adept at solving. Building on work using Sparse Distributed Memory (SDM) to connect a core neural circuit with the powerful Transformer model, we create a modified Multi-Layered Perceptron (MLP) that is a strong continual learner. We find that every component of our MLP variant translated from biology is necessary for continual learning. Our solution is also free from any memory replay or task information, and introduces novel methods to train sparse networks that may be broadly applicable.

Xiao Liu, Soumick Sarker, Ankur Sikarwar et al.

Humans rarely perceive objects in isolation but interpret scenes through relationships among co-occurring elements. How such contextual knowledge is acquired without explicit supervision remains unclear. Here we combine human psychophysics experiments with computational modelling to study the emergence of contextual reasoning. Participants were exposed to novel objects embedded in naturalistic scenes that followed predefined contextual rules capturing global context, local context and crowding. After viewing short training videos, participants completed a "lift-the-flap" task in which a hidden object had to be inferred from the surrounding context under variations in size, resolution and spatial arrangement. Humans rapidly learned these contextual associations without labels or feedback and generalised robustly across contextual changes. We then introduce SeCo (Self-supervised learning for Context Reasoning), a biologically inspired model that learns contextual relationships from complex scenes. SeCo encodes targets and context with separate vision encoders and stores latent contextual priors in a learnable external memory module. Given contextual cues, the model retrieves likely object representations to infer hidden targets. SeCo outperforms state-of-the-art self-supervised learning approaches and predicts object placements most consistent with human behaviour, highlighting the central role of contextual associations in scene understanding.

Ravi Srinivasan, Francesca Mignacco, Martino Sorbaro et al.

"Forward-only" algorithms, which train neural networks while avoiding a backward pass, have recently gained attention as a way of solving the biologically unrealistic aspects of backpropagation. Here, we first address compelling challenges related to the "forward-only" rules, which include reducing the performance gap with backpropagation and providing an analytical understanding of their dynamics. To this end, we show that the forward-only algorithm with top-down feedback is well-approximated by an "adaptive-feedback-alignment" algorithm, and we analytically track its performance during learning in a prototype high-dimensional setting. Then, we compare different versions of forward-only algorithms, focusing on the Forward-Forward and PEPITA frameworks, and we show that they share the same learning principles. Overall, our work unveils the connections between three key neuro-inspired learning rules, providing a link between "forward-only" algorithms, i.e., Forward-Forward and PEPITA, and an approximation of backpropagation, i.e., Feedback Alignment.

Mengmi Zhang, Elisa Pavarino, Xiao Liu et al.

As AI becomes increasingly embedded in daily life, ascertaining whether an agent is human is critical. We systematically benchmark AI's ability to imitate humans in three language tasks (image captioning, word association, conversation) and three vision tasks (color estimation, object detection, attention prediction), collecting data from 636 humans and 37 AI agents. Next, we conducted 72,191 Turing-like tests with 1,916 human judges and 10 AI judges. Current AIs are approaching the ability to convincingly impersonate humans and deceive human judges in both language and vision. Even simple AI judges outperformed humans in distinguishing AI from human responses. Imitation ability showed minimal correlation with conventional AI performance metrics, suggesting that passing as human is an important independent evaluation criterion. The large-scale Turing datasets and metrics introduced here offer valuable benchmarks for assessing human-likeness in AI and highlight the importance of rigorous, quantitative imitation tests for AI development.

Vedaant Jain, Felipe dos Santos Alves Feitosa, Gabriel Kreiman

Despite significant advancements in image segmentation and object detection, understanding complex scenes remains a significant challenge. Here, we focus on graphical humor as a paradigmatic example of image interpretation that requires elucidating the interaction of different scene elements in the context of prior cognitive knowledge. This paper introduces \textbf{HumorDB}, a novel, controlled, and carefully curated dataset designed to evaluate and advance visual humor understanding by AI systems. The dataset comprises diverse images spanning photos, cartoons, sketches, and AI-generated content, including minimally contrastive pairs where subtle edits differentiate between humorous and non-humorous versions. We evaluate humans, state-of-the-art vision models, and large vision-language models on three tasks: binary humor classification, funniness rating prediction, and pairwise humor comparison. The results reveal a gap between current AI systems and human-level humor understanding. While pretrained vision-language models perform better than vision-only models, they still struggle with abstract sketches and subtle humor cues. Analysis of attention maps shows that even when models correctly classify humorous images, they often fail to focus on the precise regions that make the image funny. Preliminary mechanistic interpretability studies and evaluation of model explanations provide initial insights into how different architectures process humor. Our results identify promising trends and current limitations, suggesting that an effective understanding of visual humor requires sophisticated architectures capable of detecting subtle contextual features and bridging the gap between visual perception and abstract reasoning. All the code and data are available here: \href{https://github.com/kreimanlab/HumorDB}{https://github.com/kreimanlab/HumorDB}

Stephen Casper, Max Nadeau, Dylan Hadfield-Menell et al.

The literature on adversarial attacks in computer vision typically focuses on pixel-level perturbations. These tend to be very difficult to interpret. Recent work that manipulates the latent representations of image generators to create "feature-level" adversarial perturbations gives us an opportunity to explore perceptible, interpretable adversarial attacks. We make three contributions. First, we observe that feature-level attacks provide useful classes of inputs for studying representations in models. Second, we show that these adversaries are uniquely versatile and highly robust. We demonstrate that they can be used to produce targeted, universal, disguised, physically-realizable, and black-box attacks at the ImageNet scale. Third, we show how these adversarial images can be used as a practical interpretability tool for identifying bugs in networks. We use these adversaries to make predictions about spurious associations between features and classes which we then test by designing "copy/paste" attacks in which one natural image is pasted into another to cause a targeted misclassification. Our results suggest that feature-level attacks are a promising approach for rigorous interpretability research. They support the design of tools to better understand what a model has learned and diagnose brittle feature associations. Code is available at https://github.com/thestephencasper/feature_level_adv

Shashi Kant Gupta, Mengmi Zhang, Chia-Chien Wu et al.

Visual search is a ubiquitous and often challenging daily task, exemplified by looking for the car keys at home or a friend in a crowd. An intriguing property of some classical search tasks is an asymmetry such that finding a target A among distractors B can be easier than finding B among A. To elucidate the mechanisms responsible for asymmetry in visual search, we propose a computational model that takes a target and a search image as inputs and produces a sequence of eye movements until the target is found. The model integrates eccentricity-dependent visual recognition with target-dependent top-down cues. We compared the model against human behavior in six paradigmatic search tasks that show asymmetry in humans. Without prior exposure to the stimuli or task-specific training, the model provides a plausible mechanism for search asymmetry. We hypothesized that the polarity of search asymmetry arises from experience with the natural environment. We tested this hypothesis by training the model on augmented versions of ImageNet where the biases of natural images were either removed or reversed. The polarity of search asymmetry disappeared or was altered depending on the training protocol. This study highlights how classical perceptual properties can emerge in neural network models, without the need for task-specific training, but rather as a consequence of the statistical properties of the developmental diet fed to the model. All source code and data are publicly available at https://github.com/kreimanlab/VisualSearchAsymmetry.

Philipp Bomatter, Mengmi Zhang, Dimitar Karev et al.

Context is of fundamental importance to both human and machine vision; e.g., an object in the air is more likely to be an airplane than a pig. The rich notion of context incorporates several aspects including physics rules, statistical co-occurrences, and relative object sizes, among others. While previous work has focused on crowd-sourced out-of-context photographs from the web to study scene context, controlling the nature and extent of contextual violations has been a daunting task. Here we introduce a diverse, synthetic Out-of-Context Dataset (OCD) with fine-grained control over scene context. By leveraging a 3D simulation engine, we systematically control the gravity, object co-occurrences and relative sizes across 36 object categories in a virtual household environment. We conducted a series of experiments to gain insights into the impact of contextual cues on both human and machine vision using OCD. We conducted psychophysics experiments to establish a human benchmark for out-of-context recognition, and then compared it with state-of-the-art computer vision models to quantify the gap between the two. We propose a context-aware recognition transformer model, fusing object and contextual information via multi-head attention. Our model captures useful information for contextual reasoning, enabling human-level performance and better robustness in out-of-context conditions compared to baseline models across OCD and other out-of-context datasets. All source code and data are publicly available at https://github.com/kreimanlab/WhenPigsFlyContext

Morgan B. Talbot, Rushikesh Zawar, Rohil Badkundri et al.

Our brains extract durable, generalizable knowledge from transient experiences of the world. Artificial neural networks come nowhere close to this ability. When tasked with learning to classify objects by training on non-repeating video frames in temporal order (online stream learning), models that learn well from shuffled datasets catastrophically forget old knowledge upon learning new stimuli. We propose a new continual learning algorithm, Compositional Replay Using Memory Blocks (CRUMB), which mitigates forgetting by replaying feature maps reconstructed by combining generic parts. CRUMB concatenates trainable and re-usable "memory block" vectors to compositionally reconstruct feature map tensors in convolutional neural networks. Storing the indices of memory blocks used to reconstruct new stimuli enables memories of the stimuli to be replayed during later tasks. This reconstruction mechanism also primes the neural network to minimize catastrophic forgetting by biasing it towards attending to information about object shapes more than information about image textures, and stabilizes the network during stream learning by providing a shared feature-level basis for all training examples. These properties allow CRUMB to outperform an otherwise identical algorithm that stores and replays raw images, while occupying only 3.6% as much memory. We stress-tested CRUMB alongside 13 competing methods on 7 challenging datasets. To address the limited number of existing online stream learning datasets, we introduce 2 new benchmarks by adapting existing datasets for stream learning. With only 3.7-4.1% as much memory and 15-43% as much runtime, CRUMB mitigates catastrophic forgetting more effectively than the state-of-the-art. Our code is available at https://github.com/MorganBDT/crumb.git.

Mengmi Zhang, Claire Tseng, Gabriel Kreiman

Context plays an important role in visual recognition. Recent studies have shown that visual recognition networks can be fooled by placing objects in inconsistent contexts (e.g., a cow in the ocean). To model the role of contextual information in visual recognition, we systematically investigated ten critical properties of where, when, and how context modulates recognition, including the amount of context, context and object resolution, geometrical structure of context, context congruence, and temporal dynamics of contextual modulation. The tasks involved recognizing a target object surrounded with context in a natural image. As an essential benchmark, we conducted a series of psychophysics experiments where we altered one aspect of context at a time, and quantified recognition accuracy. We propose a biologically-inspired context-aware object recognition model consisting of a two-stream architecture. The model processes visual information at the fovea and periphery in parallel, dynamically incorporates object and contextual information, and sequentially reasons about the class label for the target object. Across a wide range of behavioral tasks, the model approximates human level performance without retraining for each task, captures the dependence of context enhancement on image properties, and provides initial steps towards integrating scene and object information for visual recognition. All source code and data are publicly available: https://github.com/kreimanlab/Put-In-Context.

Morgan B. Talbot, Gabriel Kreiman, James J. DiCarlo et al.

The currently leading artificial neural network models of the visual ventral stream - which are derived from a combination of performance optimization and robustification methods - have demonstrated a remarkable degree of behavioral alignment with humans on visual categorization tasks. We show that image perturbations generated by these models can enhance the ability of humans to accurately report the ground truth class. Furthermore, we find that the same models can also be used out-of-the-box to predict the proportion of correct human responses to individual images, providing a simple, human-aligned estimator of the relative difficulty of each image. Motivated by these observations, we propose to augment visual learning in humans in a way that improves human categorization accuracy at test time. Our learning augmentation approach consists of (i) selecting images based on their model-estimated recognition difficulty, and (ii) applying image perturbations that aid recognition for novice learners. We find that combining these model-based strategies leads to categorization accuracy gains of 33-72% relative to control subjects without these interventions, on unmodified, randomly selected held-out test images. Beyond the accuracy gain, the training time for the augmented learning group was also shortened by 20-23%, despite both groups completing the same number of training trials. We demonstrate the efficacy of our approach in a fine-grained categorization task with natural images, as well as two tasks in clinically relevant image domains - histology and dermoscopy - where visual learning is notoriously challenging. To the best of our knowledge, our work is the first application of artificial neural networks to increase visual learning performance in humans by enhancing category-specific image features.

Serena Bono, Spandan Madan, Ishaan Grover et al.

Is it better to perform tennis training in a pristine indoor environment or a noisy outdoor one? To model this problem, here we investigate whether shifts in the transition probabilities between the training and testing environments in reinforcement learning problems can lead to better performance under certain conditions. We generate new Markov Decision Processes (MDPs) starting from a given MDP, by adding quantifiable, parametric noise into the transition function. We refer to this process as Noise Injection and the resulting environments as δ-environments. This process allows us to create variations of the same environment with quantitative control over noise serving as a metric of distance between environments. Conventional wisdom suggests that training and testing on the same MDP should yield the best results. In stark contrast, we observe that agents can perform better when trained on the noise-free environment and tested on the noisy δ-environments, compared to training and testing on the same δ-environments. We confirm that this finding extends beyond noise variations: it is possible to showcase the same phenomenon in ATARI game variations including varying Ghost behaviour in PacMan, and Paddle behaviour in Pong. We demonstrate this intriguing behaviour across 60 different variations of ATARI games, including PacMan, Pong, and Breakout. We refer to this phenomenon as the Indoor-Training Effect. Code to reproduce our experiments and to implement Noise Injection can be found at https://bit.ly/3X6CTYk.

Lorenzo Tausani, Paolo Muratore, Morgan B. Talbot et al.

Uncovering which features' combinations high-level visual units encode is critical to understand how images are transformed into representations that support recognition. While existing feature visualization approaches typically infer a unit's most exciting images, this is insufficient to reveal the manifold of transformations under which responses remain invariant, which is key to generalization in vision. Here we introduce Stretch-and-Squeeze (SnS), an unbiased, model-agnostic, and gradient-free framework to systematically characterize a unit's invariance landscape and its vulnerability to adversarial perturbations in both biological and artificial visual systems. SnS frames these transformations as bi-objective optimization problems. To probe invariance, SnS seeks image perturbations that maximally alter the representation of a reference stimulus in a given processing stage while preserving unit activation. To probe adversarial sensitivity, SnS seeks perturbations that minimally alter the stimulus while suppressing unit activation. Applied to convolutional neural networks (CNNs), SnS revealed image variations that were further from a reference image in pixel-space than those produced by affine transformations, while more strongly preserving the target unit's response. The discovered invariant images differed dramatically depending on the choice of image representation used for optimization: pixel-level changes primarily affected luminance and contrast, while stretching mid- and late-layer CNN representations altered texture and pose respectively. Notably, the invariant images from robust networks were more recognizable by human subjects than those from standard networks, supporting the higher fidelity of robust CNNs as models of the visual system.

Jiaqi Shang, Gabriel Kreiman, Haim Sompolinsky

Humans readily generalize abstract relations, such as recognizing "constant" in shape or color, whereas neural networks struggle, limiting their flexible reasoning. To investigate mechanisms underlying such generalization, we introduce SimplifiedRPM, a novel benchmark for systematically evaluating abstract relational reasoning, addressing limitations in prior datasets. In parallel, we conduct human experiments to quantify relational difficulty, enabling direct model-human comparisons. Testing four models, ResNet-50, Vision Transformer, Wild Relation Network, and Scattering Compositional Learner (SCL), we find that SCL generalizes best and most closely aligns with human behavior. Using a geometric approach, we identify key representation properties that accurately predict generalization and uncover a fundamental trade-off between signal and dimensionality: novel relations compress into training-induced subspaces. Layer-wise analysis reveals where relational structure emerges, highlights bottlenecks, and generates concrete hypotheses about abstract reasoning in the brain. Motivated by these insights, we propose SNRloss, a novel objective explicitly balancing representation geometry. Our results establish a geometric foundation for relational reasoning, paving the way for more human-like visual reasoning in AI and opening promising avenues for extending geometric analysis to broader cognitive tasks.

Vighnesh Subramaniam, Colin Conwell, Christopher Wang et al.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoencephalography (SEEG) recordings taken while human subjects watched movies. We operationalize sites of multimodal integration as regions where a multimodal vision-language model predicts recordings better than unimodal language, unimodal vision, or linearly-integrated language-vision models. Our target DNN models span different architectures (e.g., convolutional networks and transformers) and multimodal training techniques (e.g., cross-attention and contrastive learning). As a key enabling step, we first demonstrate that trained vision and language models systematically outperform their randomly initialized counterparts in their ability to predict SEEG signals. We then compare unimodal and multimodal models against one another. Because our target DNN models often have different architectures, number of parameters, and training sets (possibly obscuring those differences attributable to integration), we carry out a controlled comparison of two models (SLIP and SimCLR), which keep all of these attributes the same aside from input modality. Using this approach, we identify a sizable number of neural sites (on average 141 out of 1090 total sites or 12.94%) and brain regions where multimodal integration seems to occur. Additionally, we find that among the variants of multimodal training techniques we assess, CLIP-style training is the best suited for downstream prediction of the neural activity in these sites.

Spandan Madan, Will Xiao, Mingran Cao et al.

We characterized the generalization capabilities of DNN-based encoding models when predicting neuronal responses from the visual cortex. We collected \textit{MacaqueITBench}, a large-scale dataset of neural population responses from the macaque inferior temporal (IT) cortex to over $300,000$ images, comprising $8,233$ unique natural images presented to seven monkeys over $109$ sessions. Using \textit{MacaqueITBench}, we investigated the impact of distribution shifts on models predicting neural activity by dividing the images into Out-Of-Distribution (OOD) train and test splits. The OOD splits included several different image-computable types including image contrast, hue, intensity, temperature, and saturation. Compared to the performance on in-distribution test images -- the conventional way these models have been evaluated -- models performed worse at predicting neuronal responses to out-of-distribution images, retaining as little as $20\%$ of the performance on in-distribution test images. The generalization performance under OOD shifts can be well accounted by a simple image similarity metric -- the cosine distance between image representations extracted from a pre-trained object recognition model is a strong predictor of neural predictivity under different distribution shifts. The dataset of images, neuronal firing rate recordings, and computational benchmarks are hosted publicly at: https://bit.ly/3zeutVd.

Giorgia Dellaferrera, Gabriel Kreiman

Supervised learning in artificial neural networks typically relies on backpropagation, where the weights are updated based on the error-function gradients and sequentially propagated from the output layer to the input layer. Although this approach has proven effective in a wide domain of applications, it lacks biological plausibility in many regards, including the weight symmetry problem, the dependence of learning on non-local signals, the freezing of neural activity during error propagation, and the update locking problem. Alternative training schemes have been introduced, including sign symmetry, feedback alignment, and direct feedback alignment, but they invariably rely on a backward pass that hinders the possibility of solving all the issues simultaneously. Here, we propose to replace the backward pass with a second forward pass in which the input signal is modulated based on the error of the network. We show that this novel learning rule comprehensively addresses all the above-mentioned issues and can be applied to both fully connected and convolutional models. We test this learning rule on MNIST, CIFAR-10, and CIFAR-100. These results help incorporate biological principles into machine learning.

Ankur Sikarwar, Gabriel Kreiman

In recent years, multi-modal transformers have shown significant progress in Vision-Language tasks, such as Visual Question Answering (VQA), outperforming previous architectures by a considerable margin. This improvement in VQA is often attributed to the rich interactions between vision and language streams. In this work, we investigate the efficacy of co-attention transformer layers in helping the network focus on relevant regions while answering the question. We generate visual attention maps using the question-conditioned image attention scores in these co-attention layers. We evaluate the effect of the following critical components on visual attention of a state-of-the-art VQA model: (i) number of object region proposals, (ii) question part of speech (POS) tags, (iii) question semantics, (iv) number of co-attention layers, and (v) answer accuracy. We compare the neural network attention maps against human attention maps both qualitatively and quantitatively. Our findings indicate that co-attention transformer modules are crucial in attending to relevant regions of the image given a question. Importantly, we observe that the semantic meaning of the question is not what drives visual attention, but specific keywords in the question do. Our work sheds light on the function and interpretation of co-attention transformer layers, highlights gaps in current networks, and can guide the development of future VQA models and networks that simultaneously process visual and language streams.

Guy Ben-Yosef, Gabriel Kreiman, Shimon Ullman

In human vision objects and their parts can be visually recognized from purely spatial or purely temporal information but the mechanisms integrating space and time are poorly understood. Here we show that human visual recognition of objects and actions can be achieved by efficiently combining spatial and motion cues in configurations where each source on its own is insufficient for recognition. This analysis is obtained by identifying minimal videos: these are short and tiny video clips in which objects, parts, and actions can be reliably recognized, but any reduction in either space or time makes them unrecognizable. State-of-the-art deep networks for dynamic visual recognition cannot replicate human behavior in these configurations. This gap between humans and machines points to critical mechanisms in human dynamic vision that are lacking in current models.

Mengmi Zhang, Marcelo Armendariz, Will Xiao et al.

Primates constantly explore their surroundings via saccadic eye movements that bring different parts of an image into high resolution. In addition to exploring new regions in the visual field, primates also make frequent return fixations, revisiting previously foveated locations. We systematically studied a total of 44,328 return fixations out of 217,440 fixations. Return fixations were ubiquitous across different behavioral tasks, in monkeys and humans, both when subjects viewed static images and when subjects performed natural behaviors. Return fixations locations were consistent across subjects, tended to occur within short temporal offsets, and typically followed a 180-degree turn in saccadic direction. To understand the origin of return fixations, we propose a proof-of-principle, biologically-inspired and image-computable neural network model. The model combines five key modules: an image feature extractor, bottom-up saliency cues, task-relevant visual features, finite inhibition-of-return, and saccade size constraints. Even though there are no free parameters that are fine-tuned for each specific task, species, or condition, the model produces fixation sequences resembling the universal properties of return fixations. These results provide initial steps towards a mechanistic understanding of the trade-off between rapid foveal recognition and the need to scrutinize previous fixation locations.

Li Yuan, Will Xiao, Giorgia Dellaferrera et al.

Artificial neural networks (ANNs) are considered the current best models of biological vision. ANNs are the best predictors of neural activity in the ventral stream; moreover, recent work has demonstrated that ANN models fitted to neuronal activity can guide the synthesis of images that drive pre-specified response patterns in small neuronal populations. Despite the success in predicting and steering firing activity, these results have not been connected with perceptual or behavioral changes. Here we propose an array of methods for creating minimal, targeted image perturbations that lead to changes in both neuronal activity and perception as reflected in behavior. We generated 'deceptive images' of human faces, monkey faces, and noise patterns so that they are perceived as a different, pre-specified target category, and measured both monkey neuronal responses and human behavior to these images. We found several effective methods for changing primate visual categorization that required much smaller image change compared to untargeted noise. Our work shares the same goal with adversarial attack, namely the manipulation of images with minimal, targeted noise that leads ANN models to misclassify the images. Our results represent a valuable step in quantifying and characterizing the differences in perturbation robustness of biological and artificial vision.

Mengmi Zhang, Gabriel Kreiman

Can we infer intentions from a person's actions? As an example problem, here we consider how to decipher what a person is searching for by decoding their eye movement behavior. We conducted two psychophysics experiments where we monitored eye movements while subjects searched for a target object. We defined the fixations falling on \textit{non-target} objects as "error fixations". Using those error fixations, we developed a model (InferNet) to infer what the target was. InferNet uses a pre-trained convolutional neural network to extract features from the error fixations and computes a similarity map between the error fixations and all locations across the search image. The model consolidates the similarity maps across layers and integrates these maps across all error fixations. InferNet successfully identifies the subject's goal and outperforms competitive null models, even without any object-specific training on the inference task.

Vincent Jacquot, Zhuofan Ying, Gabriel Kreiman

Deep Learning has driven recent and exciting progress in computer vision, instilling the belief that these algorithms could solve any visual task. Yet, datasets commonly used to train and test computer vision algorithms have pervasive confounding factors. Such biases make it difficult to truly estimate the performance of those algorithms and how well computer vision models can extrapolate outside the distribution in which they were trained. In this work, we propose a new action classification challenge that is performed well by humans, but poorly by state-of-the-art Deep Learning models. As a proof-of-principle, we consider three exemplary tasks: drinking, reading, and sitting. The best accuracies reached using state-of-the-art computer vision models were 61.7%, 62.8%, and 76.8%, respectively, while human participants scored above 90% accuracy on the three tasks. We propose a rigorous method to reduce confounds when creating datasets, and when comparing human versus computer vision performance. Source code and datasets are publicly available.

Stephen Casper, Xavier Boix, Vanessa D'Amario et al.

A remarkable characteristic of overparameterized deep neural networks (DNNs) is that their accuracy does not degrade when the network's width is increased. Recent evidence suggests that developing compressible representations is key for adjusting the complexity of large networks to the learning task at hand. However, these compressible representations are poorly understood. A promising strand of research inspired from biology is understanding representations at the unit level as it offers a more granular and intuitive interpretation of the neural mechanisms. In order to better understand what facilitates increases in width without decreases in accuracy, we ask: Are there mechanisms at the unit level by which networks control their effective complexity as their width is increased? If so, how do these depend on the architecture, dataset, and training parameters? We identify two distinct types of "frivolous" units that proliferate when the network's width is increased: prunable units which can be dropped out of the network without significant change to the output and redundant units whose activities can be expressed as a linear combination of others. These units imply complexity constraints as the function the network represents could be expressed by a network without them. We also identify how the development of these units can be influenced by architecture and a number of training factors. Together, these results help to explain why the accuracy of DNNs does not degrade when width is increased and highlight the importance of frivolous units toward understanding implicit regularization in DNNs.

Mengmi Zhang, Tao Wang, Joo Hwee Lim et al.

Continual learning refers to the ability to acquire and transfer knowledge without catastrophically forgetting what was previously learned. In this work, we consider \emph{few-shot} continual learning in classification tasks, and we propose a novel method, Variational Prototype Replays, that efficiently consolidates and recalls previous knowledge to avoid catastrophic forgetting. In each classification task, our method learns a set of variational prototypes with their means and variances, where embedding of the samples from the same class can be represented in a prototypical distribution and class-representative prototypes are separated apart. To alleviate catastrophic forgetting, our method replays one sample per class from previous tasks, and correspondingly matches newly predicted embeddings to their nearest class-representative prototypes stored from previous tasks. Compared with recent continual learning approaches, our method can readily adapt to new tasks with more classes without requiring the addition of new units. Furthermore, our method is more memory efficient since only class-representative prototypes with their means and variances, as well as only one sample per class from previous tasks need to be stored. Without tampering with the performance on initial tasks, our method learns novel concepts given a few training examples of each class in new tasks.

Will Xiao, Gabriel Kreiman

A fundamental question for understanding brain function is what types of stimuli drive neurons to fire. In visual neuroscience, this question has also been posted as characterizing the receptive field of a neuron. The search for effective stimuli has traditionally been based on a combination of insights from previous studies, intuition, and luck. Recently, the same question has emerged in the study of units in convolutional neural networks (ConvNets), and together with this question a family of solutions were developed that are generally referred to as "feature visualization by activation maximization." We sought to bring in tools and techniques developed for studying ConvNets to the study of biological neural networks. However, one key difference that impedes direct translation of tools is that gradients can be obtained from ConvNets using backpropagation, but such gradients are not available from the brain. To circumvent this problem, we developed a method for gradient-free activation maximization by combining a generative neural network with a genetic algorithm. We termed this method XDream (EXtending DeepDream with real-time evolution for activation maximization), and we have shown that this method can reliably create strong stimuli for neurons in the macaque visual cortex (Ponce et al., 2019). In this paper, we describe extensive experiments characterizing the XDream method by using ConvNet units as in silico models of neurons. We show that XDream is applicable across network layers, architectures, and training sets; examine design choices in the algorithm; and provide practical guides for choosing hyperparameters in the algorithm. XDream is an efficient algorithm for uncovering neuronal tuning preferences in black-box networks using a vast and diverse stimulus space.

Mengmi Zhang, Claire Tseng, Karla Montejo et al.

Context reasoning is critical in a wide variety of applications where current inputs need to be interpreted in the light of previous experience and knowledge. Both spatial and temporal contextual information play a critical role in the domain of visual recognition. Here we investigate spatial constraints (what image features provide contextual information and where they are located), and temporal constraints (when different contextual cues matter) for visual recognition. The task is to reason about the scene context and infer what a target object hidden behind a flap is in a natural image. To tackle this problem, we first describe an online human psychophysics experiment recording active sampling via mouse clicks in lift-the-flap games and identify clicking patterns and features which are diagnostic for high contextual reasoning accuracy. As a proof of the usefulness of these clicking patterns and visual features, we extend a state-of-the-art recurrent model capable of attending to salient context regions, dynamically integrating useful information, making inferences, and predicting class label for the target object over multiple clicks. The proposed model achieves human-level contextual reasoning accuracy, shares human-like sampling behavior and learns interpretable features for contextual reasoning.

Mengmi Zhang, Gabriel Kreiman

Can we infer intentions from a person's actions? As an example problem, here we consider how to decipher what a person is searching for by decoding their eye movement behavior. We conducted two psychophysics experiments where we monitored eye movements while subjects searched for a target object. We defined the fixations falling on non-target objects as "error fixations". Using those error fixations, we developed a model (InferNet) to infer what the target was. InferNet uses a pre-trained convolutional neural network to extract features from the error fixations and computes a similarity map between the error fixations and all locations across the search image. The model consolidates the similarity maps across layers and integrates these maps across all error fixations. InferNet successfully identifies the subject's goal and outperforms competitive null models, even without any object-specific training on the inference task.

Mengmi Zhang, Jiashi Feng, Keng Teck Ma et al.

Searching for a target object in a cluttered scene constitutes a fundamental challenge in daily vision. Visual search must be selective enough to discriminate the target from distractors, invariant to changes in the appearance of the target, efficient to avoid exhaustive exploration of the image, and must generalize to locate novel target objects with zero-shot training. Previous work has focused on searching for perfect matches of a target after extensive category-specific training. Here we show for the first time that humans can efficiently and invariantly search for natural objects in complex scenes. To gain insight into the mechanisms that guide visual search, we propose a biologically inspired computational model that can locate targets without exhaustive sampling and generalize to novel objects. The model provides an approximation to the mechanisms integrating bottom-up and top-down signals during search in natural scenes.

William Lotter, Gabriel Kreiman, David Cox

While deep neural networks take loose inspiration from neuroscience, it is an open question how seriously to take the analogies between artificial deep networks and biological neuronal systems. Interestingly, recent work has shown that deep convolutional neural networks (CNNs) trained on large-scale image recognition tasks can serve as strikingly good models for predicting the responses of neurons in visual cortex to visual stimuli, suggesting that analogies between artificial and biological neural networks may be more than superficial. However, while CNNs capture key properties of the average responses of cortical neurons, they fail to explain other properties of these neurons. For one, CNNs typically require large quantities of labeled input data for training. Our own brains, in contrast, rarely have access to this kind of supervision, so to the extent that representations are similar between CNNs and brains, this similarity must arise via different training paths. In addition, neurons in visual cortex produce complex time-varying responses even to static inputs, and they dynamically tune themselves to temporal regularities in the visual environment. We argue that these differences are clues to fundamental differences between the computations performed in the brain and in deep networks. To begin to close the gap, here we study the emergent properties of a previously-described recurrent generative network that is trained to predict future video frames in a self-supervised manner. Remarkably, the model is able to capture a wide variety of seemingly disparate phenomena observed in visual cortex, ranging from single unit response dynamics to complex perceptual motion illusions. These results suggest potentially deep connections between recurrent predictive neural network models and the brain, providing new leads that can enrich both fields.

Kevin Wu, Eric Wu, Gabriel Kreiman

Advancements in convolutional neural networks (CNNs) have made significant strides toward achieving high performance levels on multiple object recognition tasks. While some approaches utilize information from the entire scene to propose regions of interest, the task of interpreting a particular region or object is still performed independently of other objects and features in the image. Here we demonstrate that a scene's 'gist' can significantly contribute to how well humans can recognize objects. These findings are consistent with the notion that humans foveate on an object and incorporate information from the periphery to aid in recognition. We use a biologically inspired two-part convolutional neural network ('GistNet') that models the fovea and periphery to provide a proof-of-principle demonstration that computational object recognition can significantly benefit from the gist of the scene as contextual information. Our model yields accuracy improvements of up to 50% in certain object categories when incorporating contextual gist, while only increasing the original model size by 5%. This proposed model mirrors our intuition about how the human visual system recognizes objects, suggesting specific biologically plausible constraints to improve machine vision and building initial steps towards the challenge of scene understanding.

Hanlin Tang, Martin Schrimpf, Bill Lotter et al.

Making inferences from partial information constitutes a critical aspect of cognition. During visual perception, pattern completion enables recognition of poorly visible or occluded objects. We combined psychophysics, physiology and computational models to test the hypothesis that pattern completion is implemented by recurrent computations and present three pieces of evidence that are consistent with this hypothesis. First, subjects robustly recognized objects even when rendered <15% visible, but recognition was largely impaired when processing was interrupted by backward masking. Second, invasive physiological responses along the human ventral cortex exhibited visually selective responses to partially visible objects that were delayed compared to whole objects, suggesting the need for additional computations. These physiological delays were correlated with the effects of backward masking. Third, state-of-the-art feed-forward computational architectures were not robust to partial visibility. However, recognition performance was recovered when the model was augmented with attractor-based recurrent connectivity. These results provide a strong argument of plausibility for the role of recurrent computations in making visual inferences from partial information.

Nicholas Cheney, Martin Schrimpf, Gabriel Kreiman

Deep convolutional neural networks are generally regarded as robust function approximators. So far, this intuition is based on perturbations to external stimuli such as the images to be classified. Here we explore the robustness of convolutional neural networks to perturbations to the internal weights and architecture of the network itself. We show that convolutional networks are surprisingly robust to a number of internal perturbations in the higher convolutional layers but the bottom convolutional layers are much more fragile. For instance, Alexnet shows less than a 30% decrease in classification performance when randomly removing over 70% of weight connections in the top convolutional or dense layers but performance is almost at chance with the same perturbation in the first convolutional layer. Finally, we suggest further investigations which could continue to inform the robustness of convolutional networks to internal perturbations.

William Lotter, Gabriel Kreiman, David Cox

While great strides have been made in using deep learning algorithms to solve supervised learning tasks, the problem of unsupervised learning - leveraging unlabeled examples to learn about the structure of a domain - remains a difficult unsolved challenge. Here, we explore prediction of future frames in a video sequence as an unsupervised learning rule for learning about the structure of the visual world. We describe a predictive neural network ("PredNet") architecture that is inspired by the concept of "predictive coding" from the neuroscience literature. These networks learn to predict future frames in a video sequence, with each layer in the network making local predictions and only forwarding deviations from those predictions to subsequent network layers. We show that these networks are able to robustly learn to predict the movement of synthetic (rendered) objects, and that in doing so, the networks learn internal representations that are useful for decoding latent object parameters (e.g. pose) that support object recognition with fewer training views. We also show that these networks can scale to complex natural image streams (car-mounted camera videos), capturing key aspects of both egocentric movement and the movement of objects in the visual scene, and the representation learned in this setting is useful for estimating the steering angle. Altogether, these results suggest that prediction represents a powerful framework for unsupervised learning, allowing for implicit learning of object and scene structure.

William Lotter, Gabriel Kreiman, David Cox

The ability to predict future states of the environment is a central pillar of intelligence. At its core, effective prediction requires an internal model of the world and an understanding of the rules by which the world changes. Here, we explore the internal models developed by deep neural networks trained using a loss based on predicting future frames in synthetic video sequences, using a CNN-LSTM-deCNN framework. We first show that this architecture can achieve excellent performance in visual sequence prediction tasks, including state-of-the-art performance in a standard 'bouncing balls' dataset (Sutskever et al., 2009). Using a weighted mean-squared error and adversarial loss (Goodfellow et al., 2014), the same architecture successfully extrapolates out-of-the-plane rotations of computer-generated faces. Furthermore, despite being trained end-to-end to predict only pixel-level information, our Predictive Generative Networks learn a representation of the latent structure of the underlying three-dimensional objects themselves. Importantly, we find that this representation is naturally tolerant to object transformations, and generalizes well to new tasks, such as classification of static images. Similar models trained solely with a reconstruction loss fail to generalize as effectively. We argue that prediction can serve as a powerful unsupervised loss for learning rich internal representations of high-level object features.