Michael C. Mozer

Nan Rosemary Ke, Silvia Chiappa, Jane Wang et al. · deepmind, mila

The fundamental challenge in causal induction is to infer the underlying graph structure given observational and/or interventional data. Most existing causal induction algorithms operate by generating candidate graphs and evaluating them using either score-based methods (including continuous optimization) or independence tests. In our work, we instead treat the inference process as a black box and design a neural network architecture that learns the mapping from both observational and interventional data to graph structures via supervised training on synthetic graphs. The learned model generalizes to new synthetic graphs, is robust to train-test distribution shifts, and achieves state-of-the-art performance on naturalistic graphs for low sample complexity.

Frederik Träuble, Anirudh Goyal, Nasim Rahaman et al. · mila

Deep neural networks perform well on classification tasks where data streams are i.i.d. and labeled data is abundant. Challenges emerge with non-stationary training data streams such as continual learning. One powerful approach that has addressed this challenge involves pre-training of large encoders on volumes of readily available data, followed by task-specific tuning. Given a new task, however, updating the weights of these encoders is challenging as a large number of weights needs to be fine-tuned, and as a result, they forget information about the previous tasks. In the present work, we propose a model architecture to address this issue, building upon a discrete bottleneck containing pairs of separate and learnable key-value codes. Our paradigm will be to encode; process the representation via a discrete bottleneck; and decode. Here, the input is fed to the pre-trained encoder, the output of the encoder is used to select the nearest keys, and the corresponding values are fed to the decoder to solve the current task. The model can only fetch and re-use a sparse number of these key-value pairs during inference, enabling localized and context-dependent model updates. We theoretically investigate the ability of the discrete key-value bottleneck to minimize the effect of learning under distribution shifts and show that it reduces the complexity of the hypothesis class. We empirically verify the proposed method under challenging class-incremental learning scenarios and show that the proposed model - without any task boundaries - reduces catastrophic forgetting across a wide variety of pre-trained models, outperforming relevant baselines on this task.

Vedant Shah, Dingli Yu, Kaifeng Lyu et al. · mila, tsinghua

Current LLM training positions mathematical reasoning as a core capability. With publicly available sources fully tapped, there is unmet demand for diverse and challenging math questions. Relying solely on human experts is both time-consuming and costly, while LLM-generated questions often lack the requisite diversity and difficulty. We present a design framework that combines the strengths of LLMs with a human-in-the-loop approach to generate a diverse array of challenging math questions. We leverage LLM metacognition skills [Didolkar et al., 2024] of a strong LLM to extract core "skills" from existing math datasets. These skills serve as the basis for generating novel and difficult questions by prompting the LLM with random pairs of core skills. The use of two different skills within each question makes finding such questions an "out of distribution" task for both LLMs and humans. Our pipeline employs LLMs to iteratively generate and refine questions and solutions through multiturn prompting. Human annotators then verify and further refine the questions, with their efficiency enhanced via further LLM interactions. Applying this pipeline on skills extracted from the MATH dataset [Hendrycks et al., 2021] resulted in MATH$^2$ - a dataset of higher-quality math questions, as evidenced by: (a) Lower performance of all models on MATH$^2$ than on MATH (b) Higher performance on MATH when using MATH$^2$ questions as in-context examples. Although focused on mathematics, our methodology seems applicable to other domains requiring structured reasoning, and potentially as a component of scalable oversight. Also of interest is a striking relationship observed between models' performance on the new dataset: the success rate on MATH$^2$ is the square on MATH, suggesting that successfully solving the question in MATH$^2$ requires a nontrivial combination of two distinct math skills.

Dianbo Liu, Vedant Shah, Oussama Boussif et al. · mila

In cooperative multi-agent reinforcement learning, a team of agents works together to achieve a common goal. Different environments or tasks may require varying degrees of coordination among agents in order to achieve the goal in an optimal way. The nature of coordination will depend on the properties of the environment -- its spatial layout, distribution of obstacles, dynamics, etc. We term this variation of properties within an environment as heterogeneity. Existing literature has not sufficiently addressed the fact that different environments may have different levels of heterogeneity. We formalize the notions of coordination level and heterogeneity level of an environment and present HECOGrid, a suite of multi-agent RL environments that facilitates empirical evaluation of different MARL approaches across different levels of coordination and environmental heterogeneity by providing a quantitative control over coordination and heterogeneity levels of the environment. Further, we propose a Centralized Training Decentralized Execution learning approach called Stateful Active Facilitator (SAF) that enables agents to work efficiently in high-coordination and high-heterogeneity environments through a differentiable and shared knowledge source used during training and dynamic selection from a shared pool of policies. We evaluate SAF and compare its performance against baselines IPPO and MAPPO on HECOGrid. Our results show that SAF consistently outperforms the baselines across different tasks and different heterogeneity and coordination levels. We release the code for HECOGrid as well as all our experiments.

Vedant Shah, Frederik Träuble, Ashish Malik et al. · mila

Machine \emph{unlearning}, which involves erasing knowledge about a \emph{forget set} from a trained model, can prove to be costly and infeasible by existing techniques. We propose a nearly compute-free zero-shot unlearning technique based on a discrete representational bottleneck. We show that the proposed technique efficiently unlearns the forget set and incurs negligible damage to the model's performance on the rest of the data set. We evaluate the proposed technique on the problem of \textit{class unlearning} using three datasets: CIFAR-10, CIFAR-100, and LACUNA-100. We compare the proposed technique to SCRUB, a state-of-the-art approach which uses knowledge distillation for unlearning. Across all three datasets, the proposed technique performs as well as, if not better than SCRUB while incurring almost no computational cost.

Pratyush Maini, Michael C. Mozer, Hanie Sedghi et al.

Recent efforts at explaining the interplay of memorization and generalization in deep overparametrized networks have posited that neural networks $\textit{memorize}$ "hard" examples in the final few layers of the model. Memorization refers to the ability to correctly predict on $\textit{atypical}$ examples of the training set. In this work, we show that rather than being confined to individual layers, memorization is a phenomenon confined to a small set of neurons in various layers of the model. First, via three experimental sources of converging evidence, we find that most layers are redundant for the memorization of examples and the layers that contribute to example memorization are, in general, not the final layers. The three sources are $\textit{gradient accounting}$ (measuring the contribution to the gradient norms from memorized and clean examples), $\textit{layer rewinding}$ (replacing specific model weights of a converged model with previous training checkpoints), and $\textit{retraining}$ (training rewound layers only on clean examples). Second, we ask a more generic question: can memorization be localized $\textit{anywhere}$ in a model? We discover that memorization is often confined to a small number of neurons or channels (around 5) of the model. Based on these insights we propose a new form of dropout -- $\textit{example-tied dropout}$ that enables us to direct the memorization of examples to an apriori determined set of neurons. By dropping out these neurons, we are able to reduce the accuracy on memorized examples from $100\%\to3\%$, while also reducing the generalization gap.

Sumukh Aithal, Anirudh Goyal, Alex Lamb et al. · cmu, mila

Self-Supervised Learning (SSL) methods operate on unlabeled data to learn robust representations useful for downstream tasks. Most SSL methods rely on augmentations obtained by transforming the 2D image pixel map. These augmentations ignore the fact that biological vision takes place in an immersive three-dimensional, temporally contiguous environment, and that low-level biological vision relies heavily on depth cues. Using a signal provided by a pretrained state-of-the-art monocular RGB-to-depth model (the \emph{Depth Prediction Transformer}, Ranftl et al., 2021), we explore two distinct approaches to incorporating depth signals into the SSL framework. First, we evaluate contrastive learning using an RGB+depth input representation. Second, we use the depth signal to generate novel views from slightly different camera positions, thereby producing a 3D augmentation for contrastive learning. We evaluate these two approaches on three different SSL methods -- BYOL, SimSiam, and SwAV -- using ImageNette (10 class subset of ImageNet), ImageNet-100 and ImageNet-1k datasets. We find that both approaches to incorporating depth signals improve the robustness and generalization of the baseline SSL methods, though the first approach (with depth-channel concatenation) is superior. For instance, BYOL with the additional depth channel leads to an increase in downstream classification accuracy from 85.3\% to 88.0\% on ImageNette and 84.1\% to 87.0\% on ImageNet-C.

Frieda Born, Tom Neuhäuser, Lukas Muttenthaler et al. · deepmind, stanford

Modern machine learning models typically represent inputs as fixed points in a high-dimensional embedding space. While this approach has been proven powerful for a wide range of downstream tasks, it fundamentally differs from the way humans process information. Because humans are constantly adapting to their environment, they represent objects and their relationships in a highly context-sensitive manner. To address this gap, we propose a method for context-sensitive similarity computation from neural network embeddings, applied to modeling a triplet odd-one-out task with an anchor image serving as simultaneous context. Modeling context enables us to achieve up to a 15% improvement in odd-one-out accuracy over a context-insensitive model. We find that this improvement is consistent across both original and "human-aligned" vision foundation models.

Lukas Muttenthaler, Klaus Greff, Frieda Born et al. · deepmind, stanford

Deep neural networks have achieved success across a wide range of applications, including as models of human behavior and neural representations in vision tasks. However, neural network training and human learning differ in fundamental ways, and neural networks often fail to generalize as robustly as humans do raising questions regarding the similarity of their underlying representations. What is missing for modern learning systems to exhibit more human-aligned behavior? We highlight a key misalignment between vision models and humans: whereas human conceptual knowledge is hierarchically organized from fine- to coarse-scale distinctions, model representations do not accurately capture all these levels of abstraction. To address this misalignment, we first train a teacher model to imitate human judgments, then transfer human-aligned structure from its representations to refine the representations of pretrained state-of-the-art vision foundation models via finetuning. These human-aligned models more accurately approximate human behavior and uncertainty across a wide range of similarity tasks, including a new dataset of human judgments spanning multiple levels of semantic abstractions. They also perform better on a diverse set of machine learning tasks, increasing generalization and out-of-distribution robustness. Thus, infusing neural networks with additional human knowledge yields a best-of-both-worlds representation that is both more consistent with human cognitive judgments and more practically useful, thus paving the way toward more robust, interpretable, and human-aligned artificial intelligence systems.

Gamaleldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste et al.

The visual world can be parsimoniously characterized in terms of distinct entities with sparse interactions. Discovering this compositional structure in dynamic visual scenes has proven challenging for end-to-end computer vision approaches unless explicit instance-level supervision is provided. Slot-based models leveraging motion cues have recently shown great promise in learning to represent, segment, and track objects without direct supervision, but they still fail to scale to complex real-world multi-object videos. In an effort to bridge this gap, we take inspiration from human development and hypothesize that information about scene geometry in the form of depth signals can facilitate object-centric learning. We introduce SAVi++, an object-centric video model which is trained to predict depth signals from a slot-based video representation. By further leveraging best practices for model scaling, we are able to train SAVi++ to segment complex dynamic scenes recorded with moving cameras, containing both static and moving objects of diverse appearance on naturalistic backgrounds, without the need for segmentation supervision. Finally, we demonstrate that by using sparse depth signals obtained from LiDAR, SAVi++ is able to learn emergent object segmentation and tracking from videos in the real-world Waymo Open dataset.

Dianbo Liu, Vedant Shah, Oussama Boussif et al. · mila

In Multi-Agent Reinforcement Learning (MARL), specialized channels are often introduced that allow agents to communicate directly with one another. In this paper, we propose an alternative approach whereby agents communicate through an intelligent facilitator that learns to sift through and interpret signals provided by all agents to improve the agents' collective performance. To ensure that this facilitator does not become a centralized controller, agents are incentivized to reduce their dependence on the messages it conveys, and the messages can only influence the selection of a policy from a fixed set, not instantaneous actions given the policy. We demonstrate the strength of this architecture over existing baselines on several cooperative MARL environments.

Amr Khalifa, Michael C. Mozer, Hanie Sedghi et al.

Recent works demonstrate that early layers in a neural network contain useful information for prediction. Inspired by this, we show that extending temperature scaling across all layers improves both calibration and accuracy. We call this procedure "layer-stack temperature scaling" (LATES). Informally, LATES grants each layer a weighted vote during inference. We evaluate it on five popular convolutional neural network architectures both in- and out-of-distribution and observe a consistent improvement over temperature scaling in terms of accuracy, calibration, and AUC. All conclusions are supported by comprehensive statistical analyses. Since LATES neither retrains the architecture nor introduces many more parameters, its advantages can be reaped without requiring additional data beyond what is used in temperature scaling. Finally, we show that combining LATES with Monte Carlo Dropout matches state-of-the-art results on CIFAR10/100.

Tyler R. Scott, Ting Liu, Michael C. Mozer et al.

Recent research in clustering face embeddings has found that unsupervised, shallow, heuristic-based methods -- including $k$-means and hierarchical agglomerative clustering -- underperform supervised, deep, inductive methods. While the reported improvements are indeed impressive, experiments are mostly limited to face datasets, where the clustered embeddings are highly discriminative or well-separated by class (Recall@1 above 90% and often nearing ceiling), and the experimental methodology seemingly favors the deep methods. We conduct a large-scale empirical study of 17 clustering methods across three datasets and obtain several robust findings. Notably, deep methods are surprisingly fragile for embeddings with more uncertainty, where they match or even perform worse than shallow, heuristic-based methods. When embeddings are highly discriminative, deep methods do outperform the baselines, consistent with past results, but the margin between methods is much smaller than previously reported. We believe our benchmarks broaden the scope of supervised clustering methods beyond the face domain and can serve as a foundation on which these methods could be improved. To enable reproducibility, we include all necessary details in the appendices, and plan to release the code.

Shruthi Sukumar, Adrian F. Ward, Camden Elliott-Williams et al.

Individuals are often faced with temptations that can lead them astray from long-term goals. We're interested in developing interventions that steer individuals toward making good initial decisions and then maintaining those decisions over time. In the realm of financial decision making, a particularly successful approach is the prize-linked savings account: individuals are incentivized to make deposits by tying deposits to a periodic lottery that awards bonuses to the savers. Although these lotteries have been very effective in motivating savers across the globe, they are a one-size-fits-all solution. We investigate whether customized bonuses can be more effective. We formalize a delayed-gratification task as a Markov decision problem and characterize individuals as rational agents subject to temporal discounting, a cost associated with effort, and fluctuations in willpower. Our theory is able to explain key behavioral findings in intertemporal choice. We created an online delayed-gratification game in which the player scores points by selecting a queue to wait in and then performing a series of actions to advance to the front. Data collected from the game is fit to the model, and the instantiated model is then used to optimize predicted player performance over a space of incentives. We demonstrate that customized incentive structures can improve an individual's goal-directed decision making.

Eleni Triantafillou, Ahmed Imtiaz Humayun, Monica Ribero et al.

As models are getting larger and are trained on increasing amounts of data, there has been an explosion of interest into how we can ``delete'' specific data points or behaviours from a trained model, after the fact. This goal has been referred to as ``machine unlearning''. In this note, we argue that the term ``unlearning'' has been overloaded, with different research efforts spanning two distinct problem formulations, but without that distinction having been observed or acknowledged in the literature. This causes various issues, including ambiguity around when an algorithm is expected to work, use of inappropriate metrics and baselines when comparing different algorithms to one another, difficulty in interpreting results, as well as missed opportunities for pursuing critical research directions. In this note, we address this issue by establishing a fundamental distinction between two notions that we identify as \unlearning and \untraining, illustrated in Figure 1. In short, \untraining aims to reverse the effect of having trained on a given forget set, i.e. to remove the influence that that specific forget set examples had on the model during training. On the other hand, the goal of \unlearning is not just to remove the influence of those given examples, but to use those examples for the purpose of more broadly removing the entire underlying distribution from which those examples were sampled (e.g. the concept or behaviour that those examples represent). We discuss technical definitions of these problems and map problem settings studied in the literature to each. We hope to initiate discussions on disambiguating technical definitions and identify a set of overlooked research questions, as we believe that this a key missing step for accelerating progress in the field of ``unlearning''.

Bernd Bohnet, Michael C. Mozer, Kevin Swersky et al.

Classic AI planning problems have been revisited in the Large Language Model (LLM) era, with a focus of recent benchmarks on success rates rather than plan efficiency. We examine the degree to which frontier models reason optimally versus relying on simple, heuristic, and possibly inefficient strategies. We focus on the Blocksworld domain involving towers of labeled blocks which have to be moved from an initial to a goal configuration via a set of primitive actions. We also study a formally equivalent task, the generalized Path-Star ($P^*$) graph, in order to isolate true topological reasoning from semantic priors. We systematically manipulate problem depth (the height of block towers), width (the number of towers), and compositionality (the number of goal blocks). Reasoning-enhanced LLMs significantly outperform traditional satisficing planners (e.g., LAMA) in complex, multi-goal configurations. Although classical search algorithms hit a wall as the search space expands, LLMs track theoretical optimality limits with near-perfect precision, even when domain-specific semantic hints are stripped away. To explain these surprising findings, we consider (and find evidence to support) two hypotheses: an active Algorithmic Simulation executed via reasoning tokens and a Geometric Memory that allows models to represent the $P^*$ topology as a navigable global geometry, effectively bypassing exponential combinatorial complexity.

Michael C. Mozer, Shoaib Ahmed Siddiqui, Rosanne Liu

Transformers encode structure in sequences via an expanding contextual history. However, their purely feedforward architecture fundamentally limits dynamic state tracking. State tracking -- the iterative updating of latent variables reflecting an evolving environment -- involves inherently sequential dependencies that feedforward networks struggle to maintain. Consequently, feedforward models push evolving state representations deeper into their layer stack with each new input step, rendering information inaccessible in shallow layers and ultimately exhausting the model's depth. While this depth limit can be bypassed by dynamic depth models and by explicit or latent thinking that externalizes state representations, these solutions are computationally and memory inefficient. In this article, we argue that temporally extended cognition requires refocusing from explicit thought traces to implicit activation dynamics via recurrent architectures. We introduce a taxonomy of recurrent and continuous-thought transformer architectures, categorizing them by their recurrence axis (depth versus step) and their ratio of input tokens to recurrence steps. Finally, we outline promising research directions, including enhanced state-space models and coarse-grained recurrence, to better integrate state tracking into modern foundation models.

Katherine L. Hermann, Hossein Mobahi, Thomas Fel et al.

Deep-learning models can extract a rich assortment of features from data. Which features a model uses depends not only on \emph{predictivity} -- how reliably a feature indicates training-set labels -- but also on \emph{availability} -- how easily the feature can be extracted from inputs. The literature on shortcut learning has noted examples in which models privilege one feature over another, for example texture over shape and image backgrounds over foreground objects. Here, we test hypotheses about which input properties are more available to a model, and systematically study how predictivity and availability interact to shape models' feature use. We construct a minimal, explicit generative framework for synthesizing classification datasets with two latent features that vary in predictivity and in factors we hypothesize to relate to availability, and we quantify a model's shortcut bias -- its over-reliance on the shortcut (more available, less predictive) feature at the expense of the core (less available, more predictive) feature. We find that linear models are relatively unbiased, but introducing a single hidden layer with ReLU or Tanh units yields a bias. Our empirical findings are consistent with a theoretical account based on Neural Tangent Kernels. Finally, we study how models used in practice trade off predictivity and availability in naturalistic datasets, discovering availability manipulations which increase models' degree of shortcut bias. Taken together, these findings suggest that the propensity to learn shortcut features is a fundamental characteristic of deep nonlinear architectures warranting systematic study given its role in shaping how models solve tasks.

Aniket Didolkar, Anirudh Goyal, Nan Rosemary Ke et al. · mila

Metacognitive knowledge refers to humans' intuitive knowledge of their own thinking and reasoning processes. Today's best LLMs clearly possess some reasoning processes. The paper gives evidence that they also have metacognitive knowledge, including ability to name skills and procedures to apply given a task. We explore this primarily in context of math reasoning, developing a prompt-guided interaction procedure to get a powerful LLM to assign sensible skill labels to math questions, followed by having it perform semantic clustering to obtain coarser families of skill labels. These coarse skill labels look interpretable to humans. To validate that these skill labels are meaningful and relevant to the LLM's reasoning processes we perform the following experiments. (a) We ask GPT-4 to assign skill labels to training questions in math datasets GSM8K and MATH. (b) When using an LLM to solve the test questions, we present it with the full list of skill labels and ask it to identify the skill needed. Then it is presented with randomly selected exemplar solved questions associated with that skill label. This improves accuracy on GSM8k and MATH for several strong LLMs, including code-assisted models. The methodology presented is domain-agnostic, even though this article applies it to math problems.

Haonan Wang, James Zou, Michael Mozer et al.

Creativity serves as a cornerstone for societal progress and innovation. With the rise of advanced generative AI models capable of tasks once reserved for human creativity, the study of AI's creative potential becomes imperative for its responsible development and application. In this paper, we prove in theory that AI can be as creative as humans under the condition that it can properly fit the data generated by human creators. Therefore, the debate on AI's creativity is reduced into the question of its ability to fit a sufficient amount of data. To arrive at this conclusion, this paper first addresses the complexities in defining creativity by introducing a new concept called Relative Creativity. Rather than attempting to define creativity universally, we shift the focus to whether AI can match the creative abilities of a hypothetical human. The methodological shift leads to a statistically quantifiable assessment of AI's creativity, term Statistical Creativity. This concept, statistically comparing the creative abilities of AI with those of specific human groups, facilitates theoretical exploration of AI's creative potential. Our analysis reveals that by fitting extensive conditional data without marginalizing out the generative conditions, AI can emerge as a hypothetical new creator. The creator possesses the same creative abilities on par with the human creators it was trained on. Building on theoretical findings, we discuss the application in prompt-conditioned autoregressive models, providing a practical means for evaluating creative abilities of generative AI models, such as Large Language Models (LLMs). Additionally, this study provides an actionable training guideline, bridging the theoretical quantification of creativity with practical model training.

Rushi Shah, Mingyuan Yan, Michael Curtis Mozer et al.

Discrete representations play a crucial role in many deep learning architectures, yet their non-differentiable nature poses significant challenges for gradient-based optimization. To address this issue, various gradient estimators have been developed, including the Straight-Through Gumbel-Softmax (ST-GS) estimator, which combines the Straight-Through Estimator (STE) and the Gumbel-based reparameterization trick. However, the performance of ST-GS is highly sensitive to temperature, with its selection often compromising gradient fidelity. In this work, we propose a simple yet effective extension to ST-GS by employing decoupled temperatures for forward and backward passes, which we refer to as "Decoupled ST-GS". We show that our approach significantly enhances the original ST-GS through extensive experiments across multiple tasks and datasets. We further investigate the impact of our method on gradient fidelity from multiple perspectives, including the gradient gap and the bias-variance trade-off of estimated gradients. Our findings contribute to the ongoing effort to improve discrete optimization in deep learning, offering a practical solution that balances simplicity and effectiveness.

Danny P. Sawyer, Nan Rosemary Ke, Hubert Soyer et al. · deepmind

Foundation models excel at single-turn reasoning but struggle with multi-turn exploration in dynamic environments, a requirement for many real-world challenges. We evaluated these models on their ability to learn from experience, adapt, and gather information. First, in "Feature World," a simple setting for testing information gathering, models performed near-optimally. However, to test more complex, multi-trial learning, we implemented a text-based version of the "Alchemy" environment, a benchmark for meta-learning. Here, agents must deduce a latent causal structure by integrating information across many trials. In this setting, recent foundation models initially failed to improve their performance over time. Crucially, we found that prompting the models to summarize their observations at regular intervals enabled an emergent meta-learning process. This allowed them to improve across trials and even adaptively re-learn when the environment's rules changed unexpectedly. While most models handled the simple task, Alchemy revealed stark differences in robustness: Gemini 2.5 performed best, followed by Claude 3.7, while ChatGPT-4o and o4-mini struggled. This underscores Alchemy's value as a benchmark. Our findings demonstrate that the biggest challenge for foundation models is not selecting informative actions in the moment, but integrating knowledge through adaptive strategies over time. Encouragingly, there appears to be no intrinsic barrier to future models mastering these abilities.

Shoaib Ahmed Siddiqui, Adrian Weller, David Krueger et al.

Recent unlearning methods for LLMs are vulnerable to relearning attacks: knowledge believed-to-be-unlearned re-emerges by fine-tuning on a small set of (even seemingly-unrelated) examples. We study this phenomenon in a controlled setting for example-level unlearning in vision classifiers. We make the surprising discovery that forget-set accuracy can recover from around 50% post-unlearning to nearly 100% with fine-tuning on just the retain set -- i.e., zero examples of the forget set. We observe this effect across a wide variety of unlearning methods, whereas for a model retrained from scratch excluding the forget set (gold standard), the accuracy remains at 50%. We observe that resistance to relearning attacks can be predicted by weight-space properties, specifically, $L_2$-distance and linear mode connectivity between the original and the unlearned model. Leveraging this insight, we propose a new class of methods that achieve state-of-the-art resistance to relearning attacks.

Anand Gopalakrishnan, Robert Csordás, Jürgen Schmidhuber et al.

The attention mechanism in a Transformer architecture matches key to query based on both content -- the what -- and position in a sequence -- the where. We present an analysis indicating that what and where are entangled in the popular RoPE rotary position embedding. This entanglement can impair performance particularly when decisions require independent matches on these two factors. We propose an improvement to RoPE, which we call Polar Coordinate Position Embeddings or PoPE, that eliminates the what-where confound. PoPE is far superior on a diagnostic task requiring indexing solely by position or by content. On autoregressive sequence modeling in music, genomic, and natural language domains, Transformers using PoPE as the positional encoding scheme outperform baselines using RoPE with respect to evaluation loss (perplexity) and downstream task performance. On language modeling, these gains persist across model scale, from 124M to 774M parameters. Crucially, PoPE shows strong zero-shot length extrapolation capabilities, whereas RoPE's performance degrades significantly on longer sequences at test time without fine tuning or the use of position-interpolation methods.

Yanlai Yang, Matt Jones, Michael C. Mozer et al.

We explore the training dynamics of neural networks in a structured non-IID setting where documents are presented cyclically in a fixed, repeated sequence. Typically, networks suffer from catastrophic interference when training on a sequence of documents; however, we discover a curious and remarkable property of LLMs finetuned sequentially in this setting: they exhibit anticipatory behavior, recovering from the forgetting on documents before encountering them again. This behavior occurs even though the documents are never presented in context together. The behavior emerges and becomes more robust as the architecture scales up its number of parameters. Through comprehensive experiments and visualizations, we demonstrate a new mechanism by which over-parametrized neural networks can recover from catastrophic interference and uncover new insights into training over-parameterized networks in cyclically structured environments.

Ayush Chakravarthy, Trang Nguyen, Anirudh Goyal et al.

The aim of object-centric vision is to construct an explicit representation of the objects in a scene. This representation is obtained via a set of interchangeable modules called \emph{slots} or \emph{object files} that compete for local patches of an image. The competition has a weak inductive bias to preserve spatial continuity; consequently, one slot may claim patches scattered diffusely throughout the image. In contrast, the inductive bias of human vision is strong, to the degree that attention has classically been described with a spotlight metaphor. We incorporate a spatial-locality prior into state-of-the-art object-centric vision models and obtain significant improvements in segmenting objects in both synthetic and real-world datasets. Similar to human visual attention, the combination of image content and spatial constraints yield robust unsupervised object-centric learning, including less sensitivity to model hyperparameters.

Utku Evci, Vincent Dumoulin, Hugo Larochelle et al.

Transfer-learning methods aim to improve performance in a data-scarce target domain using a model pretrained on a data-rich source domain. A cost-efficient strategy, linear probing, involves freezing the source model and training a new classification head for the target domain. This strategy is outperformed by a more costly but state-of-the-art method -- fine-tuning all parameters of the source model to the target domain -- possibly because fine-tuning allows the model to leverage useful information from intermediate layers which is otherwise discarded by the later pretrained layers. We explore the hypothesis that these intermediate layers might be directly exploited. We propose a method, Head-to-Toe probing (Head2Toe), that selects features from all layers of the source model to train a classification head for the target-domain. In evaluations on the VTAB-1k, Head2Toe matches performance obtained with fine-tuning on average while reducing training and storage cost hundred folds or more, but critically, for out-of-distribution transfer, Head2Toe outperforms fine-tuning.

Mengye Ren, Tyler R. Scott, Michael L. Iuzzolino et al.

Real world learning scenarios involve a nonstationary distribution of classes with sequential dependencies among the samples, in contrast to the standard machine learning formulation of drawing samples independently from a fixed, typically uniform distribution. Furthermore, real world interactions demand learning on-the-fly from few or no class labels. In this work, we propose an unsupervised model that simultaneously performs online visual representation learning and few-shot learning of new categories without relying on any class labels. Our model is a prototype-based memory network with a control component that determines when to form a new class prototype. We formulate it as an online mixture model, where components are created with only a single new example, and assignments do not have to be balanced, which permits an approximation to natural imbalanced distributions from uncurated raw data. Learning includes a contrastive loss that encourages different views of the same image to be assigned to the same prototype. The result is a mechanism that forms categorical representations of objects in nonstationary environments. Experiments show that our method can learn from an online stream of visual input data and its learned representations are significantly better at category recognition compared to state-of-the-art self-supervised learning methods.

Nino Scherrer, Olexa Bilaniuk, Yashas Annadani et al.

Discovering causal structures from data is a challenging inference problem of fundamental importance in all areas of science. The appealing properties of neural networks have recently led to a surge of interest in differentiable neural network-based methods for learning causal structures from data. So far, differentiable causal discovery has focused on static datasets of observational or fixed interventional origin. In this work, we introduce an active intervention targeting (AIT) method which enables a quick identification of the underlying causal structure of the data-generating process. Our method significantly reduces the required number of interactions compared with random intervention targeting and is applicable for both discrete and continuous optimization formulations of learning the underlying directed acyclic graph (DAG) from data. We examine the proposed method across multiple frameworks in a wide range of settings and demonstrate superior performance on multiple benchmarks from simulated to real-world data.

Archit Karandikar, Nicholas Cain, Dustin Tran et al.

Optimal decision making requires that classifiers produce uncertainty estimates consistent with their empirical accuracy. However, deep neural networks are often under- or over-confident in their predictions. Consequently, methods have been developed to improve the calibration of their predictive uncertainty both during training and post-hoc. In this work, we propose differentiable losses to improve calibration based on a soft (continuous) version of the binning operation underlying popular calibration-error estimators. When incorporated into training, these soft calibration losses achieve state-of-the-art single-model ECE across multiple datasets with less than 1% decrease in accuracy. For instance, we observe an 82% reduction in ECE (70% relative to the post-hoc rescaled ECE) in exchange for a 0.7% relative decrease in accuracy relative to the cross entropy baseline on CIFAR-100. When incorporated post-training, the soft-binning-based calibration error objective improves upon temperature scaling, a popular recalibration method. Overall, experiments across losses and datasets demonstrate that using calibration-sensitive procedures yield better uncertainty estimates under dataset shift than the standard practice of using a cross entropy loss and post-hoc recalibration methods.

Dianbo Liu, Alex Lamb, Kenji Kawaguchi et al.

Deep learning has advanced from fully connected architectures to structured models organized into components, e.g., the transformer composed of positional elements, modular architectures divided into slots, and graph neural nets made up of nodes. In structured models, an interesting question is how to conduct dynamic and possibly sparse communication among the separate components. Here, we explore the hypothesis that restricting the transmitted information among components to discrete representations is a beneficial bottleneck. The motivating intuition is human language in which communication occurs through discrete symbols. Even though individuals have different understandings of what a "cat" is based on their specific experiences, the shared discrete token makes it possible for communication among individuals to be unimpeded by individual differences in internal representation. To discretize the values of concepts dynamically communicated among specialist components, we extend the quantization mechanism from the Vector-Quantized Variational Autoencoder to multi-headed discretization with shared codebooks and use it for discrete-valued neural communication (DVNC). Our experiments show that DVNC substantially improves systematic generalization in a variety of architectures -- transformers, modular architectures, and graph neural networks. We also show that the DVNC is robust to the choice of hyperparameters, making the method very useful in practice. Moreover, we establish a theoretical justification of our discretization process, proving that it has the ability to increase noise robustness and reduce the underlying dimensionality of the model.

Nan Rosemary Ke, Aniket Didolkar, Sarthak Mittal et al.

Inducing causal relationships from observations is a classic problem in machine learning. Most work in causality starts from the premise that the causal variables themselves are observed. However, for AI agents such as robots trying to make sense of their environment, the only observables are low-level variables like pixels in images. To generalize well, an agent must induce high-level variables, particularly those which are causal or are affected by causal variables. A central goal for AI and causality is thus the joint discovery of abstract representations and causal structure. However, we note that existing environments for studying causal induction are poorly suited for this objective because they have complicated task-specific causal graphs which are impossible to manipulate parametrically (e.g., number of nodes, sparsity, causal chain length, etc.). In this work, our goal is to facilitate research in learning representations of high-level variables as well as causal structures among them. In order to systematically probe the ability of methods to identify these variables and structures, we design a suite of benchmarking RL environments. We evaluate various representation learning algorithms from the literature and find that explicitly incorporating structure and modularity in models can help causal induction in model-based reinforcement learning.

Tyler R. Scott, Andrew C. Gallagher, Michael C. Mozer

Recent work has argued that classification losses utilizing softmax cross-entropy are superior not only for fixed-set classification tasks, but also by outperforming losses developed specifically for open-set tasks including few-shot learning and retrieval. Softmax classifiers have been studied using different embedding geometries -- Euclidean, hyperbolic, and spherical -- and claims have been made about the superiority of one or another, but they have not been systematically compared with careful controls. We conduct an empirical investigation of embedding geometry on softmax losses for a variety of fixed-set classification and image retrieval tasks. An interesting property observed for the spherical losses lead us to propose a probabilistic classifier based on the von Mises-Fisher distribution, and we show that it is competitive with state-of-the-art methods while producing improved out-of-the-box calibration. We provide guidance regarding the trade-offs between losses and how to choose among them.

Piotr Teterwak, Chiyuan Zhang, Dilip Krishnan et al.

A discriminatively trained neural net classifier can fit the training data perfectly if all information about its input other than class membership has been discarded prior to the output layer. Surprisingly, past research has discovered that some extraneous visual detail remains in the logit vector. This finding is based on inversion techniques that map deep embeddings back to images. We explore this phenomenon further using a novel synthesis of methods, yielding a feedforward inversion model that produces remarkably high fidelity reconstructions, qualitatively superior to those of past efforts. When applied to an adversarially robust classifier model, the reconstructions contain sufficient local detail and global structure that they might be confused with the original image in a quick glance, and the object category can clearly be gleaned from the reconstruction. Our approach is based on BigGAN (Brock, 2019), with conditioning on logits instead of one-hot class labels. We use our reconstruction model as a tool for exploring the nature of representations, including: the influence of model architecture and training objectives (specifically robust losses), the forms of invariance that networks achieve, representational differences between correctly and incorrectly classified images, and the effects of manipulating logits and images. We believe that our method can inspire future investigations into the nature of information flow in a neural net and can provide diagnostics for improving discriminative models.

Anirudh Goyal, Aniket Didolkar, Nan Rosemary Ke et al.

Visual environments are structured, consisting of distinct objects or entities. These entities have properties -- both visible and latent -- that determine the manner in which they interact with one another. To partition images into entities, deep-learning researchers have proposed structural inductive biases such as slot-based architectures. To model interactions among entities, equivariant graph neural nets (GNNs) are used, but these are not particularly well suited to the task for two reasons. First, GNNs do not predispose interactions to be sparse, as relationships among independent entities are likely to be. Second, GNNs do not factorize knowledge about interactions in an entity-conditional manner. As an alternative, we take inspiration from cognitive science and resurrect a classic approach, production systems, which consist of a set of rule templates that are applied by binding placeholder variables in the rules to specific entities. Rules are scored on their match to entities, and the best fitting rules are applied to update entity properties. In a series of experiments, we demonstrate that this architecture achieves a flexible, dynamic flow of control and serves to factorize entity-specific and rule-based information. This disentangling of knowledge achieves robust future-state prediction in rich visual environments, outperforming state-of-the-art methods using GNNs, and allows for the extrapolation from simple (few object) environments to more complex environments.

Anirudh Goyal, Aniket Didolkar, Alex Lamb et al.

Deep learning has seen a movement away from representing examples with a monolithic hidden state towards a richly structured state. For example, Transformers segment by position, and object-centric architectures decompose images into entities. In all these architectures, interactions between different elements are modeled via pairwise interactions: Transformers make use of self-attention to incorporate information from other positions; object-centric architectures make use of graph neural networks to model interactions among entities. However, pairwise interactions may not achieve global coordination or a coherent, integrated representation that can be used for downstream tasks. In cognitive science, a global workspace architecture has been proposed in which functionally specialized components share information through a common, bandwidth-limited communication channel. We explore the use of such a communication channel in the context of deep learning for modeling the structure of complex environments. The proposed method includes a shared workspace through which communication among different specialist modules takes place but due to limits on the communication bandwidth, specialist modules must compete for access. We show that capacity limitations have a rational basis in that (1) they encourage specialization and compositionality and (2) they facilitate the synchronization of otherwise independent specialists.

Michael L. Iuzzolino, Michael C. Mozer, Samy Bengio

Although deep feedforward neural networks share some characteristics with the primate visual system, a key distinction is their dynamics. Deep nets typically operate in serial stages wherein each layer completes its computation before processing begins in subsequent layers. In contrast, biological systems have cascaded dynamics: information propagates from neurons at all layers in parallel but transmission occurs gradually over time, leading to speed-accuracy trade offs even in feedforward architectures. We explore the consequences of biologically inspired parallel hardware by constructing cascaded ResNets in which each residual block has propagation delays but all blocks update in parallel in a stateful manner. Because information transmitted through skip connections avoids delays, the functional depth of the architecture increases over time, yielding anytime predictions that improve with internal-processing time. We introduce a temporal-difference training loss that achieves a strictly superior speed-accuracy profile over standard losses and enables the cascaded architecture to outperform state-of-the-art anytime-prediction methods. The cascaded architecture has intriguing properties, including: it classifies typical instances more rapidly than atypical instances; it is more robust to both persistent and transient noise than is a conventional ResNet; and its time-varying output trace provides a signal that can be exploited to improve information processing and inference.

Rebecca Roelofs, Nicholas Cain, Jonathon Shlens et al.

For an AI system to be reliable, the confidence it expresses in its decisions must match its accuracy. To assess the degree of match, examples are typically binned by confidence and the per-bin mean confidence and accuracy are compared. Most research in calibration focuses on techniques to reduce this empirical measure of calibration error, ECE_bin. We instead focus on assessing statistical bias in this empirical measure, and we identify better estimators. We propose a framework through which we can compute the bias of a particular estimator for an evaluation data set of a given size. The framework involves synthesizing model outputs that have the same statistics as common neural architectures on popular data sets. We find that binning-based estimators with bins of equal mass (number of instances) have lower bias than estimators with bins of equal width. Our results indicate two reliable calibration-error estimators: the debiased estimator (Brocker, 2012; Ferro and Fricker, 2012) and a method we propose, ECE_sweep, which uses equal-mass bins and chooses the number of bins to be as large as possible while preserving monotonicity in the calibration function. With these estimators, we observe improvements in the effectiveness of recalibration methods and in the detection of model miscalibration.

Alex Lamb, Anirudh Goyal, Agnieszka Słowik et al.

Feed-forward neural networks consist of a sequence of layers, in which each layer performs some processing on the information from the previous layer. A downside to this approach is that each layer (or module, as multiple modules can operate in parallel) is tasked with processing the entire hidden state, rather than a particular part of the state which is most relevant for that module. Methods which only operate on a small number of input variables are an essential part of most programming languages, and they allow for improved modularity and code re-usability. Our proposed method, Neural Function Modules (NFM), aims to introduce the same structural capability into deep learning. Most of the work in the context of feed-forward networks combining top-down and bottom-up feedback is limited to classification problems. The key contribution of our work is to combine attention, sparsity, top-down and bottom-up feedback, in a flexible algorithm which, as we show, improves the results in standard classification, out-of-domain generalization, generative modeling, and learning representations in the context of reinforcement learning.

Maria Attarian, Brett D. Roads, Michael C. Mozer

Deep-learning vision models have shown intriguing similarities and differences with respect to human vision. We investigate how to bring machine visual representations into better alignment with human representations. Human representations are often inferred from behavioral evidence such as the selection of an image most similar to a query image. We find that with appropriate linear transformations of deep embeddings, we can improve prediction of human binary choice on a data set of bird images from 72% at baseline to 89%. We hypothesized that deep embeddings have redundant, high (4096) dimensional representations; however, reducing the rank of these representations results in a loss of explanatory power. We hypothesized that the dilation transformation of representations explored in past research is too restrictive, and indeed we found that model explanatory power can be significantly improved with a more expressive linear transform. Most surprising and exciting, we found that, consistent with classic psychological literature, human similarity judgments are asymmetric: the similarity of X to Y is not necessarily equal to the similarity of Y to X, and allowing models to express this asymmetry improves explanatory power.

Mengye Ren, Michael L. Iuzzolino, Michael C. Mozer et al.

We aim to bridge the gap between typical human and machine-learning environments by extending the standard framework of few-shot learning to an online, continual setting. In this setting, episodes do not have separate training and testing phases, and instead models are evaluated online while learning novel classes. As in the real world, where the presence of spatiotemporal context helps us retrieve learned skills in the past, our online few-shot learning setting also features an underlying context that changes throughout time. Object classes are correlated within a context and inferring the correct context can lead to better performance. Building upon this setting, we propose a new few-shot learning dataset based on large scale indoor imagery that mimics the visual experience of an agent wandering within a world. Furthermore, we convert popular few-shot learning approaches into online versions and we also propose a new contextual prototypical memory model that can make use of spatiotemporal contextual information from the recent past.

Sarthak Mittal, Alex Lamb, Anirudh Goyal et al.

Robust perception relies on both bottom-up and top-down signals. Bottom-up signals consist of what's directly observed through sensation. Top-down signals consist of beliefs and expectations based on past experience and short-term memory, such as how the phrase `peanut butter and~...' will be completed. The optimal combination of bottom-up and top-down information remains an open question, but the manner of combination must be dynamic and both context and task dependent. To effectively utilize the wealth of potential top-down information available, and to prevent the cacophony of intermixed signals in a bidirectional architecture, mechanisms are needed to restrict information flow. We explore deep recurrent neural net architectures in which bottom-up and top-down signals are dynamically combined using attention. Modularity of the architecture further restricts the sharing and communication of information. Together, attention and modularity direct information flow, which leads to reliable performance improvements in perceptual and language tasks, and in particular improves robustness to distractions and noisy data. We demonstrate on a variety of benchmarks in language modeling, sequential image classification, video prediction and reinforcement learning that the \emph{bidirectional} information flow can improve results over strong baselines.

Anirudh Goyal, Alex Lamb, Phanideep Gampa et al.

Modeling a structured, dynamic environment like a video game requires keeping track of the objects and their states declarative knowledge) as well as predicting how objects behave (procedural knowledge). Black-box models with a monolithic hidden state often fail to apply procedural knowledge consistently and uniformly, i.e., they lack systematicity. For example, in a video game, correct prediction of one enemy's trajectory does not ensure correct prediction of another's. We address this issue via an architecture that factorizes declarative and procedural knowledge and that imposes modularity within each form of knowledge. The architecture consists of active modules called object files that maintain the state of a single object and invoke passive external knowledge sources called schemata that prescribe state updates. To use a video game as an illustration, two enemies of the same type will share schemata but will have separate object files to encode their distinct state (e.g., health, position). We propose to use attention to determine which object files to update, the selection of schemata, and the propagation of information between object files. The resulting architecture is a drop-in replacement conforming to the same input-output interface as normal recurrent networks (e.g., LSTM, GRU) yet achieves substantially better generalization on environments that have multiple object tokens of the same type, including a challenging intuitive physics benchmark.

Zeqian Li, Michael C. Mozer, Jacob Whitehill

We present a compositional embedding framework that infers not just a single class per input image, but a set of classes, in the setting of one-shot learning. Specifically, we propose and evaluate several novel models consisting of (1) an embedding function f trained jointly with a "composition" function g that computes set union operations between the classes encoded in two embedding vectors; and (2) embedding f trained jointly with a "query" function h that computes whether the classes encoded in one embedding subsume the classes encoded in another embedding. In contrast to prior work, these models must both perceive the classes associated with the input examples and encode the relationships between different class label sets, and they are trained using only weak one-shot supervision consisting of the label-set relationships among training examples. Experiments on the OmniGlot, Open Images, and COCO datasets show that the proposed compositional embedding models outperform existing embedding methods. Our compositional embedding models have applications to multi-label object recognition for both one-shot and supervised learning.

Ziheng Jiang, Chiyuan Zhang, Kunal Talwar et al.

Humans are accustomed to environments that contain both regularities and exceptions. For example, at most gas stations, one pays prior to pumping, but the occasional rural station does not accept payment in advance. Likewise, deep neural networks can generalize across instances that share common patterns or structures, yet have the capacity to memorize rare or irregular forms. We analyze how individual instances are treated by a model via a consistency score. The score characterizes the expected accuracy for a held-out instance given training sets of varying size sampled from the data distribution. We obtain empirical estimates of this score for individual instances in multiple data sets, and we show that the score identifies out-of-distribution and mislabeled examples at one end of the continuum and strongly regular examples at the other end. We identify computationally inexpensive proxies to the consistency score using statistics collected during training. We show examples of potential applications to the analysis of deep-learning systems.

Tyler R. Scott, Karl Ridgeway, Michael C. Mozer

Supervised deep-embedding methods project inputs of a domain to a representational space in which same-class instances lie near one another and different-class instances lie far apart. We propose a probabilistic method that treats embeddings as random variables. Extending a state-of-the-art deterministic method, Prototypical Networks (Snell et al., 2017), our approach supposes the existence of a class prototype around which class instances are Gaussian distributed. The prototype posterior is a product distribution over labeled instances, and query instances are classified by marginalizing relative prototype proximity over embedding uncertainty. We describe an efficient sampler for approximate inference that allows us to train the model at roughly the same space and time cost as its deterministic sibling. Incorporating uncertainty improves performance on few-shot learning and gracefully handles label noise and out-of-distribution inputs. Compared to the state-of-the-art stochastic method, Hedged Instance Embeddings (Oh et al., 2019), we achieve superior large- and open-set classification accuracy. Our method also aligns class-discriminating features with the axes of the embedding space, yielding an interpretable, disentangled representation.

Michael Iuzzolino, Yoram Singer, Michael C. Mozer

In human perception and cognition, a fundamental operation that brains perform is interpretation: constructing coherent neural states from noisy, incomplete, and intrinsically ambiguous evidence. The problem of interpretation is well matched to an early and often overlooked architecture, the attractor network---a recurrent neural net that performs constraint satisfaction, imputation of missing features, and clean up of noisy data via energy minimization dynamics. We revisit attractor nets in light of modern deep learning methods and propose a convolutional bipartite architecture with a novel training loss, activation function, and connectivity constraints. We tackle larger problems than have been previously explored with attractor nets and demonstrate their potential for image completion and super-resolution. We argue that this architecture is better motivated than ever-deeper feedforward models and is a viable alternative to more costly sampling-based generative methods on a range of supervised and unsupervised tasks.

Guy Davidson, Michael C. Mozer

We explore the behavior of a standard convolutional neural net in a continual-learning setting that introduces visual classification tasks sequentially and requires the net to master new tasks while preserving mastery of previously learned tasks. This setting corresponds to that which human learners face as they acquire domain expertise serially, for example, as an individual studies a textbook. Through simulations involving sequences of ten related visual tasks, we find reason for optimism that nets will scale well as they advance from having a single skill to becoming multi-skill domain experts. We observe two key phenomena. First, \emph{forward facilitation}---the accelerated learning of task $n+1$ having learned $n$ previous tasks---grows with $n$. Second, \emph{backward interference}---the forgetting of the $n$ previous tasks when learning task $n+1$---diminishes with $n$. Amplifying forward facilitation is the goal of research on metalearning, and attenuating backward interference is the goal of research on catastrophic forgetting. We find that both of these goals are attained simply through broader exposure to a domain.

Alex Lamb, Jonathan Binas, Anirudh Goyal et al.

Machine learning promises methods that generalize well from finite labeled data. However, the brittleness of existing neural net approaches is revealed by notable failures, such as the existence of adversarial examples that are misclassified despite being nearly identical to a training example, or the inability of recurrent sequence-processing nets to stay on track without teacher forcing. We introduce a method, which we refer to as \emph{state reification}, that involves modeling the distribution of hidden states over the training data and then projecting hidden states observed during testing toward this distribution. Our intuition is that if the network can remain in a familiar manifold of hidden space, subsequent layers of the net should be well trained to respond appropriately. We show that this state-reification method helps neural nets to generalize better, especially when labeled data are sparse, and also helps overcome the challenge of achieving robust generalization with adversarial training.

Been Kim, Emily Reif, Martin Wattenberg et al.

The Gestalt laws of perceptual organization, which describe how visual elements in an image are grouped and interpreted, have traditionally been thought of as innate despite their ecological validity. We use deep-learning methods to investigate whether natural scene statistics might be sufficient to derive the Gestalt laws. We examine the law of closure, which asserts that human visual perception tends to "close the gap" by assembling elements that can jointly be interpreted as a complete figure or object. We demonstrate that a state-of-the-art convolutional neural network, trained to classify natural images, exhibits closure on synthetic displays of edge fragments, as assessed by similarity of internal representations. This finding provides support for the hypothesis that the human perceptual system is even more elegant than the Gestaltists imagined: a single law---adaptation to the statistical structure of the environment---might suffice as fundamental.