People infer rich social information from others' actions. These inferences are often constrained by the physical world: what agents can do, what obstacles permit, and how the physical actions of agents causally change an environment and other agents' mental states and behavior. We propose that such rich social perception is more than visual pattern matching, but rather a reasoning process grounded in an integration of intuitive psychology with intuitive physics. To test this hypothesis, we introduced PHASE (PHysically grounded Abstract Social Events), a large dataset of procedurally generated animations, depicting physically simulated two-agent interactions on a 2D surface. Each animation follows the style of the Heider and Simmel movie, with systematic variation in environment geometry, object dynamics, agent capacities, goals, and relationships (friendly/adversarial/neutral). We then present a computational model, SIMPLE, a physics-grounded Bayesian inverse planning model that integrates planning, probabilistic planning, and physics simulation to infer agents' goals and relations from their trajectories. Our experimental results showed that SIMPLE achieved high accuracy and agreement with human judgments across diverse scenarios, while feedforward baseline models – including strong vision-language models – and physics-agnostic inverse planning failed to achieve human-level performance and did not align with human judgments. These results suggest that our model provides a computational account for how people understand physically grounded social scenes by inverting a generative model of physics and agents.

As Artificial Neural Networks (ANNs) reach human and super-human performance, the convergence between the visual strategies learned by ANNs and human perception is a fundamental step to allow their deployment as models of biological vision and their integration into assistive and rehabilitation technologies. In this work, we investigate the alignment between visual strategies learned by an ANN, specifically the Temporal Shift Module (TSM), and human gaze-based visual attention, during classification of social interactions from videos. We present preliminary results comparing the similarity between saliency maps generated by the ANN model after fine-tuning on the social interaction classification task and ground-truth fixation maps of human gaze data.

We demonstrate that architectures which traditionally are considered to be ill-suited for a task can be trained using inductive biases from another architecture. We call a network untrainable when it overfits, underfits, or converges to poor results even when tuning their hyperparameters. For example, fully connected networks overfit on object recognition while deep convolutional networks without residual connections underfit. The traditional answer is to change the architecture to impose some inductive bias, although the nature of that bias is unknown. We introduce guidance, where a guide network steers a target network using a neural distance function. The target minimizes its task loss plus a layerwise representational similarity against the frozen guide. If the guide is trained, this transfers over the architectural prior and knowledge of the guide to the target. If the guide is untrained, this transfers over only part of the architectural prior of the guide. We show that guidance prevents FCN overfitting on ImageNet, narrows the vanilla RNN–Transformer gap, boosts plain CNNs toward ResNet accuracy, and aids Transformers on RNN-favored tasks. We further identify that guidance-driven initialization alone can mitigate FCN overfitting. Our method provides a mathematical tool to investigate priors and architectures, and in the long term, could automate architecture design.

We present a self-supervised framework that learns population-level codes for arbitrary ensembles of neural recordings at scale. We address key challenges in scaling models with neural time-series data, namely, sparse and variable electrode distribution across subjects and datasets. The Population Transformer (PopT) stacks on top of pretrained temporal embeddings and enhances downstream decoding by enabling learned aggregation of multiple spatially-sparse data channels. The pretrained PopT lowers the amount of data required for downstream decoding experiments, while increasing accuracy, even on held-out subjects and tasks. Compared to end-to-end methods, this approach is computationally lightweight, while achieving similar or better decoding performance. We further show how our framework is generalizable to multiple time-series embeddings and neural data modalities. Beyond decoding, we interpret the pretrained and fine-tuned PopT models to show how they can be used to extract neuroscience insights from large amounts of data. We release our code as well as a pretrained PopT to enable off-the-shelf improvements in multi-channel intracranial data decoding and interpretability. Code is available at https://github.com/czlwang/PopulationTransformer.

A standard assumption in deep learning is that the inductive bias introduced by a neural network architecture must persist from training through inference. The architecture you train with is the architecture you deploy. This assumption constrains the community from selecting architectures that may have desirable efficiency or design properties due to difficulties with optimization. We challenge this assumption with Network of Theseus (NoT), a method for progressively converting a trained, or even untrained, guide network architecture part-by-part into an entirely different target network architecture while preserving the performance of the guide network. At each stage, components in the guide network architecture are incrementally replaced with target architecture modules and aligned via representational similarity metrics. This procedure largely preserves the functionality of the guide network even under substantial architectural changes-for example, converting a convolutional network into a multilayer perceptron, or GPT-2 into a recurrent neural network. By decoupling optimization from deployment, NoT expands the space of viable inference-time architectures, opening opportunities for better accuracy-efficiency tradeoffs and enabling more directed exploration of the architectural design space.

Artificial Neural Networks (ANNs) are considered the best models of human vision in many visual tasks. However, there is still a significant gap between humans and ANNs in visual tasks involving social perception and cognition. In this paper, we investigate the alignment between humans and different ANNs architectures, i.e., convolutional neural networks (CNNs) and transformers, during the classification of social interactions from videos. Specifically, we provide a novel dataset of videos and human gaze data during the classification of help and hinder social interactions, and we evaluate the human-model alignment in terms of classification accuracy and visual attention. We show that our proposed dataset and experimental protocol can enable comparison between different models' alignment to humans in terms of both static and dynamic visual attention. Our results suggest that a higher classification accuracy may be correlated with a higher similarity of the model's attention to human gaze-based saliency maps. However, a stronger alignment of human-model attention is not always predictive of a better classification. Data and code are available at: humanmodelvsa.github.io/HumanModelVSA.

High-resolution neural datasets enable foundation models for the next generation of brain-computer interfaces and neurological treatments. The community requires rigorous benchmarks to discriminate between competing modeling approaches, yet no standardized evaluation frameworks exist for intracranial EEG (iEEG) recordings. To address this gap, we present Neuroprobe: a suite of decoding tasks for studying multi-modal language processing in the brain. Unlike scalp EEG, intracranial EEG requires invasive surgery to implant electrodes that record neural activity directly from the brain with minimal signal distortion. Neuroprobe is built on the BrainTreebank dataset, which consists of over 40 hours of iEEG recordings from 10 human subjects performing a naturalistic movie viewing task. Neuroprobe serves two critical functions. First, it is a source from which neuroscience insights can be drawn. The high temporal and spatial resolution of the labeled iEEG allows researchers to systematically determine when and where computations for each aspect of language processing occur in the brain by measuring the decodability of each feature across time and all electrode locations. Using Neuroprobe, we visualize how information flows from key language and audio processing sites in the superior temporal gyrus to sites in the prefrontal cortex. We also demonstrate the time evolution of processing from simple auditory features (e.g., pitch and volume) to more complex language features (e.g., part of speech) in a purely data-driven manner. Second, as the field moves toward neural foundation models trained on large-scale datasets, Neuroprobe provides a rigorous framework for comparing competing architectures and training protocols. We make the code for Neuroprobe openly available, aiming to enable rapid progress in the field of iEEG foundation models. Public leaderboard: https://neuroprobe.dev/

Even on out-of-domain image captioning datasets such as nocaps, models often outperform humans according to captioning metrics like CIDEr. Yet, in real world conditions, model captions are often wrong. We demonstrate that this performance deficit exists by introducing a new dataset and a new captioning metric. We introduce a new dataset, called ObjectNet Captions, that reduces spurious correlations which machines often exploit. We show the shortcomings of current captioning metrics with a head-to-head experiment against humans, where we find that humans rate human-generated captions as being of much higher quality than machine captions. Driven by this, we introduce HUMANr, a new, highly robust, easy to replicate, and consistent metric, computed from head-to-head comparisons, which can be crowdsourced at low cost. We also develop tooling to automatically compute HUMANr. HUMANr is an absolute performance metric: driving it to 0 means that humans can no longer distinguish machine captions from human captions. No current metric provides such a fixed target to aim for along with knowledge of when captioning is solved in this sense. Moreover, HUMANr can reveal that humans still outperform machines, which no current metric is able to demonstrate. Existing metrics both overstate the performance of machine models and, at the same time, they inherently limit it. While most current metrics are saturated, HUMANr provides significant opportunities for further captioning research, thereby opening the door to new advances. ObjectNet Captions and HUMANr are made available to the research community.

Traditional security mechanisms isolate resources from users who should not access them. We reflect the compositional nature of such security mechanisms back into the structure of LLMs to build a provably secure LLM; that we term SecureLLM. Other approaches to LLM safety attempt to protect against bad actors or bad outcomes, but can only do so to an extent making them inappropriate for sensitive data. SecureLLM blends access security with fine-tuning methods. Each data silo has associated with it a separate fine-tuning and a user has access only to the collection of fine-tunings that they have permission for. The model must then perform on compositional tasks at the intersection of those data silos with the combination of those individual fine-tunings. While applicable to any task like document QA or making API calls, in this work we concern ourselves with models that learn the layouts of new SQL databases to provide natural-language-to-SQL translation capabilities. Existing fine-tuning composition methods fail in this challenging environment, as they are not well-equipped for handling compositional tasks. Compositionality remains a challenge for LLMs. We contribute both a difficult new compositional natural-language-to-SQL translation task and a new perspective on LLM security that allows models to be deployed to secure environments today.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoen-cephalography (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.

When evaluating stimuli reconstruction results it is tempting to assume that higher fidelity text and image generation is due to an improved understanding of the brain or more powerful signal extraction from neural recordings. However, in practice, new reconstruction methods could improve performance for at least three other reasons: learning more about the distribution of stimuli, becoming better at reconstructing text or images in general, or exploiting weaknesses in current image and/or text evaluation metrics. Here we disentangle how much of the reconstruction is due to these other factors vs. productively using the neural recordings. We introduce BrainBits, a method that uses a bottleneck to quantify the amount of signal extracted from neural recordings that is actually necessary to reproduce a method's reconstruction fidelity. We find that it takes surprisingly little information from the brain to produce reconstructions with high fidelity. In these cases, it is clear that the priors of the methods' generative models are so powerful that the outputs they produce extrapolate far beyond the neural signal they decode. Given that reconstructing stimuli can be improved independently by either improving signal extraction from the brain or by building more powerful generative models, improving the latter may fool us into thinking we are improving the former. We propose that methods should report a method-specific random baseline, a reconstruction ceiling, and a curve of performance as a function of bottleneck size, with the ultimate goal of using more of the neural recordings.

We present the Brain Treebank, a large-scale dataset of electrophysiological neural responses, recorded from intracranial probes while 10 subjects watched one or more Hollywood movies. Subjects watched on average 2.6 Hollywood movies, for an average viewing time of 4.3 hours, and a total of 43 hours. The audio track for each movie was transcribed with manual corrections. Word onsets were manually annotated on spectrograms of the audio track for each movie. Each transcript was automatically parsed and manually corrected into the universal dependencies (UD) formalism, assigning a part of speech to every word and a dependency parse to every sentence. In total, subjects heard over 38,000 sentences (223,000 words), while they had on average 168 electrodes implanted. This is the largest dataset of intracranial recordings featuring grounded naturalistic language, one of the largest English UD treebanks in general, and one of only a few UD treebanks aligned to multimodal features. We hope that this dataset serves as a bridge between linguistic concepts, perception, and their neural representations. To that end, we present an analysis of which electrodes are sensitive to language features while also mapping out a rough time course of language processing across these electrodes. The Brain Treebank is available at https://BrainTreebank.dev/

The migration of conversational AI agents across different embodiments in order to maintain the continuity of the task has been recently explored to further improve user experience. However, these migratable agents lack contextual understanding of the user information and the migrated device during the dialog conversations with the user. This opens the question of how an agent might behave when migrated into an embodiment for contextually predicting the next utterance. We collected a dataset from the dialog conversations between crowdsourced workers with the migration context involving personal and non-personal utterances in different settings (public or private) of embodiment into which the agent migrated. We trained the generative and information retrieval models on the dataset using with and without migration context and report the results of both qualitative metrics and human evaluation. We believe that the migration dataset would be useful for training future migratable AI systems.

Humans and animals engage in rich social interactions. It is often theorized that a relatively small number of basic social interactions give rise to the full range of behavior observed. But no computational theory explaining how social interactions combine together has been proposed before. We do so here. We take a model, the Social MDP, which is able to express a range of social interactions, and extend it to represent linear combinations of social interactions. Practically for robotics applications, such models are now able to not just express that an agent should help another agent, but to express goal-centric social interactions. Perhaps an agent is helping someone get dressed, but preventing them from falling, and is happy to exchange stories in the meantime. How an agent responds socially, should depend on what it thinks the other agent is doing at that point in time. To encode this notion, we take linear combinations of social interactions as defined in Social MDPs, and compute the weights on those combinations on the fly depending on the estimated goals of other agents. This new model, the Linear Social MDP, enables zero-shot reasoning about complex social interactions, provides a mathematical basis for the long-standing intuition that social interactions should compose, and leads to interesting new behaviors that we validate using human observers. Complex social interactions are part of the future of intelligent agents, and having principled mathematical models built on a foundation like MDPs will make it possible to bring social interactions to every robotic application.

Curricula for goal-conditioned reinforcement learning agents typically rely on poor estimates of the agent's epistemic uncertainty or fail to consider the agents' epistemic uncertainty altogether, resulting in poor sample efficiency. We propose a novel algorithm, Query The Agent (QTA), which significantly improves sample efficiency by estimating the agent's epistemic uncertainty throughout the state space and setting goals in highly uncertain areas. Encouraging the agent to collect data in highly uncertain states allows the agent to improve its estimation of the value function rapidly. QTA utilizes a novel technique for estimating epistemic uncertainty, Predictive Uncertainty Networks (PUN), to allow QTA to assess the agent's uncertainty in all previously observed states. We demonstrate that QTA offers decisive sample efficiency improvements over preexisting methods.

Humans outperform object recognizers despite the fact that models perform well on current datasets. Numerous attempts have been made to create more challenging datasets by scaling them up from the web, exploring distribution shift, or adding controls for biases. The difficulty of each image in each dataset is not independently evaluated, nor is the concept of dataset difficulty as a whole well-posed. We develop a new dataset difficulty metric based on how long humans must view an image in order to classify a target object. Images whose objects can be recognized in 17ms are considered to be easier than those which require seconds of viewing time. Using 133,588 judgments on two major datasets, ImageNet and ObjectNet, we determine the distribution of image difficulties in those datasets, which we find varies wildly, but significantly undersamples hard images. Rather than hoping that distribution shift or other approaches will lead to hard datasets, we should measure the difficulty of datasets and seek to explicitly fill out the class of difficult examples. Analyzing model performance guided by image difficulty reveals that models tend to have lower performance and a larger generalization gap on harder images. Encouragingly for the biological validity of current architectures, much of the variance in human difficulty can be accounted for given an object recognizer by computing a combination of prediction depth, c-score, and adversarial robustness. We release a dataset of such judgments as a complementary metric to raw performance and a network’s ability to explain neural recordings. Such experiments with humans allow us to create a metric for progress in object recognition datasets, which we find are skewed toward easy examples, to test the biological validity of models in a novel way, and to develop tools for shaping datasets as they are being gathered to focus them on filling out the missing class of hard examples from today’s datasets. Dataset and analysis code can be found at https://github.com/image-flash/image-flash-2022.

We present an approach to bridge the gap between the computational models of human vision and the clinical practice on visual impairments (VI). In a nutshell, we propose to connect advances in neuroscience and machine learning to study the impact of VI on key functional competencies and improve treatment strategies. We review related literature, with the goal of promoting the full exploitation of Artificial Neural Network (ANN) models in meeting the needs of visually impaired individuals and the operators working in the field of visual rehabilitation. We first summarize the existing types of visual issues, the key functional vision-related tasks, and the current methodologies used for the assessment of both. Second, we explore the ANNs best suitable to model visual issues and to predict their impact on functional vision-related tasks, at a behavioral (including performance and attention measures) and neural level. We provide guidelines to inform the future research about developing and deploying ANNs for clinical applications targeting individuals affected by VI.