Large language models (LLMs) are increasingly used to evolve solutions to problems. However, most LLM-evolution frameworks solve static optimization problems, overlooking the open-ended adversarial dynamics that characterize real-world evolutionary processes. Here, we study Digital Red Queen (DRQ), a simple self-play algorithm that embraces these "Red Queen" dynamics via a changing objective. DRQ uses an LLM to evolve assembly programs (warriors) which compete for control of a virtual machine in the game of Core War, a Turing-complete environment studied in artificial life and connected to cybersecurity. In each round of DRQ, the model evolves a new warrior to defeat all previous ones. Over many rounds, warriors become increasingly general (relative to a set of held-out human warriors). Interestingly, across independent runs, we observe a convergence pressure toward a single generalpurpose behavioral strategy, much like convergent evolution in nature. Our work positions Core War as a rich, controllable sandbox for studying adversarial adaptation in artificial systems and for evaluating LLM-based evolution methods. More broadly, the simplicity and effectiveness of DRQ suggest that similarly minimal self-play approaches could prove useful in practical multi-agent adversarial domains, like real-world cybersecurity or combating drug resistance.

Measuring alignment between language and vision is a fundamental challenge, especially as multimodal data becomes increasingly detailed and complex. Existing methods often rely on collecting human or AI preferences, which can be costly and time-intensive. We propose an alternative approach that leverages cycle consistency as a supervisory signal. Given an image and generated text, we map the text back to image space using a text-to-image model and compute the similarity between the original image and its reconstruction. Analogously, for text-to-image generation, we measure the textual similarity between an input caption and its reconstruction through the cycle. We use the cycle consistency score to rank candidates and construct a preference dataset of 866K comparison pairs. The reward model trained on our dataset, CycleReward, outperforms state-of-the-art alignment metrics on detailed captioning, with superior inference-time scalability when used as a verifier for Best-of-N sampling, while maintaining speed and differentiability. Furthermore, performing DPO and Diffusion DPO using our dataset enhances performance across a wide range of vision-language tasks and text-to-image generation. Our dataset, model, and code are publicly released at https://cyclereward.github.io.

Current vision systems typically assign fixed-length representations to images, regardless of the information content. This contrasts with human intelligence —and even large language models—which allocate varying representational capacities based on entropy, context and familiarity. Inspired by this, we propose an approach to learn variable-length token representations for 2D images. Our encoder-decoder architecture recursively processes 2D image tokens, distilling them into 1D latent tokens over multiple iterations of recurrent rollouts. Each iteration refines the 2D tokens, updates the existing 1D latent tokens, and adaptively increases representational capacity by adding new tokens. This enables compression of images into a variable number of tokens, ranging from 32 to 256. We validate our tokenizer using reconstruction loss and FID metrics, demonstrating that token count aligns with image entropy, familiarity and downstream task requirements. Recurrent token processing with increasing representational capacity in each iteration shows signs of token specialization, revealing potential for object / part discovery.

Modern vision models excel at general purpose downstream tasks. It is unclear, however, how they may be used for personalized vision tasks, which are both fine-grained and data-scarce. Recent works have successfully applied synthetic data to general-purpose representation learning, while advances in T2I diffusion models have enabled the generation of personalized images from just a few real examples. Here, we explore a potential connection between these ideas, and formalize the challenge of using personalized synthetic data to learn personalized representations, which encode knowledge about an object of interest and may be flexibly applied to any downstream task relating to the target object. We introduce an evaluation suite for this challenge, including reformulations of two existing datasets and a novel dataset explicitly constructed for this purpose, and propose a contrastive learning approach that makes creative use of image generators. We show that our method improves personalized representation learning for diverse downstream tasks, from recognition to segmentation, and analyze characteristics of image generation approaches that are key to this gain.

We train representation models with procedural data only, and apply them on visual similarity, classification, and semantic segmentation tasks without further training by using visual memory—an explicit database of reference image embeddings. Unlike prior work on visual memory, our approach achieves full compartmentalization with respect to all real-world images while retaining strong performance. Compared to a model trained on Places, our procedural model performs within 1% on NIGHTS visual similarity, outperforms by 8% and 15% on CUB200 and Flowers102 fine-grained classification, and is within 10% on ImageNet-1K classification. It also demonstrates strong zero-shot segmentation, achieving an $R^2$ on COCO within 10% of the models trained on real data. Finally, we analyze procedural versus real data models, showing that parts of the same object have dissimilar representations in procedural models, resulting in incorrect searches in memory and explaining the remaining performance gap.

Simulating object dynamics from real-world perception shows great promise for digital twins and robotic manipulation but often demands labor-intensive measurements and expertise. We present a fully automated Real2Sim pipeline that generates simulation-ready assets for real-world objects through robotic interaction. Using only a robot’s joint torque sensors and an external camera, the pipeline identifies visual geometry, collision geometry, and physical properties such as inertial parameters. Our approach introduces a general method for extracting high-quality, object-centric meshes from photometric reconstruction techniques (e.g., NeRF, Gaussian Splatting) by employing alpha-transparent training while explicitly distinguishing foreground occlusions from background subtraction. We validate the full pipeline through extensive experiments, demonstrating its effectiveness across diverse objects. By eliminating the need for manual intervention or environment modifications, our pipeline can be integrated directly into existing pick-and-place setups, enabling scalable and efficient dataset creation. Project page (with code and data): https://scalable-real2sim.github.io/.

According to Algorithmic Information Theory (AIT), intelligent representations compress data into the shortest possible program while remaining predictive of its content—exhibiting low Kolmogorov Complexity (KC). In contrast, most visual representation learning systems assign fixed-length representations to all inputs, ignoring variations in complexity or familiarity. Recent adaptive tokenization methods address this by allocating variable-length representations but typically require test-time search over multiple hypotheses to identify the most predictive one. Inspired by KC principles, we propose a one-shot adaptive tokenizer, KARL, that predicts the appropriate number of tokens for an image in a single forward pass, halting once its approximate KC is reached. The token count serves as a proxy for the minimum description length. KARL performs comparably to recent adaptive tokenizers while operating in a one-pass manner. Additionally, we present a conceptual study showing a correlation between adaptive tokenization and core ideas from AIT. We demonstrate that adaptive tokenization not only aligns with KC but also reveals empirical signals approximating AIT concepts such as sophistication and logical depth. Finally, we analyze predicted image complexity and interestingness across axes such as structure vs. noise and in-distribution vs. out-of-distribution familiarity, highlighting alignment with human annotations.

Recent significant advances in text-to-image models un-lock the possibility of training vision systems using synthetic images, potentially overcoming the difficulty of collecting curated data at scale. It is unclear, however, how these models behave at scale, as more synthetic data is added to the training set. In this paper we study the scaling laws of synthetic images generated by state of the art text-to-image models, for the training of supervised models: image classifiers with label supervision, and CLIP with language super-vision. We identify several factors, including text prompts, classifier-free guidance scale, and types of text-to-image models, that significantly affect scaling behavior. After tuning these factors, we observe that synthetic images demon-strate a scaling trend similar to, but slightly less effective than, real images in CLIP training, while they significantly underperform in scaling when training supervised image classifiers. Our analysis indicates that the main reason for this underperformance is the inability of off-the-shelf text-to-image models to generate certain concepts, a limitation that significantly impairs the training of image classifiers. Our findings also suggest that scaling synthetic data can be particularly effective in scenarios such as: (1) when there is a limited supply of real images for a supervised problem (e.g., fewer than 0.5 million images in ImageNet), (2) when the evaluation dataset diverges significantly from the training data, indicating the out-of-distribution scenario, or (3) when synthetic data is used in conjunction with real images, as demonstrated in the training of CLIP models.

What does learning to model relationships between strings teach Large Language Models (LLMs) about the visual world? We systematically evaluate LLMs' abilities to generate and recognize an assortment of visual concepts of increasing complexity and then demonstrate how a preliminary visual representation learning system can be trained using models of text. As language models lack the ability to consume or output visual information as pixels, we use code to represent images in our study. Although LLM-generated images do not look like natural images, results on image generation and the ability of models to correct these generated images indicate that precise modeling of strings can teach language models about numerous aspects of the visual world. Furthermore, experiments on self-supervised visual representation learning, utilizing images generated with text models, highlight the potential to train vision models capable of making semantic assessments of natural images using just LLMs.

We introduce SynCLR, a novel approach for learning visual representations exclusively from synthetic images and synthetic captions, without any real data. We synthesize a large dataset of image captions using LLMs, then use an off-the-shelf text-to-image model to generate multiple images corresponding to each synthetic caption. We perform visual representation learning on these synthetic images via contrastive learning, treating images sharing the same caption as positive pairs. The resulting representations transfer well to many downstream tasks, competing favorably with other general-purpose visual representation learners such as CLIP and DINO v2 in image classification tasks. Furthermore, in dense prediction tasks such as semantic segmentation, SynCLR outperforms previous self-supervised methods by a significant margin, e.g., improving over MAE and iBOT by 6.2 and 4.3 mIoU on ADE20k for ViT-B/16.

We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato’s concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.

To improve performance in contemporary deep learning, one is interested in scaling up the neural network in terms of both the number and the size of the layers. When ramping up the width of a single layer, graceful scaling of training has been linked to the need to normalize the weights and their updates in the "natural norm" particular to that layer. In this paper, we significantly generalize this idea by defining the modular norm, which is the natural norm on the full weight space of any neural network architecture. The modular norm is defined recursively in tandem with the network architecture itself. We show that the modular norm has several promising applications. On the practical side, the modular norm can be used to normalize the updates of any base optimizer so that the learning rate becomes transferable across width and depth. This means that the user does not need to compute optimizer-specific scale factors in order to scale training. On the theoretical side, we show that for any neural network built from "well-behaved" atomic modules, the gradient of the network is Lipschitz-continuous in the modular norm, with the Lipschitz constant admitting a simple recursive formula. This characterization opens the door to porting standard ideas in optimization theory over to deep learning. We have created a Python package called Modula that automatically normalizes weight updates in the modular norm of the architecture. Both the Modula package and code for our experiments are provided in the supplementary material.

Humans judge perceptual similarity according to diverse visual attributes, including scene layout, subject location, and camera pose. Existing vision models understand a wide range of semantic abstractions but improperly weigh these attributes and thus make inferences misaligned with human perception. While vision representations have previously benefited from human preference alignment in contexts like image generation, the utility of perceptually aligned representations in more general-purpose settings remains unclear. Here, we investigate how aligning vision model representations to human perceptual judgments impacts their usability in standard computer vision tasks. We finetune state-of-the-art models on a dataset of human similarity judgments for synthetic image triplets and evaluate them across diverse computer vision tasks. We find that aligning models to perceptual judgments yields representations that improve upon the original backbones across many downstream tasks, including counting, semantic segmentation, depth estimation, instance retrieval, and retrieval-augmented generation. In addition, we find that performance is widely preserved on other tasks, including specialized out-of-distribution domains such as in medical imaging and 3D environment frames. Our results suggest that injecting an inductive bias about human perceptual knowledge into vision models can make them better representation learners.

Despite increasingly realistic image quality, recent 3D image generative models often operate on 3D volumes of fixed extent with limited camera motions. We investigate the task of unconditionally synthesizing unbounded nature scenes, enabling arbitrarily large camera motion while maintaining a persistent 3D world model. Our scene representation consists of an extendable, planar scene layout grid, which can be rendered from arbitrary camera poses via a 3D decoder and volume rendering, and a panoramic sky-dome. Based on this representation, we learn a generative world model solely from single-view internet photos. Our method enables simulating long flights through 3D landscapes, while maintaining global scene consistency-for instance, returning to the starting point yields the same view of the scene. Our approach enables scene extrapolation beyond the fixed bounds of current 3D generative models, while also supporting a persistent, camera-independent world representation that stands in contrast to auto-regressive 3D prediction models. Our project page: https://chail.github.io/persistent-nature/.

In goal-reaching reinforcement learning (RL), the optimal value function has a particular geometry, called quasimetrics structure. This paper introduces Quasimetric Reinforcement Learning (QRL), a new RL method that utilizes quasimetric models to learn optimal value functions. Distinct from prior approaches, the QRL objective is specifically designed for quasimetrics, and provides strong theoretical recovery guarantees. Empirically, we conduct thorough analyses on a discretized MountainCar environment, identifying properties of QRL and its advantages over alternatives. On offline and online goal-reaching benchmarks, QRL also demonstrates improved sample efficiency and performance, across both state-based and image-based observations.

This work examines the challenges of training neural networks using vector quantization using straight-through estimation. We find that the main cause of training instability is the discrepancy between the model embedding and the code-vector distribution. We identify the factors that contribute to this issue, including the codebook gradient sparsity and the asymmetric nature of the commitment loss, which leads to misaligned code-vector assignments. We propose to address this issue via affine re-parameterization of the code vectors. Additionally, we introduce an alternating optimization to reduce the gradient error introduced by the straight-through estimation. Moreover, we propose an improvement to the commitment loss to ensure better alignment between the codebook representation and the model embedding. These optimization methods improve the mathematical approximation of the straight-through estimation and, ultimately, the model performance. We demonstrate the effectiveness of our methods on several common model architectures, such as AlexNet, ResNet, and ViT, across various tasks, including image classification and generative modeling.

Contrastive Language-Image Pre-training (CLIP) stands as one of the most effective and scalable methods for training transferable vision models using paired image and text data. CLIP models are trained using contrastive loss, which typically relies on data augmentations to prevent overfitting and shortcuts. However, in the CLIP training paradigm, data augmentations are exclusively applied to image inputs, while language inputs remain unchanged throughout the entire training process, limiting the exposure of diverse texts to the same image. In this paper, we introduce Language augmented CLIP (LaCLIP), a simple yet highly effective approach to enhance CLIP training through language rewrites. Leveraging the in-context learning capability of large language models, we rewrite the text descriptions associated with each image. These rewritten texts exhibit diversity in sentence structure and vocabulary while preserving the original key concepts and meanings. During training, LaCLIP randomly selects either the original texts or the rewritten versions as text augmentations for each image. Extensive experiments on CC3M, CC12M, RedCaps and LAION-400M datasets show that CLIP pre-training with language rewrites significantly improves the transfer performance without computation or memory overhead during training. Specifically for ImageNet zero-shot accuracy, LaCLIP outperforms CLIP by 8.2% on CC12M and 2.4% on LAION-400M.

Current perceptual similarity metrics operate at the level of pixels and patches. These metrics compare images in terms of their low-level colors and textures, but fail to capture mid-level similarities and differences in image layout, object pose, and semantic content. In this paper, we develop a perceptual metric that assesses images holistically. Our first step is to collect a new dataset of human similarity judgments over image pairs that are alike in diverse ways. Critical to this dataset is that judgments are nearly automatic and shared by all observers. To achieve this we use recent text-to-image models to create synthetic pairs that are perturbed along various dimensions. We observe that popular perceptual metrics fall short of explaining our new data, and we introduce a new metric, DreamSim, tuned to better align with human perception. We analyze how our metric is affected by different visual attributes, and find that it focuses heavily on foreground objects and semantic content while also being sensitive to color and layout. Notably, despite being trained on synthetic data, our metric generalizes to real images, giving strong results on retrieval and reconstruction tasks. Furthermore, our metric outperforms both prior learned metrics and recent large vision models on these tasks. Our project page: https://dreamsim-nights.github.io/

Neural MMO 2.0 is a massively multi-agent and multi-task environment for reinforcement learning research. This version features a novel task-system that broadens the range of training settings and poses a new challenge in generalization: evaluation on and against tasks, maps, and opponents never seen during training. Maps are procedurally generated with 128 agents in the standard setting and 1-1024 supported overall. Version 2.0 is a complete rewrite of its predecessor with three-fold improved performance, effectively addressing simulation bottlenecks in online training. Enhancements to compatibility enable training with standard reinforcement learning frameworks designed for much simpler environments. Neural MMO 2.0 is free and open-source with comprehensive documentation available at neuralmmo.github.io and an active community Discord. To spark initial research on this new platform, we are concurrently running a competition at NeurIPS 2023.

We investigate the potential of learning visual representations using synthetic images generated by text-to-image models. This is a natural question in the light of the excellent performance of such models in generating high-quality images. We consider specifically the Stable Diffusion, one of the leading open source text-to-image models. We show that (1) when the generative model is properly configured, training self-supervised methods on synthetic images can match or beat the real image counterpart; (2) by treating the multiple images generated from the same text prompt as positives for each other, we develop a multi-positive contrastive learning method, which we call StableRep. With solely synthetic images, the representations learned by StableRep surpass the performance of representations learned by SimCLR and CLIP using the same set of text prompts and corresponding real images, on large scale datasets. When we further add language supervision, \name~trained with 20M synthetic images (10M captions) achieves better accuracy than CLIP trained with 50M real images (50M captions).

This paper presents an unpaired method for creating line drawings from photographs. Current methods often rely on high quality paired datasets to generate line drawings. However, these datasets often have limitations due to the subjects of the drawings belonging to a specific domain, or in the amount of data collected. Although recent work in unsupervised image-to-image translation has shown much progress, the latest methods still struggle to generate compelling line drawings. We observe that line drawings are encodings of scene information and seek to convey 3D shape and semantic meaning. We build these observations into a set of objectives and train an image translation to map photographs into line drawings. We introduce a geometry loss which predicts depth information from the image features of a line drawing, and a semantic loss which matches the CLIP features of a line drawing with its corresponding photograph. Our approach outperforms state-of-the-art un-paired image translation and line drawing generation methods on creating line drawings from arbitrary photographs.

The ability to separate signal from noise, and reason with clean abstractions, is critical to intelligence. With this ability, humans can efficiently perform real world tasks without considering all possible nuisance factors. How can artificial agents do the same? What kind of information can agents safely discard as noises? In this work, we categorize information out in the wild into four types based on controllability and relation with reward, and formulate useful information as that which is both controllable and reward-relevant. This framework clarifies the kinds information removed by various prior work on representation learning in reinforcement learning (RL), and leads to our proposed approach of learning a Denoised MDP that explicitly factors out certain noise distractors. Extensive experiments on variants of DeepMind Control Suite and RoboDesk demonstrate superior performance of our denoised world model over using raw observations alone, and over prior works, across policy optimization control tasks as well as the non-control task of joint position regression. Project Page: https://ssnl.github.io/denoised_mdp/ Code: https://github.com/facebookresearch/denoised_mdp/

Thin, reflective objects such as forks and whisks are common in our daily lives, but they are particularly chal-lenging for robot perception because it is hard to reconstruct them using commodity RGB-D cameras or multi-view stereo techniques. While traditional pipelines struggle with objects like these, Neural Radiance Fields (NeRFs) have recently been shown to be remarkably effective for performing view synthesis on objects with thin structures or reflective materials. In this paper we explore the use of NeRF as a new source of supervision for robust robot vision systems. In particular, we demonstrate that a NeRF representation of a scene can be used to train dense object descriptors. We use an optimized NeRF to extract dense correspondences between multiple views of an object, and then use these correspondences as training data for learning a view-invariant representation of the object. NeRF's usage of a density field allows us to reformulate the correspondence problem with a novel distribution-of-depths formulation, as opposed to the conventional approach of using a depth map. Dense correspondence models supervised with our method significantly outperform off-the-shelf learned descriptors by 106% (PCK@3px metric, more than doubling performance) and outperform our baseline supervised with multi-view stereo by 29%. Furthermore, we demonstrate the learned dense descriptors enable robots to perform accurate 6-degree of freedom (6-DoF) pick and place of thin and reflective objects.

Humans integrate multiple sensory modalities (e.g., visual and audio) to build a causal understanding of the physical world. In this work, we propose a novel type of intrinsic motivation for Reinforcement Learning (RL) that encourages the agent to understand the causal effect of its actions through auditory event prediction. First, we allow the agent to collect a small amount of acoustic data and use K-means to discover underlying auditory event clusters. We then train a neural network to predict the auditory events and use the prediction errors as intrinsic rewards to guide RL exploration. We first conduct proof-of-concept experiments using a set of Atari games for an in-depth analysis of our module. We then apply our model to embodied audio-visual exploration using the Habitat simulator and active exploration with a rolling robot using the ThreeDWorld (TDW) simulator. Experimental results demonstrate the advantages of using audio signals over vision-based models as intrinsic rewards to guide RL explorations.

Learning image representations using synthetic data allows training neural networks without some of the concerns associated with real images, such as privacy and bias. Existing work focuses on a handful of curated generative processes which require expert knowledge to design, making it hard to scale up. To overcome this, we propose training with a large dataset of twenty-one thousand programs, each one generating a diverse set of synthetic images. These programs are short code snippets, which are easy to modify and fast to execute using OpenGL. The proposed dataset can be used for both supervised and unsupervised representation learning, and reduces the gap between pre-training with real and procedurally generated images by 38%.

Meaningful uncertainty quantification in computer vision requires reasoning about semantic information---say, the hair color of the person in a photo or the location of a car on the street. To this end, recent breakthroughs in generative modeling allow us to represent semantic information in disentangled latent spaces, but providing uncertainties on the semantic latent variables has remained challenging. In this work, we provide principled uncertainty intervals that are guaranteed to contain the true semantic factors for any underlying generative model. The method does the following: (1) it uses quantile regression to output a heuristic uncertainty interval for each element in the latent space (2) calibrates these uncertainties such that they contain the true value of the latent for a new, unseen input. The endpoints of these calibrated intervals can then be propagated through the generator to produce interpretable uncertainty visualizations for each semantic factor. This technique reliably communicates semantically meaningful, principled, and instance-adaptive uncertainty in inverse problems like image super-resolution and image completion. Project page: https://swamiviv.github.io/semanticuncertaintyintervals/

We introduce an offline multi-agent reinforcement learning ( offline MARL) framework that utilizes previously collected data without additional online data collection. Our method reformulates offline MARL as a sequence modeling problem and thus builds on top of the simplicity and scalability of the Transformer architecture. In the fashion of centralized training and decentralized execution, we propose to first train a teacher policy as if the MARL dataset is generated by a single agent. After the teacher policy has identified and recombined the "good" behavior in the dataset, we create separate student policies and distill not only the teacher policy's features but also its structural relations among different agents' features to student policies. Despite its simplicity, the proposed method outperforms state-of-the-art model-free offline MARL baselines while being more robust to demonstration's quality on several environments.

Recent generative models can synthesize "views" of artificial images that mimic real-world variations, such as changes in color or pose, simply by learning from unlabeled image collections. Here, we investigate whether such views can be applied to real images to benefit downstream analysis tasks such as image classification. Using a pre-trained generator, we first find the latent code corresponding to a given real input image. Applying perturbations to the code creates natural variations of the image, which can then be ensembled together at test-time. We use StyleGAN2 as the source of generative augmentations and investigate this setup on classification tasks involving facial attributes, cat faces, and cars. Critically, we find that several design decisions are required towards making this process work; the perturbation procedure, weighting between the augmentations and original image, and training the classifier on synthesized images can all impact the result. Currently, we find that while test-time ensembling with GAN-based augmentations can offer some small improvements, the remaining bottlenecks are the efficiency and accuracy of the GAN reconstructions, coupled with classifier sensitivities to artifacts in GAN-generated images.

Self-supervised representation learning has achieved remarkable success in recent years. By subverting the need for supervised labels, such approaches are able to utilize the numerous unlabeled images that exist on the Internet and in photographic datasets. Yet to build truly intelligent agents, we must construct representation learning algorithms that can learn not only from datasets but also learn from environments. An agent in a natural environment will not typically be fed curated data. Instead, it must explore its environment to acquire the data it will learn from. We propose a framework, curious representation learning (CRL), which jointly learns a reinforcement learning policy and a visual representation model. The policy is trained to maximize the error of the representation learner, and in doing so is incentivized to explore its environment. At the same time, the learned representation becomes stronger and stronger as the policy feeds it ever harder data to learn from. Our learned representations enable promising transfer to downstream navigation tasks, performing better than or comparably to ImageNet pretraining without using any supervision at all. In addition, despite being trained in simulation, our learned representations can obtain interpretable results on real images. Code is available at https://yilundu.github.io/crl/.

Image classification models can depend on multiple different semantic attributes of the image. An explanation of the decision of the classifier needs to both discover and visualize these properties. Here we present StylEx, a method for doing this, by training a generative model to specifically explain multiple attributes that underlie classifier decisions. A natural source for such attributes is the StyleSpace of StyleGAN, which is known to generate semantically meaningful dimensions in the image. However, because standard GAN training is not dependent on the classifier, it may not represent those attributes which are important for the classifier decision, and the dimensions of StyleSpace may represent irrelevant at-tributes. To overcome this, we propose a training procedure for a StyleGAN, which incorporates the classifier model, in order to learn a classifier-specific StyleSpace. Explanatory attributes are then selected from this space. These can be used to visualize the effect of changing multiple attributes per image, thus providing image-specific explanations. We apply StylEx to multiple domains, including animals, leaves, faces and retinal images. For these, we show how an image can be modified in different ways to change its classifier output. Our results show that the method finds attributes that align well with semantic ones, generate meaningful image-specific explanations, and are human-interpretable as measured in user-studies.1

Humans have mental models that allow them to plan, experiment, and reason in the physical world. How should an intelligent agent go about learning such models? In this paper, we will study if models of the world learned in an open-ended physics environment, without any specific tasks, can be reused for downstream physics reasoning tasks. To this end, we build a benchmark Open-ended Physics Environment (OPEn) and also design several tasks to test learning representations in this environment explicitly. This setting reflects the conditions in which real agents (i.e. rolling robots) find themselves, where they may be placed in a new kind of environment and must adapt without any teacher to tell them how this environment works. This setting is challenging because it requires solving an exploration problem in addition to a model building and representation learning problem. We test several existing RL-based exploration methods on this benchmark and find that an agent using unsupervised contrastive learning for representation learning, and impact-driven learning for exploration, achieved the best results. However, all models still fall short in sample efficiency when transferring to the downstream tasks. We expect that OPEn will encourage the development of novel rolling robot agents that can build reusable mental models of the world that facilitate many tasks.

We present iNeRF, a framework that performs mesh-free pose estimation by "inverting" a Neural Radiance Field (NeRF). NeRFs have been shown to be remarkably effective for the task of view synthesis — synthesizing photorealistic novel views of real-world scenes or objects. In this work, we investigate whether we can apply analysis-by-synthesis via NeRF for mesh-free, RGB-only 6DoF pose estimation – given an image, find the translation and rotation of a camera relative to a 3D object or scene. Our method assumes that no object mesh models are available during either training or test time. Starting from an initial pose estimate, we use gradient descent to minimize the residual between pixels rendered from a NeRF and pixels in an observed image. In our experiments, we first study 1) how to sample rays during pose refinement for iNeRF to collect informative gradients and 2) how different batch sizes of rays affect iNeRF on a synthetic dataset. We then show that for complex real-world scenes from the LLFF dataset [21], iNeRF can improve NeRF by estimating the camera poses of novel images and using these images as additional training data for NeRF. Finally, we show iNeRF can perform categorylevel object pose estimation, including object instances not seen during training, with RGB images by inverting a NeRF model inferred from a single view.

In the natural world, life has found innumerable ways to survive and often thrive. Between and even within species, each individual is in some manner unique, and this diversity lends adaptability and robustness to life. In this work, we aim to learn a space of diverse and high-reward policies in a given environment. To this end, we introduce a generative model of policies for reinforcement learning, which maps a low-dimensional latent space to an agent policy space. Our method enables learning an entire population of agent policies, without requiring the use of separate policy parameters. Just as real world populations can adapt and evolve via natural selection, our method is able to adapt to changes in our environment solely by selecting for policies in latent space. We test our generative model’s capabilities in a variety of environments, including an open-ended grid-world and a two-player soccer environment. Code, visualizations, and additional experiments can be found at https://kennyderek.github.io/adap/.

Current vision systems are trained on huge datasets, and these datasets come with costs: curation is expensive, they inherit human biases, and there are concerns over privacy and usage rights. To counter these costs, interest has surged in learning from cheaper data sources, such as unlabeled images. In this paper we go a step further and ask if we can do away with real image datasets entirely, instead learning from procedural noise processes. We investigate a suite of image generation models that produce images from simple random processes. These are then used as training data for a visual representation learner with a contrastive loss. In particular, we study statistical image models, randomly initialized deep generative models, and procedural graphics models.Our findings show that it is important for the noise to capture certain structural properties of real data but that good performance can be achieved even with processes that are far from realistic. We also find that diversity is a key property to learn good representations.

Communication requires having a common language, a lingua franca, between agents. This language could emerge via a consensus process, but it may require many generations of trial and error. Alternatively, the lingua franca can be given by the environment, where agents ground their language in representations of the observed world. We demonstrate a simple way to ground language in learned representations, which facilitates decentralized multi-agent communication and coordination. We find that a standard representation learning algorithm -- autoencoding -- is sufficient for arriving at a grounded common language. When agents broadcast these representations, they learn to understand and respond to each other's utterances and achieve surprisingly strong task performance across a variety of multi-agent communication environments.

Contrastive representation learning has been outstandingly successful in practice. In this work, we identify two key properties related to the contrastive loss: (1) alignment (closeness) of features from positive pairs, and (2) uniformity of the induced distribution of the (normalized) features on the hypersphere. We prove that, asymptotically, the contrastive loss optimizes these properties, and analyze their positive effects on downstream tasks. Empirically, we introduce an optimizable metric to quantify each property. Extensive experiments on standard vision and language datasets confirm the strong agreement between both metrics and downstream task performance. Directly optimizing for these two metrics leads to representations with comparable or better performance at downstream tasks than contrastive learning.

Does having visual priors (e.g. the ability to detect objects) facilitate learning to perform vision-based manipulation (e.g. picking up objects)? We study this problem under the framework of transfer learning, where the model is first trained on a passive vision task (i.e., the data distribution does not depend on the agent’s decisions), then adapted to perform an active manipulation task (i.e., the data distribution does depend on the agent’s decisions). We find that pre-training on vision tasks significantly improves generalization and sample efficiency for learning to manipulate objects. However, realizing these gains requires careful selection of which parts of the model to transfer. Our key insight is that outputs of standard vision models highly correlate with affordance maps commonly used in manipulation. Therefore, we explore directly transferring model parameters from vision networks to affordance prediction networks, and show that this can result in successful zero-shot adaptation, where a robot can pick up certain objects with zero robotic experience. With just a small amount of robotic experience, we can further fine-tune the affordance model to achieve better results. With just 10 minutes of suction experience or 1 hour of grasping experience, our method achieves ∼ 80% success rate at picking up novel objects.

Contrastive learning applied to self-supervised representation learning has seen a resurgence in recent years, leading to state of the art performance in the unsupervised training of deep image models. Modern batch contrastive approaches subsume or significantly outperform traditional contrastive losses such as triplet, max-margin and the N-pairs loss. In this work, we extend the self-supervised batch contrastive approach to the fully-supervised setting, allowing us to effectively leverage label information. Clusters of points belonging to the same class are pulled together in embedding space, while simultaneously pushing apart clusters of samples from different classes. We analyze two possible versions of the supervised contrastive (SupCon) loss, identifying the best-performing formulation of the loss. On ResNet-200, we achieve top-1 accuracy of 81.4% on the ImageNet dataset, which is 0.8% above the best number reported for this architecture. We show consistent outperformance over cross-entropy on other datasets and two ResNet variants. The loss shows benefits for robustness to natural corruptions, and is more stable to hyperparameter settings such as optimizers and data augmentations. In reduced data settings, it outperforms cross-entropy significantly. Our loss function is simple to implement and reference TensorFlow code is released at https://t.ly/supcon.

Contrastive learning between multiple views of the data has recently achieved state of the art performance in the field of self-supervised representation learning. Despite its success, the influence of different view choices has been less studied. In this paper, we use theoretical and empirical analysis to better understand the importance of view selection, and argue that we should reduce the mutual information (MI) between views while keeping task-relevant information intact. To verify this hypothesis, we devise unsupervised and semi-supervised frameworks that learn effective views by aiming to reduce their MI. We also consider data augmentation as a way to reduce MI, and show that increasing data augmentation indeed leads to decreasing MI and improves downstream classification accuracy. As a by-product, we achieve a new state-of-the-art accuracy on unsupervised pre-training for ImageNet classification (73% top-1 linear readout with a ResNet-50).

We introduce a framework that uses Generative Adversarial Networks (GANs) to study cognitive properties like memorability. These attributes are of interest because we do not have a concrete visual definition of what they entail. What does it look like for a dog to be more memorable? GANs allow us to generate a manifold of natural-looking images with fine-grained differences in their visual attributes. By navigating this manifold in directions that increase memorability, we can visualize what it looks like for a particular generated image to become more memorable. The resulting ``visual definitions" surface image properties (like ``object size") that may underlie memorability. Through behavioral experiments, we verify that our method indeed discovers image manipulations that causally affect human memory performance. We further demonstrate that the same framework can be used to analyze image aesthetics and emotional valence. ganalyze.csail.mit.edu.

Generative Adversarial Networks (GANs) typically learn a distribution of images in a large image dataset, and are then able to generate new images from this distribution. However, each natural image has its own internal statistics, captured by its unique distribution of patches. In this paper we propose an "Internal GAN'' (InGAN) - an image-specific GAN - which trains on a single input image and learns its internal distribution of patches. It is then able to synthesize a plethora of new natural images of significantly different sizes, shapes and aspect-ratios - all with the same internal patch-distribution (same "DNA'') as the input image. In particular, despite large changes in global size/shape of the image, all elements inside the image maintain their local size/shape. InGAN is fully unsupervised, requiring no additional data other than the input image itself. Once trained on the input image, it can remap the input to any size or shape in a single feedforward pass, while preserving the same internal patch distribution. InGAN provides a unified framework for a variety of tasks, bridging the gap between textures and natural images.

Pushing is a fundamental robotic skill. Existing work has shown how to exploit models of pushing to achieve a variety of tasks, including grasping under uncertainty, in-hand manipulation and clearing clutter. Such models, however, are approximate, which limits their applicability.Learning-based methods can reason directly from raw sensory data with accuracy, and have the potential to generalize to a wider diversity of scenarios. However, developing and testing such methods requires rich-enough datasets. In this paper we introduce Omnipush, a dataset with high variety of planar pushing behavior.In particular, we provide 250 pushes for each of 250 objects, all recorded with RGB-D and a high precision tracking system. The objects are constructed so as to systematically explore key factors that affect pushing-the shape of the object and its mass distribution-which have not been broadly explored in previous datasets, and allow to study generalization in model learning.Omnipush includes a benchmark for meta-learning dynamic models, which requires algorithms that make good predictions and estimate their own uncertainty. We also provide an RGB video prediction benchmark and propose other relevant tasks that can be suited with this dataset. Data and code are available at https://web.mit.edu/mcube/omnipush-dataset/.

Contemporary sensorimotor learning approaches typically start with an existing complex agent (e.g., a robotic arm), which they learn to control. In contrast, this paper investigates a modular co-evolution strategy: a collection of primitive agents learns to dynamically self-assemble into composite bodies while also learning to coordinate their behavior to control these bodies. Each primitive agent consists of a limb with a motor attached at one end. Limbs may choose to link up to form collectives. When a limb initiates a link-up action and there is another limb nearby, the latter is magnetically connected to the 'parent' limb's motor. This forms a new single agent, which may further link with other agents. In this way, complex morphologies can emerge, controlled by a policy whose architecture is in explicit correspondence with the morphology. We evaluate the performance of these dynamic and modular agents in simulated environments. We demonstrate better generalization to test-time changes both in the environment, as well as in the structure of the agent, compared to static and monolithic baselines. Project videos and source code are provided in the supplementary material.

While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on ImageNet classification has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called "perceptual losses"? What elements are critical for their success? To answer these questions, we introduce a new dataset of human perceptual similarity judgments. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by large margins on our dataset. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations.

Domain adaptation is critical for success in new, unseen environments. Adversarial adaptation models have shown tremendous progress towards adapting to new environments by focusing either on discovering domain invariant representations or by mapping between unpaired image domains. While feature space methods are difficult to interpret and sometimes fail to capture pixel-level and low-level domain shifts, image space methods sometimes fail to incorporate high level semantic knowledge relevant for the end task. We propose a model which adapts between domains using both generative image space alignment and latent representation space alignment. Our approach, Cycle-Consistent Adversarial Domain Adaptation (CyCADA), guides transfer between domains according to a specific discriminatively trained task and avoids divergence by enforcing consistency of the relevant semantics before and after adaptation. We evaluate our method on a variety of visual recognition and prediction settings, including digit classification and semantic segmentation of road scenes, advancing state-of-the-art performance for unsupervised adaptation from synthetic to real world driving domains.

We propose a metalearning approach for learning gradient-based reinforcement learning (RL) algorithms. The idea is to evolve a differentiable loss function, such that an agent, which optimizes its policy to minimize this loss, will achieve high rewards. The loss is parametrized via temporal convolutions over the agent's experience. Because this loss is highly flexible in its ability to take into account the agent's history, it enables fast task learning. Empirical results show that our evolved policy gradient algorithm (EPG) achieves faster learning on several randomized environments compared to an off-the-shelf policy gradient method. We also demonstrate that EPG's learned loss can generalize to out-of-distribution test time tasks, and exhibits qualitatively different behavior from other popular metalearning algorithms.

We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. Moreover, since the release of the pi×2pi× software associated with this paper, hundreds of twitter users have posted their own artistic experiments using our system. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without handengineering our loss functions either.

We propose split-brain autoencoders, a straightforward modification of the traditional autoencoder architecture, for unsupervised representation learning. The method adds a split to the network, resulting in two disjoint sub-networks. Each sub-network is trained to perform a difficult task - predicting one subset of the data channels from another. Together, the sub-networks extract features from the entire input signal. By forcing the network to solve crosschannel prediction tasks, we induce a representation within the network which transfers well to other, unseen tasks. This method achieves state-of-the-art performance on several large-scale transfer learning benchmarks.

Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs. However, for many tasks, paired training data will not be available. We present an approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples. Our goal is to learn a mapping G : X → Y such that the distribution of images from G(X) is indistinguishable from the distribution Y using an adversarial loss. Because this mapping is highly under-constrained, we couple it with an inverse mapping F : Y → X and introduce a cycle consistency loss to push F(G(X)) ≈ X (and vice versa). Qualitative results are presented on several tasks where paired training data does not exist, including collection style transfer, object transfiguration, season transfer, photo enhancement, etc. Quantitative comparisons against several prior methods demonstrate the superiority of our approach.

Manipulation of deformable objects, such as ropes and cloth, is an important but challenging problem in robotics. We present a learning-based system where a robot takes as input a sequence of images of a human manipulating a rope from an initial to goal configuration, and outputs a sequence of actions that can reproduce the human demonstration, using only monocular images as input. To perform this task, the robot learns a pixel-level inverse dynamics model of rope manipulation directly from images in a self-supervised manner, using about 60K interactions with the rope collected autonomously by the robot. The human demonstration provides a high-level plan of what to do and the low-level inverse model is used to execute the plan. We show that by combining the high and low-level plans, the robot can successfully manipulate a rope into a variety of target shapes using only a sequence of human-provided images for direction.

We propose a deep learning approach for user-guided image colorization. The system directly maps a grayscale image, along with sparse, local user "hints" to an output colorization with a Convolutional Neural Network (CNN). Rather than using hand-defined rules, the network propagates user edits by fusing low-level cues along with high-level semantic information, learned from large-scale data. We train on a million images, with simulated user inputs. To guide the user towards efficient input selection, the system recommends likely colors based on the input image and current user inputs. The colorization is performed in a single feed-forward pass, enabling real-time use. Even with randomly simulated user inputs, we show that the proposed system helps novice users quickly create realistic colorizations, and offers large improvements in colorization quality with just a minute of use. In addition, we demonstrate that the framework can incorporate other user "hints" to the desired colorization, showing an application to color histogram transfer.

Objects make distinctive sounds when they are hit or scratched. These sounds reveal aspects of an object's material properties, as well as the actions that produced them. In this paper, we propose the task of predicting what sound an object makes when struck as a way of studying physical interactions within a visual scene. We present an algorithm that synthesizes sound from silent videos of people hitting and scratching objects with a drumstick. This algorithm uses a recurrent neural network to predict sound features from videos and then produces a waveform from these features with an example-based synthesis procedure. We show that the sounds predicted by our model are realistic enough to fool participants in a "real or fake" psychophysical experiment, and that they convey significant information about material properties and physical interactions.

Objects in visual scenes come in a rich variety of transformed states. A few classes of transformation have been heavily studied in computer vision: mostly simple, parametric changes in color and geometry. However, transformations in the physical world occur in many more flavors, and they come with semantic meaning: e.g., bending, folding, aging, etc. The transformations an object can undergo tell us about its physical and functional properties. In this paper, we introduce a dataset of objects, scenes, and materials, each of which is found in a variety of transformed states. Given a novel collection of images, we show how to explain the collection in terms of the states and transformations it depicts. Our system works by generalizing across object classes: states and transformations learned on one set of objects are used to interpret the image collection for an entirely new object class.

We propose a framework that infers mid-level visual properties of an image by learning about ordinal relationships. Instead of estimating metric quantities directly, the system proposes pairwise relationship estimates for points in the input image. These sparse probabilistic ordinal measurements are globalized to create a dense output map of continuous metric measurements. Estimating order relationships between pairs of points has several advantages over metric estimation: it solves a simpler problem than metric regression, humans are better at relative judgements, so data collection is easier, ordinal relationships are invariant to monotonic transformations of the data, thereby increasing the robustness of the system and providing qualitatively different information. We demonstrate that this frame-work works well on two important mid-level vision tasks: intrinsic image decomposition and depth from an RGB image. We train two systems with the same architecture on data from these two modalities. We provide an analysis of the resulting models, showing that they learn a number of simple rules to make ordinal decisions. We apply our algorithm to depth estimation, with good results, and intrinsic image decomposition, with state-of-the-art results.

When glancing at a magazine, or browsing the Internet, we are continuously exposed to photographs. Despite this overflow of visual information, humans are extremely good at remembering thousands of pictures along with some of their visual details. But not all images are equal in memory. Some stick in our minds while others are quickly forgotten. In this paper, we focus on the problem of predicting how memorable an image will be. We show that memorability is an intrinsic and stable property of an image that is shared across different viewers, and remains stable across delays. We introduce a database for which we have measured the probability that each picture will be recognized after a single view. We analyze a collection of image features, labels, and attributes that contribute to making an image memorable, and we train a predictor based on global image descriptors. We find that predicting image memorability is a task that can be addressed with current computer vision techniques. While making memorable images is a challenging task in visualization, photography, and education, this work is a first attempt to quantify this useful property of images.

To quickly synthesize complex scenes, digital artists often collage together visual elements from multiple sources: for example, mountains from New Zealand behind a Scottish castle with wisps of Saharan sand in front. In this paper, we propose to use a similar process in order to parse a scene. We model a scene as a collage of warped, layered objects sampled from labeled, reference images. Each object is related to the rest by a set of support constraints. Scene parsing is achieved through analysis-by-synthesis. Starting with a dataset of labeled exemplar scenes, we retrieve a dictionary of candidate object segments that match a query image. We then combine elements of this set into a "scene collage" that explains the query image. Beyond just assigning object labels to pixels, scene collaging produces a lot more information such as the number of each type of object in the scene, how they support one another, the ordinal depth of each object, and, to some degree, occluded content. We exploit this representation for several applications: image editing, random scene synthesis, and image-to-anaglyph.

An ongoing debate in the Visualization community concerns the role that visualization types play in data understanding. In human cognition, understanding and memorability are intertwined. As a first step towards being able to ask questions about impact and effectiveness, here we ask: 'What makes a visualization memorable?' We ran the largest scale visualization study to date using 2,070 single-panel visualizations, categorized with visualization type (e.g., bar chart, line graph, etc.), collected from news media sites, government reports, scientific journals, and infographic sources. Each visualization was annotated with additional attributes, including ratings for data-ink ratios and visual densities. Using Amazon's Mechanical Turk, we collected memorability scores for hundreds of these visualizations, and discovered that observers are consistent in which visualizations they find memorable and forgettable. We find intuitive results (e.g., attributes like color and the inclusion of a human recognizable object enhance memorability) and less intuitive results (e.g., common graphs are less memorable than unique visualization types). Altogether our findings suggest that quantifying memorability is a general metric of the utility of information, an essential step towards determining how to design effective visualizations.

When glancing at a magazine, or browsing the Internet, we are continuously being exposed to photographs. Despite of this overflow of visual information, humans are extremely good at remembering thousands of pictures along with some of their visual details. But not all images are equal in memory. Some stitch to our minds, and other are forgotten. In this paper we focus on the problem of predicting how memorable an image will be. We show that memorability is a stable property of an image that is shared across different viewers. We introduce a database for which we have measured the probability that each picture will be remembered after a single view. We analyze image features and labels that contribute to making an image memorable, and we train a predictor based on global image descriptors. We find that predicting image memorability is a task that can be addressed with current computer vision techniques. Whereas making memorable images is a challenging task in visualization and photography, this work is a first attempt to quantify this useful quality of images.

Artists, advertisers, and photographers are routinely presented with the task of creating an image that a viewer will remember. While it may seem like image memorability is purely subjective, recent work shows that it is not an inexplicable phenomenon: variation in memorability of images is consistent across subjects, suggesting that some images are intrinsically more memorable than others, independent of a subjects' contexts and biases. In this paper, we used the publicly available memorability dataset of Isola et al., and augmented the object and scene annotations with interpretable spatial, content, and aesthetic image properties. We used a feature-selection scheme with desirable explaining-away properties to determine a compact set of attributes that characterizes the memorability of any individual image. We find that images of enclosed spaces containing people with visible faces are memorable, while images of vistas and peaceful scenes are not. Contrary to popular belief, unusual or aesthetically pleasing scenes do not tend to be highly memorable. This work represents one of the first attempts at understanding intrinsic image memorability, and opens a new domain of investigation at the interface between human cognition and computer vision.