Amir Zamir

Yajie Bao, Yang Li, Shao-Lun Huang et al.

Task transfer learning is a popular technique in image processing applications that uses pre-trained models to reduce the supervision cost of related tasks. An important question is to determine task transferability, i.e. given a common input domain, estimating to what extent representations learned from a source task can help in learning a target task. Typically, transferability is either measured experimentally or inferred through task relatedness, which is often defined without a clear operational meaning. In this paper, we present a novel metric, H-score, an easily-computable evaluation function that estimates the performance of transferred representations from one task to another in classification problems using statistical and information theoretic principles. Experiments on real image data show that our metric is not only consistent with the empirical transferability measurement, but also useful to practitioners in applications such as source model selection and task transfer curriculum learning.

Roman Bachmann, David Mizrahi, Andrei Atanov et al.

We propose a pre-training strategy called Multi-modal Multi-task Masked Autoencoders (MultiMAE). It differs from standard Masked Autoencoding in two key aspects: I) it can optionally accept additional modalities of information in the input besides the RGB image (hence "multi-modal"), and II) its training objective accordingly includes predicting multiple outputs besides the RGB image (hence "multi-task"). We make use of masking (across image patches and input modalities) to make training MultiMAE tractable as well as to ensure cross-modality predictive coding is indeed learned by the network. We show this pre-training strategy leads to a flexible, simple, and efficient framework with improved transfer results to downstream tasks. In particular, the same exact pre-trained network can be flexibly used when additional information besides RGB images is available or when no information other than RGB is available - in all configurations yielding competitive to or significantly better results than the baselines. To avoid needing training datasets with multiple modalities and tasks, we train MultiMAE entirely using pseudo labeling, which makes the framework widely applicable to any RGB dataset. The experiments are performed on multiple transfer tasks (image classification, semantic segmentation, depth estimation) and datasets (ImageNet, ADE20K, Taskonomy, Hypersim, NYUv2). The results show an intriguingly impressive capability by the model in cross-modal/task predictive coding and transfer.

Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov et al.

We introduce a set of image transformations that can be used as corruptions to evaluate the robustness of models as well as data augmentation mechanisms for training neural networks. The primary distinction of the proposed transformations is that, unlike existing approaches such as Common Corruptions, the geometry of the scene is incorporated in the transformations -- thus leading to corruptions that are more likely to occur in the real world. We also introduce a set of semantic corruptions (e.g. natural object occlusions). We show these transformations are `efficient' (can be computed on-the-fly), `extendable' (can be applied on most image datasets), expose vulnerability of existing models, and can effectively make models more robust when employed as `3D data augmentation' mechanisms. The evaluations on several tasks and datasets suggest incorporating 3D information into benchmarking and training opens up a promising direction for robustness research.

Marius Memmel, Roman Bachmann, Amir Zamir

Effectively localizing an agent in a realistic, noisy setting is crucial for many embodied vision tasks. Visual Odometry (VO) is a practical substitute for unreliable GPS and compass sensors, especially in indoor environments. While SLAM-based methods show a solid performance without large data requirements, they are less flexible and robust w.r.t. to noise and changes in the sensor suite compared to learning-based approaches. Recent deep VO models, however, limit themselves to a fixed set of input modalities, e.g., RGB and depth, while training on millions of samples. When sensors fail, sensor suites change, or modalities are intentionally looped out due to available resources, e.g., power consumption, the models fail catastrophically. Furthermore, training these models from scratch is even more expensive without simulator access or suitable existing models that can be fine-tuned. While such scenarios get mostly ignored in simulation, they commonly hinder a model's reusability in real-world applications. We propose a Transformer-based modality-invariant VO approach that can deal with diverse or changing sensor suites of navigation agents. Our model outperforms previous methods while training on only a fraction of the data. We hope this method opens the door to a broader range of real-world applications that can benefit from flexible and learned VO models.

Andrei Atanov, Andrei Filatov, Teresa Yeo et al.

When developing deep learning models, we usually decide what task we want to solve then search for a model that generalizes well on the task. An intriguing question would be: what if, instead of fixing the task and searching in the model space, we fix the model and search in the task space? Can we find tasks that the model generalizes on? How do they look, or do they indicate anything? These are the questions we address in this paper. We propose a task discovery framework that automatically finds examples of such tasks via optimizing a generalization-based quantity called agreement score. We demonstrate that one set of images can give rise to many tasks on which neural networks generalize well. These tasks are a reflection of the inductive biases of the learning framework and the statistical patterns present in the data, thus they can make a useful tool for analysing the neural networks and their biases. As an example, we show that the discovered tasks can be used to automatically create adversarial train-test splits which make a model fail at test time, without changing the pixels or labels, but by only selecting how the datapoints should be split between the train and test sets. We end with a discussion on human-interpretability of the discovered tasks.

Teresa Yeo, Oğuzhan Fatih Kar, Zahra Sodagar et al.

We propose a method for adapting neural networks to distribution shifts at test-time. In contrast to training-time robustness mechanisms that attempt to anticipate and counter the shift, we create a closed-loop system and make use of a test-time feedback signal to adapt a network on the fly. We show that this loop can be effectively implemented using a learning-based function, which realizes an amortized optimizer for the network. This leads to an adaptation method, named Rapid Network Adaptation (RNA), that is notably more flexible and orders of magnitude faster than the baselines. Through a broad set of experiments using various adaptation signals and target tasks, we study the efficiency and flexibility of this method. We perform the evaluations using various datasets (Taskonomy, Replica, ScanNet, Hypersim, COCO, ImageNet), tasks (depth, optical flow, semantic segmentation, classification), and distribution shifts (Cross-datasets, 2D and 3D Common Corruptions) with promising results. We end with a discussion on general formulations for handling distribution shifts and our observations from comparing with similar approaches from other domains.

Onur Beker, Mohammad Mohammadi, Amir Zamir

To achieve autonomy in a priori unknown real-world scenarios, agents should be able to: i) act from high-dimensional sensory observations (e.g., images), ii) learn from past experience to adapt and improve, and iii) be capable of long horizon planning. Classical planning algorithms (e.g. PRM, RRT) are proficient at handling long-horizon planning. Deep learning based methods in turn can provide the necessary representations to address the others, by modeling statistical contingencies between observations. In this direction, we introduce a general-purpose planning algorithm called PALMER that combines classical sampling-based planning algorithms with learning-based perceptual representations. For training these perceptual representations, we combine Q-learning with contrastive representation learning to create a latent space where the distance between the embeddings of two states captures how easily an optimal policy can traverse between them. For planning with these perceptual representations, we re-purpose classical sampling-based planning algorithms to retrieve previously observed trajectory segments from a replay buffer and restitch them into approximately optimal paths that connect any given pair of start and goal states. This creates a tight feedback loop between representation learning, memory, reinforcement learning, and sampling-based planning. The end result is an experiential framework for long-horizon planning that is significantly more robust and sample efficient compared to existing methods.

Andrei Atanov, Jesse Allardice, Roman Bachmann et al.

Visual tokenizers map high-dimensional raw pixels into a compressed representation for downstream modeling. Beyond compression, tokenizers dictate what information is preserved and how it is organized. A de facto standard approach to video tokenization is to represent a video as a spatiotemporal 3D grid of tokens, each capturing the corresponding local information in the original signal. This requires the downstream model that consumes the tokens, e.g., a text-to-video model, to learn to predict all low-level details "pixel-by-pixel" irrespective of the video's inherent complexity, leading to high learning complexity. We present VideoFlexTok, which represents videos with a variable-length sequence of tokens structured in a coarse-to-fine manner -- where the first tokens (emergently) capture abstract information, such as semantics and motion, and later tokens add fine-grained details. The generative flow decoder enables realistic video reconstructions from any token count. This representation structure allows adapting the token count according to downstream needs and encoding videos longer than the baselines with the same budget. We evaluate VideoFlexTok on class- and text-to-video generative tasks and show that it leads to more efficient training compared to 3D grid tokens, e.g., achieving comparable generation quality (gFVD and ViCLIP Score) with a 5x smaller model (1.1B vs 5.2B). Finally, we demonstrate how VideoFlexTok can enable long video generation without prohibitive computational cost by training a text-to-video model on 10-second 81-frame videos with only 672 tokens, 8x fewer than a comparable 3D grid tokenizer.

Sogand Salehi, Mahdi Shafiei, Teresa Yeo et al.

Different users find different images generated for the same prompt desirable. This gives rise to personalized image generation which involves creating images aligned with an individual's visual preference. Current generative models are, however, unpersonalized, as they are tuned to produce outputs that appeal to a broad audience. Using them to generate images aligned with individual users relies on iterative manual prompt engineering by the user which is inefficient and undesirable. We propose to personalize the image generation process by first capturing the generic preferences of the user in a one-time process by inviting them to comment on a small selection of images, explaining why they like or dislike each. Based on these comments, we infer a user's structured liked and disliked visual attributes, i.e., their visual preference, using a large language model. These attributes are used to guide a text-to-image model toward producing images that are tuned towards the individual user's visual preference. Through a series of user studies and large language model guided evaluations, we demonstrate that the proposed method results in generations that are well aligned with individual users' visual preferences.

Zhitong Gao, Parham Rezaei, Ali Cy et al.

Tokenization is a key component of autoregressive (AR) generative models, converting raw data into more manageable units for modeling. Commonly, tokens describe local information, such as regions of pixels in images or word pieces in text, and AR generation predicts these tokens in a fixed order. A worthwhile question is whether token structures affect the ability to steer the generation through test-time search, where multiple candidate generations are explored and evaluated by a verifier. Using image generation as our testbed, we hypothesize that recent 1D ordered tokenizers with coarse-to-fine structure can be more amenable to search than classical 2D grid structures. This is rooted in the fact that the intermediate states in coarse-to-fine sequences carry semantic meaning that verifiers can reliably evaluate, enabling effective steering during generation. Through controlled experiments, we find that AR models trained on coarse-to-fine ordered tokens exhibit improved test-time scaling behavior compared to grid-based counterparts. Moreover, we demonstrate that, thanks to the ordered structure, pure test-time search over token sequences (i.e., without training an AR model) can perform training-free text-to-image generation when guided by an image-text verifier. Beyond this, we systematically study how classical search algorithms (best-of-N, beam search, lookahead search) interact with different token structures, as well as the role of different verifiers and AR priors. Our results highlight the impact of token structure on inference-time scalability and provide practical guidance for test-time scaling in AR models.

David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar et al.

Current machine learning models for vision are often highly specialized and limited to a single modality and task. In contrast, recent large language models exhibit a wide range of capabilities, hinting at a possibility for similarly versatile models in computer vision. In this paper, we take a step in this direction and propose a multimodal training scheme called 4M. It consists of training a single unified Transformer encoder-decoder using a masked modeling objective across a wide range of input/output modalities - including text, images, geometric, and semantic modalities, as well as neural network feature maps. 4M achieves scalability by unifying the representation space of all modalities through mapping them into discrete tokens and performing multimodal masked modeling on a small randomized subset of tokens. 4M leads to models that exhibit several key capabilities: (1) they can perform a diverse set of vision tasks out of the box, (2) they excel when fine-tuned for unseen downstream tasks or new input modalities, and (3) they can function as a generative model that can be conditioned on arbitrary modalities, enabling a wide variety of expressive multimodal editing capabilities with remarkable flexibility. Through experimental analyses, we demonstrate the potential of 4M for training versatile and scalable foundation models for vision tasks, setting the stage for further exploration in multimodal learning for vision and other domains.

Roman Bachmann, Oğuzhan Fatih Kar, David Mizrahi et al.

Current multimodal and multitask foundation models like 4M or UnifiedIO show promising results, but in practice their out-of-the-box abilities to accept diverse inputs and perform diverse tasks are limited by the (usually rather small) number of modalities and tasks they are trained on. In this paper, we expand upon the capabilities of them by training a single model on tens of highly diverse modalities and by performing co-training on large-scale multimodal datasets and text corpora. This includes training on several semantic and geometric modalities, feature maps from recent state of the art models like DINOv2 and ImageBind, pseudo labels of specialist models like SAM and 4DHumans, and a range of new modalities that allow for novel ways to interact with the model and steer the generation, for example image metadata or color palettes. A crucial step in this process is performing discrete tokenization on various modalities, whether they are image-like, neural network feature maps, vectors, structured data like instance segmentation or human poses, or data that can be represented as text. Through this, we expand on the out-of-the-box capabilities of multimodal models and specifically show the possibility of training one model to solve at least 3x more tasks/modalities than existing ones and doing so without a loss in performance. This enables more fine-grained and controllable multimodal generation capabilities and allows us to study the distillation of models trained on diverse data and objectives into a unified model. We successfully scale the training to a three billion parameter model using tens of modalities and different datasets. The resulting models and training code are open sourced at 4m.epfl.ch.

Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar et al.

Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. "blindness" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.

Roman Bachmann, Jesse Allardice, David Mizrahi et al.

Image tokenization has enabled major advances in autoregressive image generation by providing compressed, discrete representations that are more efficient to process than raw pixels. While traditional approaches use 2D grid tokenization, recent methods like TiTok have shown that 1D tokenization can achieve high generation quality by eliminating grid redundancies. However, these methods typically use a fixed number of tokens and thus cannot adapt to an image's inherent complexity. We introduce FlexTok, a tokenizer that projects 2D images into variable-length, ordered 1D token sequences. For example, a 256x256 image can be resampled into anywhere from 1 to 256 discrete tokens, hierarchically and semantically compressing its information. By training a rectified flow model as the decoder and using nested dropout, FlexTok produces plausible reconstructions regardless of the chosen token sequence length. We evaluate our approach in an autoregressive generation setting using a simple GPT-style Transformer. On ImageNet, this approach achieves an FID<2 across 8 to 128 tokens, outperforming TiTok and matching state-of-the-art methods with far fewer tokens. We further extend the model to support to text-conditioned image generation and examine how FlexTok relates to traditional 2D tokenization. A key finding is that FlexTok enables next-token prediction to describe images in a coarse-to-fine "visual vocabulary", and that the number of tokens to generate depends on the complexity of the generation task.

Rahul Ramachandran, Ali Garjani, Roman Bachmann et al.

Multimodal foundation models, such as GPT-4o, have recently made remarkable progress, but it is not clear where exactly these models stand in terms of understanding vision. In this paper, we benchmark the performance of popular multimodal foundation models (GPT-4o, o4-mini, Gemini 1.5 Pro and Gemini 2.0 Flash, Claude 3.5 Sonnet, Qwen2-VL, Llama 3.2) on standard computer vision tasks (semantic segmentation, object detection, image classification, depth and surface normal prediction) using established datasets (e.g., COCO, ImageNet and its variants, etc). The main challenges to performing this are: 1) most models are trained to output text and cannot natively express versatile domains, such as segments or 3D geometry, and 2) many leading models are proprietary and accessible only at an API level, i.e., there is no weight access to adapt them. We address these challenges by translating standard vision tasks into equivalent text-promptable and API-compatible tasks via prompt chaining to create a standardized benchmarking framework. We observe that 1) the models are not close to the state-of-the-art specialist models at any task. However, 2) they are respectable generalists; this is remarkable as they are presumably trained on primarily image-text-based tasks. 3) They perform semantic tasks notably better than geometric ones. 4) While the prompt-chaining techniques affect performance, better models exhibit less sensitivity to prompt variations. 5) GPT-4o performs the best among non-reasoning models, securing the top position in 4 out of 6 tasks, 6) reasoning models, e.g. o3, show improvements in geometric tasks, and 7) a preliminary analysis of models with native image generation, like the latest GPT-4o, shows they exhibit quirks like hallucinations and spatial misalignments.

Teresa Yeo, Andrei Atanov, Harold Benoit et al.

We present a method to control a text-to-image generative model to produce training data useful for supervised learning. Unlike previous works that employ an open-loop approach and pre-define prompts to generate new data using either a language model or human expertise, we develop an automated closed-loop system which involves two feedback mechanisms. The first mechanism uses feedback from a given supervised model and finds adversarial prompts that result in image generations that maximize the model loss. While these adversarial prompts result in diverse data informed by the model, they are not informed of the target distribution, which can be inefficient. Therefore, we introduce the second feedback mechanism that guides the generation process towards a certain target distribution. We call the method combining these two mechanisms Guided Adversarial Prompts. We perform our evaluations on different tasks, datasets and architectures, with different types of distribution shifts (spuriously correlated data, unseen domains) and demonstrate the efficiency of the proposed feedback mechanisms compared to open-loop approaches.

Harold Benoit, Liangze Jiang, Andrei Atanov et al. · ibm-research

Supervised learning datasets may contain multiple cues that explain the training set equally well, i.e., learning any of them would lead to the correct predictions on the training data. However, many of them can be spurious, i.e., lose their predictive power under a distribution shift and consequently fail to generalize to out-of-distribution (OOD) data. Recently developed "diversification" methods (Lee et al., 2023; Pagliardini et al., 2023) approach this problem by finding multiple diverse hypotheses that rely on different features. This paper aims to study this class of methods and identify the key components contributing to their OOD generalization abilities. We show that (1) diversification methods are highly sensitive to the distribution of the unlabeled data used for diversification and can underperform significantly when away from a method-specific sweet spot. (2) Diversification alone is insufficient for OOD generalization. The choice of the used learning algorithm, e.g., the model's architecture and pretraining, is crucial. In standard experiments (classification on Waterbirds and Office-Home datasets), using the second-best choice leads to an up to 20\% absolute drop in accuracy. (3) The optimal choice of learning algorithm depends on the unlabeled data and vice versa i.e. they are co-dependent. (4) Finally, we show that, in practice, the above pitfalls cannot be alleviated by increasing the number of diverse hypotheses, the major feature of diversification methods. These findings provide a clearer understanding of the critical design factors influencing the OOD generalization abilities of diversification methods. They can guide practitioners in how to use the existing methods best and guide researchers in developing new, better ones.

Artyom Gadetsky, Andrei Atanov, Yulun Jiang et al.

Recent advances in large language and vision-language models have enabled zero-shot inference, allowing models to solve new tasks without task-specific training. Various adaptation techniques such as prompt engineering, In-Context Learning (ICL), and supervised fine-tuning can further enhance the model's performance on a downstream task, but they require substantial manual effort to construct effective prompts or labeled examples. In this work, we introduce a joint inference framework for fully unsupervised adaptation, eliminating the need for manual prompt engineering and labeled examples. Unlike zero-shot inference, which makes independent predictions, the joint inference makes predictions simultaneously for all inputs in a given task. Since direct joint inference involves computationally expensive optimization, we develop efficient approximation techniques, leading to two unsupervised adaptation methods: unsupervised fine-tuning and unsupervised ICL. We demonstrate the effectiveness of our methods across diverse tasks and models, including language-only Llama-3.1 on natural language processing tasks, reasoning-oriented Qwen2.5-Math on grade school math problems, vision-language OpenFlamingo on vision tasks, and the API-only access GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate substantial improvements over the standard zero-shot approach, including 39% absolute improvement on the challenging GSM8K math reasoning dataset. Remarkably, despite being fully unsupervised, our framework often performs on par with supervised approaches that rely on ground truth labels.

Andrei Atanov, Jiawei Fu, Rishubh Singh et al.

A de facto standard in solving computer vision problems is to use a common high-resolution camera and choose its placement on an agent (i.e., position and orientation) based on human intuition. On the other hand, extremely simple and well-designed visual sensors found throughout nature allow many organisms to perform diverse, complex behaviors. In this work, motivated by these examples, we raise the following questions: 1. How effective simple visual sensors are in solving vision tasks? 2. What role does their design play in their effectiveness? We explore simple sensors with resolutions as low as one-by-one pixel, representing a single photoreceptor First, we demonstrate that just a few photoreceptors can be enough to solve many tasks, such as visual navigation and continuous control, reasonably well, with performance comparable to that of a high-resolution camera. Second, we show that the design of these simple visual sensors plays a crucial role in their ability to provide useful information and successfully solve these tasks. To find a well-performing design, we present a computational design optimization algorithm and evaluate its effectiveness across different tasks and domains, showing promising results. Finally, we perform a human survey to evaluate the effectiveness of intuitive designs devised manually by humans, showing that the computationally found design is among the best designs in most cases.

Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo et al.

Abstraction is at the heart of sketching due to the simple and minimal nature of line drawings. Abstraction entails identifying the essential visual properties of an object or scene, which requires semantic understanding and prior knowledge of high-level concepts. Abstract depictions are therefore challenging for artists, and even more so for machines. We present CLIPasso, an object sketching method that can achieve different levels of abstraction, guided by geometric and semantic simplifications. While sketch generation methods often rely on explicit sketch datasets for training, we utilize the remarkable ability of CLIP (Contrastive-Language-Image-Pretraining) to distill semantic concepts from sketches and images alike. We define a sketch as a set of Bézier curves and use a differentiable rasterizer to optimize the parameters of the curves directly with respect to a CLIP-based perceptual loss. The abstraction degree is controlled by varying the number of strokes. The generated sketches demonstrate multiple levels of abstraction while maintaining recognizability, underlying structure, and essential visual components of the subject drawn.

Andrei Atanov, Shijian Xu, Onur Beker et al.

Transfer learning has witnessed remarkable progress in recent years, for example, with the introduction of augmentation-based contrastive self-supervised learning methods. While a number of large-scale empirical studies on the transfer performance of such models have been conducted, there is not yet an agreed-upon set of control baselines, evaluation practices, and metrics to report, which often hinders a nuanced and calibrated understanding of the real efficacy of the methods. We share an evaluation standard that aims to quantify and communicate transfer learning performance in an informative and accessible setup. This is done by baking a number of simple yet critical control baselines in the evaluation method, particularly the blind-guess (quantifying the dataset bias), scratch-model (quantifying the architectural contribution), and maximal-supervision (quantifying the upper-bound). To demonstrate how the evaluation standard can be employed, we provide an example empirical study investigating a few basic questions about self-supervised learning. For example, using this standard, the study shows the effectiveness of existing self-supervised pre-training methods is skewed towards image classification tasks versus dense pixel-wise predictions. In general, we encourage using/reporting the suggested control baselines in evaluating transfer learning in order to gain a more meaningful and informative understanding.

Ainaz Eftekhar, Alexander Sax, Roman Bachmann et al.

This paper introduces a pipeline to parametrically sample and render multi-task vision datasets from comprehensive 3D scans from the real world. Changing the sampling parameters allows one to "steer" the generated datasets to emphasize specific information. In addition to enabling interesting lines of research, we show the tooling and generated data suffice to train robust vision models. Common architectures trained on a generated starter dataset reached state-of-the-art performance on multiple common vision tasks and benchmarks, despite having seen no benchmark or non-pipeline data. The depth estimation network outperforms MiDaS and the surface normal estimation network is the first to achieve human-level performance for in-the-wild surface normal estimation -- at least according to one metric on the OASIS benchmark. The Dockerized pipeline with CLI, the (mostly python) code, PyTorch dataloaders for the generated data, the generated starter dataset, download scripts and other utilities are available through our project website, https://omnidata.vision.

Teresa Yeo, Oğuzhan Fatih Kar, Alexander Sax et al.

We present a method for making neural network predictions robust to shifts from the training data distribution. The proposed method is based on making predictions via a diverse set of cues (called 'middle domains') and ensembling them into one strong prediction. The premise of the idea is that predictions made via different cues respond differently to a distribution shift, hence one should be able to merge them into one robust final prediction. We perform the merging in a straightforward but principled manner based on the uncertainty associated with each prediction. The evaluations are performed using multiple tasks and datasets (Taskonomy, Replica, ImageNet, CIFAR) under a wide range of adversarial and non-adversarial distribution shifts which demonstrate the proposed method is considerably more robust than its standard learning counterpart, conventional deep ensembles, and several other baselines.

Bryan Chen, Alexander Sax, Gene Lewis et al.

Vision-based robotics often separates the control loop into one module for perception and a separate module for control. It is possible to train the whole system end-to-end (e.g. with deep RL), but doing it "from scratch" comes with a high sample complexity cost and the final result is often brittle, failing unexpectedly if the test environment differs from that of training. We study the effects of using mid-level visual representations (features learned asynchronously for traditional computer vision objectives), as a generic and easy-to-decode perceptual state in an end-to-end RL framework. Mid-level representations encode invariances about the world, and we show that they aid generalization, improve sample complexity, and lead to a higher final performance. Compared to other approaches for incorporating invariances, such as domain randomization, asynchronously trained mid-level representations scale better: both to harder problems and to larger domain shifts. In practice, this means that mid-level representations could be used to successfully train policies for tasks where domain randomization and learning-from-scratch failed. We report results on both manipulation and navigation tasks, and for navigation include zero-shot sim-to-real experiments on real robots.

Amir Zamir, Alexander Sax, Teresa Yeo et al.

Visual perception entails solving a wide set of tasks, e.g., object detection, depth estimation, etc. The predictions made for multiple tasks from the same image are not independent, and therefore, are expected to be consistent. We propose a broadly applicable and fully computational method for augmenting learning with Cross-Task Consistency. The proposed formulation is based on inference-path invariance over a graph of arbitrary tasks. We observe that learning with cross-task consistency leads to more accurate predictions and better generalization to out-of-distribution inputs. This framework also leads to an informative unsupervised quantity, called Consistency Energy, based on measuring the intrinsic consistency of the system. Consistency Energy correlates well with the supervised error (r=0.67), thus it can be employed as an unsupervised confidence metric as well as for detection of out-of-distribution inputs (ROC-AUC=0.95). The evaluations are performed on multiple datasets, including Taskonomy, Replica, CocoDoom, and ApolloScape, and they benchmark cross-task consistency versus various baselines including conventional multi-task learning, cycle consistency, and analytical consistency.

Jeffrey O Zhang, Alexander Sax, Amir Zamir et al.

When training a neural network for a desired task, one may prefer to adapt a pre-trained network rather than starting from randomly initialized weights. Adaptation can be useful in cases when training data is scarce, when a single learner needs to perform multiple tasks, or when one wishes to encode priors in the network. The most commonly employed approaches for network adaptation are fine-tuning and using the pre-trained network as a fixed feature extractor, among others. In this paper, we propose a straightforward alternative: side-tuning. Side-tuning adapts a pre-trained network by training a lightweight "side" network that is fused with the (unchanged) pre-trained network via summation. This simple method works as well as or better than existing solutions and it resolves some of the basic issues with fine-tuning, fixed features, and other common approaches. In particular, side-tuning is less prone to overfitting, is asymptotically consistent, and does not suffer from catastrophic forgetting in incremental learning. We demonstrate the performance of side-tuning under a diverse set of scenarios, including incremental learning (iCIFAR, iTaskonomy), reinforcement learning, imitation learning (visual navigation in Habitat), NLP question-answering (SQuAD v2), and single-task transfer learning (Taskonomy), with consistently promising results.

Alexander Sax, Jeffrey O. Zhang, Bradley Emi et al.

How much does having visual priors about the world (e.g. the fact that the world is 3D) assist in learning to perform downstream motor tasks (e.g. navigating a complex environment)? What are the consequences of not utilizing such visual priors in learning? We study these questions by integrating a generic perceptual skill set (a distance estimator, an edge detector, etc.) within a reinforcement learning framework (see Fig. 1). This skill set ("mid-level vision") provides the policy with a more processed state of the world compared to raw images. Our large-scale study demonstrates that using mid-level vision results in policies that learn faster, generalize better, and achieve higher final performance, when compared to learning from scratch and/or using state-of-the-art visual and non-visual representation learning methods. We show that conventional computer vision objectives are particularly effective in this regard and can be conveniently integrated into reinforcement learning frameworks. Finally, we found that no single visual representation was universally useful for all downstream tasks, hence we computationally derive a task-agnostic set of representations optimized to support arbitrary downstream tasks.

Fei Xia, Amir Zamir, Zhi-Yang He et al.

Developing visual perception models for active agents and sensorimotor control are cumbersome to be done in the physical world, as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given rise to learning-in-simulation which consequently casts a question on whether the results transfer to real-world. In this paper, we are concerned with the problem of developing real-world perception for active agents, propose Gibson Virtual Environment for this purpose, and showcase sample perceptual tasks learned therein. Gibson is based on virtualizing real spaces, rather than using artificially designed ones, and currently includes over 1400 floor spaces from 572 full buildings. The main characteristics of Gibson are: I. being from the real-world and reflecting its semantic complexity, II. having an internal synthesis mechanism, "Goggles", enabling deploying the trained models in real-world without needing further domain adaptation, III. embodiment of agents and making them subject to constraints of physics and space.

Amir Zamir, Alexander Sax, William Shen et al.

Do visual tasks have a relationship, or are they unrelated? For instance, could having surface normals simplify estimating the depth of an image? Intuition answers these questions positively, implying existence of a structure among visual tasks. Knowing this structure has notable values; it is the concept underlying transfer learning and provides a principled way for identifying redundancies across tasks, e.g., to seamlessly reuse supervision among related tasks or solve many tasks in one system without piling up the complexity. We proposes a fully computational approach for modeling the structure of space of visual tasks. This is done via finding (first and higher-order) transfer learning dependencies across a dictionary of twenty six 2D, 2.5D, 3D, and semantic tasks in a latent space. The product is a computational taxonomic map for task transfer learning. We study the consequences of this structure, e.g. nontrivial emerged relationships, and exploit them to reduce the demand for labeled data. For example, we show that the total number of labeled datapoints needed for solving a set of 10 tasks can be reduced by roughly 2/3 (compared to training independently) while keeping the performance nearly the same. We provide a set of tools for computing and probing this taxonomical structure including a solver that users can employ to devise efficient supervision policies for their use cases.

Francesco Castaldo, Amir Zamir, Roland Angst et al.

Matching cross-view images is challenging because the appearance and viewpoints are significantly different. While low-level features based on gradient orientations or filter responses can drastically vary with such changes in viewpoint, semantic information of images however shows an invariant characteristic in this respect. Consequently, semantically labeled regions can be used for performing cross-view matching. In this paper, we therefore explore this idea and propose an automatic method for detecting and representing the semantic information of an RGB image with the goal of performing cross-view matching with a (non-RGB) geographic information system (GIS). A segmented image forms the input to our system with segments assigned to semantic concepts such as traffic signs, lakes, roads, foliage, etc. We design a descriptor to robustly capture both, the presence of semantic concepts and the spatial layout of those segments. Pairwise distances between the descriptors extracted from the GIS map and the query image are then used to generate a shortlist of the most promising locations with similar semantic concepts in a consistent spatial layout. An experimental evaluation with challenging query images and a large urban area shows promising results.

Ozan Sener, Amir Zamir, Silvio Savarese et al.

Human communication typically has an underlying structure. This is reflected in the fact that in many user generated videos, a starting point, ending, and certain objective steps between these two can be identified. In this paper, we propose a method for parsing a video into such semantic steps in an unsupervised way. The proposed method is capable of providing a semantic "storyline" of the video composed of its objective steps. We accomplish this using both visual and language cues in a joint generative model. The proposed method can also provide a textual description for each of the identified semantic steps and video segments. We evaluate this method on a large number of complex YouTube videos and show results of unprecedented quality for this intricate and impactful problem.