Stefan Lee

Erik Wijmans, Manolis Savva, Irfan Essa et al.

Animal navigation research posits that organisms build and maintain internal spatial representations, or maps, of their environment. We ask if machines -- specifically, artificial intelligence (AI) navigation agents -- also build implicit (or 'mental') maps. A positive answer to this question would (a) explain the surprising phenomenon in recent literature of ostensibly map-free neural-networks achieving strong performance, and (b) strengthen the evidence of mapping as a fundamental mechanism for navigation by intelligent embodied agents, whether they be biological or artificial. Unlike animal navigation, we can judiciously design the agent's perceptual system and control the learning paradigm to nullify alternative navigation mechanisms. Specifically, we train 'blind' agents -- with sensing limited to only egomotion and no other sensing of any kind -- to perform PointGoal navigation ('go to $Δ$ x, $Δ$ y') via reinforcement learning. Our agents are composed of navigation-agnostic components (fully-connected and recurrent neural networks), and our experimental setup provides no inductive bias towards mapping. Despite these harsh conditions, we find that blind agents are (1) surprisingly effective navigators in new environments (~95% success); (2) they utilize memory over long horizons (remembering ~1,000 steps of past experience in an episode); (3) this memory enables them to exhibit intelligent behavior (following walls, detecting collisions, taking shortcuts); (4) there is emergence of maps and collision detection neurons in the representations of the environment built by a blind agent as it navigates; and (5) the emergent maps are selective and task dependent (e.g. the agent 'forgets' exploratory detours). Overall, this paper presents no new techniques for the AI audience, but a surprising finding, an insight, and an explanation.

Jacob Krantz, Stefan Lee

Recent work in Vision-and-Language Navigation (VLN) has presented two environmental paradigms with differing realism -- the standard VLN setting built on topological environments where navigation is abstracted away, and the VLN-CE setting where agents must navigate continuous 3D environments using low-level actions. Despite sharing the high-level task and even the underlying instruction-path data, performance on VLN-CE lags behind VLN significantly. In this work, we explore this gap by transferring an agent from the abstract environment of VLN to the continuous environment of VLN-CE. We find that this sim-2-sim transfer is highly effective, improving over the prior state of the art in VLN-CE by +12% success rate. While this demonstrates the potential for this direction, the transfer does not fully retain the original performance of the agent in the abstract setting. We present a sequence of experiments to identify what differences result in performance degradation, providing clear directions for further improvement.

Eric Slyman, Minsuk Kahng, Stefan Lee

Recent work in vision-and-language demonstrates that large-scale pretraining can learn generalizable models that are efficiently transferable to downstream tasks. While this may improve dataset-scale aggregate metrics, analyzing performance around hand-crafted subgroups targeting specific bias dimensions reveals systemic undesirable behaviors. However, this subgroup analysis is frequently stalled by annotation efforts, which require extensive time and resources to collect the necessary data. Prior art attempts to automatically discover subgroups to circumvent these constraints but typically leverages model behavior on existing task-specific annotations and rapidly degrades on more complex inputs beyond "tabular" data, none of which study vision-and-language models. This paper presents VLSlice, an interactive system enabling user-guided discovery of coherent representation-level subgroups with consistent visiolinguistic behavior, denoted as vision-and-language slices, from unlabeled image sets. We show that VLSlice enables users to quickly generate diverse high-coherency slices in a user study (n=22) and release the tool publicly.

Zijiao Yang, Arjun Majumdar, Stefan Lee

To be successful, Vision-and-Language Navigation (VLN) agents must be able to ground instructions to actions based on their surroundings. In this work, we develop a methodology to study agent behavior on a skill-specific basis -- examining how well existing agents ground instructions about stopping, turning, and moving towards specified objects or rooms. Our approach is based on generating skill-specific interventions and measuring changes in agent predictions. We present a detailed case study analyzing the behavior of a recent agent and then compare multiple agents in terms of skill-specific competency scores. This analysis suggests that biases from training have lasting effects on agent behavior and that existing models are able to ground simple referring expressions. Our comparisons between models show that skill-specific scores correlate with improvements in overall VLN task performance.

Zetian Wu, Bowen Xie, Wuyang Meng et al.

The field of sign language translation has witnessed significant progress in the translation between sign and spoken languages, but the translation between sign languages remains largely unexplored and out of reach. The latter can help 1.5 billion deaf and hard-of-hearing (DHH) people worldwide communicate across language barriers without relying on hearing interpreters or written-language fluency. The cascade approach composing separate sign-to-text, text-to-text, and text-to-sign systems suffers from error propagation and extra latency as well as the loss of information unique in the visual modality. We aim to develop direct sign-to-sign translation. However, a large-scale open-domain parallel corpus has not been curated between sign languages. To enable direct translation between sign language utterances, we use back-translation to produce synthetic sign-sign pairs from unaligned individual language utterance-sign corpora. Using this data, we jointly train a single MBART-based model for both text->sign (T2S) and sign->sign (S2S). On synthetically generated paired sets between American Sign Language (ASL), Chinese Sign Language (CSL), and German Sign Language (DGS), our direct S2S method outperforms the cascaded baseline on geometric sign error metrics (20% lower DTW-aligned MPJPE) and language matching metrics after predicted sign utterances are translated back to sentences (50% high BLEU-4) while achieving a roughly 2.3* speedup. On a small set of pre-existing cross-lingual sign data, we find similar improvements for our proposed method.

Yilin Yang, Stefan Lee, Prasad Tadepalli

Beam search decoding is the de-facto method for decoding auto-regressive Neural Machine Translation (NMT) models, including multilingual NMT where the target language is specified as an input. However, decoding multilingual NMT models commonly produces ``off-target'' translations -- yielding translation outputs not in the intended language. In this paper, we first conduct an error analysis of off-target translations for a strong multilingual NMT model and identify how these decodings are produced during beam search. We then propose Language-informed Beam Search (LiBS), a general decoding algorithm incorporating an off-the-shelf Language Identification (LiD) model into beam search decoding to reduce off-target translations. LiBS is an inference-time procedure that is NMT-model agnostic and does not require any additional parallel data. Results show that our proposed LiBS algorithm on average improves +1.1 BLEU and +0.9 BLEU on WMT and OPUS datasets, and reduces off-target rates from 22.9\% to 7.7\% and 65.8\% to 25.3\% respectively.

Michael Cogswell, Jiasen Lu, Rishabh Jain et al.

Can we develop visually grounded dialog agents that can efficiently adapt to new tasks without forgetting how to talk to people? Such agents could leverage a larger variety of existing data to generalize to new tasks, minimizing expensive data collection and annotation. In this work, we study a setting we call "Dialog without Dialog", which requires agents to develop visually grounded dialog models that can adapt to new tasks without language level supervision. By factorizing intention and language, our model minimizes linguistic drift after fine-tuning for new tasks. We present qualitative results, automated metrics, and human studies that all show our model can adapt to new tasks and maintain language quality. Baselines either fail to perform well at new tasks or experience language drift, becoming unintelligible to humans. Code has been made available at https://github.com/mcogswell/dialog_without_dialog

Deshraj Yadav, Rishabh Jain, Harsh Agrawal et al.

We introduce EvalAI, an open source platform for evaluating and comparing machine learning (ML) and artificial intelligence algorithms (AI) at scale. EvalAI is built to provide a scalable solution to the research community to fulfill the critical need of evaluating machine learning models and agents acting in an environment against annotations or with a human-in-the-loop. This will help researchers, students, and data scientists to create, collaborate, and participate in AI challenges organized around the globe. By simplifying and standardizing the process of benchmarking these models, EvalAI seeks to lower the barrier to entry for participating in the global scientific effort to push the frontiers of machine learning and artificial intelligence, thereby increasing the rate of measurable progress in this domain.

Ramprasaath R. Selvaraju, Prithvijit Chattopadhyay, Mohamed Elhoseiny et al.

Individual neurons in convolutional neural networks supervised for image-level classification tasks have been shown to implicitly learn semantically meaningful concepts ranging from simple textures and shapes to whole or partial objects - forming a "dictionary" of concepts acquired through the learning process. In this work we introduce a simple, efficient zero-shot learning approach based on this observation. Our approach, which we call Neuron Importance-AwareWeight Transfer (NIWT), learns to map domain knowledge about novel "unseen" classes onto this dictionary of learned concepts and then optimizes for network parameters that can effectively combine these concepts - essentially learning classifiers by discovering and composing learned semantic concepts in deep networks. Our approach shows improvements over previous approaches on the CUBirds and AWA2 generalized zero-shot learning benchmarks. We demonstrate our approach on a diverse set of semantic inputs as external domain knowledge including attributes and natural language captions. Moreover by learning inverse mappings, NIWT can provide visual and textual explanations for the predictions made by the newly learned classifiers and provide neuron names. Our code is available at https://github.com/ramprs/neuron-importance-zsl.

Skand Peri, Iain Lee, Chanho Kim et al. · nvidia

Visual control policies can encounter significant performance degradation when visual conditions like lighting or camera position differ from those seen during training -- often exhibiting sharp declines in capability even for minor differences. In this work, we examine robustness to a suite of these types of visual changes for RGB-D and point cloud based visual control policies. To perform these experiments on both model-free and model-based reinforcement learners, we introduce a novel Point Cloud World Model (PCWM) and point cloud based control policies. Our experiments show that policies that explicitly encode point clouds are significantly more robust than their RGB-D counterparts. Further, we find our proposed PCWM significantly outperforms prior works in terms of sample efficiency during training. Taken together, these results suggest reasoning about the 3D scene through point clouds can improve performance, reduce learning time, and increase robustness for robotic learners. Project Webpage: https://pvskand.github.io/projects/PCWM

Eric Slyman, Stefan Lee, Scott Cohen et al.

Recent dataset deduplication techniques have demonstrated that content-aware dataset pruning can dramatically reduce the cost of training Vision-Language Pretrained (VLP) models without significant performance losses compared to training on the original dataset. These results have been based on pruning commonly used image-caption datasets collected from the web -- datasets that are known to harbor harmful social biases that may then be codified in trained models. In this work, we evaluate how deduplication affects the prevalence of these biases in the resulting trained models and introduce an easy-to-implement modification to the recent SemDeDup algorithm that can reduce the negative effects that we observe. When examining CLIP-style models trained on deduplicated variants of LAION-400M, we find our proposed FairDeDup algorithm consistently leads to improved fairness metrics over SemDeDup on the FairFace and FACET datasets while maintaining zero-shot performance on CLIP benchmarks.

Zijiao Yang, Xiangxi Shi, Eric Slyman et al.

Assistive embodied agents that can be instructed in natural language to perform tasks in open-world environments have the potential to significantly impact labor tasks like manufacturing or in-home care -- benefiting the lives of those who come to depend on them. In this work, we consider how this benefit might be hijacked by local modifications in the appearance of the agent's operating environment. Specifically, we take the popular Vision-and-Language Navigation (VLN) task as a representative setting and develop a whitebox adversarial attack that optimizes a 3D attack object's appearance to induce desired behaviors in pretrained VLN agents that observe it in the environment. We demonstrate that the proposed attack can cause VLN agents to ignore their instructions and execute alternative actions after encountering the attack object -- even for instructions and agent paths not considered when optimizing the attack. For these novel settings, we find our attacks can induce early-termination behaviors or divert an agent along an attacker-defined multi-step trajectory. Under both conditions, environmental attacks significantly reduce agent capabilities to successfully follow user instructions.

Eric Slyman, Anirudh Kanneganti, Sanghyun Hong et al.

We study the impact of a standard practice in compressing foundation vision-language models - quantization - on the models' ability to produce socially-fair outputs. In contrast to prior findings with unimodal models that compression consistently amplifies social biases, our extensive evaluation of four quantization settings across three datasets and three CLIP variants yields a surprising result: while individual models demonstrate bias, we find no consistent change in bias magnitude or direction across a population of compressed models due to quantization.

Rajesh Mangannavar, Stefan Lee, Alan Fern et al.

We propose a novel approach to learn relational policies for classical planning based on learning to rank actions. We introduce a new graph representation that explicitly captures action information and propose a Graph Neural Network (GNN) architecture augmented with Gated Recurrent Units (GRUs) to learn action rankings. Unlike value-function based approaches that must learn a globally consistent function, our action ranking method only needs to learn locally consistent ranking. Our model is trained on data generated from small problem instances that are easily solved by planners and is applied to significantly larger instances where planning is computationally prohibitive. Experimental results across standard planning benchmarks demonstrate that our action-ranking approach not only achieves better generalization to larger problems than those used in training but also outperforms multiple baselines (value function and action ranking) methods in terms of success rate and plan quality.

Anh Nguyen, Stefan Lee

To interact effectively with humans in the real world, it is important for agents to understand language that describes the dynamics of the environment--that is, how the environment behaves--rather than just task instructions specifying "what to do". Understanding this dynamics-descriptive language is important for human-agent interaction and agent behavior. Recent work address this problem using a model-based approach: language is incorporated into a world model, which is then used to learn a behavior policy. However, these existing methods either do not demonstrate policy generalization to unseen games or rely on limiting assumptions. For instance, assuming that the latency induced by inference-time planning is tolerable for the target task or expert demonstrations are available. Expanding on this line of research, we focus on improving policy generalization from a language-conditioned world model while dropping these assumptions. We propose a model-based reinforcement learning approach, where a language-conditioned world model is trained through interaction with the environment, and a policy is learned from this model--without planning or expert demonstrations. Our method proposes Language-aware Encoder for Dreamer World Model (LED-WM) built on top of DreamerV3. LED-WM features an observation encoder that uses an attention mechanism to explicitly ground language descriptions to entities in the observation. We show that policies trained with LED-WM generalize more effectively to unseen games described by novel dynamics and language compared to other baselines in several settings in two environments: MESSENGER and MESSENGER-WM.To highlight how the policy can leverage the trained world model before real-world deployment, we demonstrate the policy can be improved through fine-tuning on synthetic test trajectories generated by the world model.

Eric Slyman, Mehrab Tanjim, Kushal Kafle et al.

Multimodal large language models (MLLMs) are increasingly used to evaluate text-to-image (TTI) generation systems, providing automated judgments based on visual and textual context. However, these "judge" models often suffer from biases, overconfidence, and inconsistent performance across diverse image domains. While prompt ensembling has shown promise for mitigating these issues in unimodal, text-only settings, our experiments reveal that standard ensembling methods fail to generalize effectively for TTI tasks. To address these limitations, we propose a new multimodal-aware method called Multimodal Mixture-of-Bayesian Prompt Ensembles (MMB). Our method uses a Bayesian prompt ensemble approach augmented by image clustering, allowing the judge to dynamically assign prompt weights based on the visual characteristics of each sample. We show that MMB improves accuracy in pairwise preference judgments and greatly enhances calibration, making it easier to gauge the judge's true uncertainty. In evaluations on two TTI benchmarks, HPSv2 and MJBench, MMB outperforms existing baselines in alignment with human annotations and calibration across varied image content. Our findings highlight the importance of multimodal-specific strategies for judge calibration and suggest a promising path forward for reliable large-scale TTI evaluation.

Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal et al.

Natural language instructions for visual navigation often use scene descriptions (e.g., "bedroom") and object references (e.g., "green chairs") to provide a breadcrumb trail to a goal location. This work presents a transformer-based vision-and-language navigation (VLN) agent that uses two different visual encoders -- a scene classification network and an object detector -- which produce features that match these two distinct types of visual cues. In our method, scene features contribute high-level contextual information that supports object-level processing. With this design, our model is able to use vision-and-language pretraining (i.e., learning the alignment between images and text from large-scale web data) to substantially improve performance on the Room-to-Room (R2R) and Room-Across-Room (RxR) benchmarks. Specifically, our approach leads to improvements of 1.8% absolute in SPL on R2R and 3.7% absolute in SR on RxR. Our analysis reveals even larger gains for navigation instructions that contain six or more object references, which further suggests that our approach is better able to use object features and align them to references in the instructions.

Jacob Krantz, Aaron Gokaslan, Dhruv Batra et al.

Little inquiry has explicitly addressed the role of action spaces in language-guided visual navigation -- either in terms of its effect on navigation success or the efficiency with which a robotic agent could execute the resulting trajectory. Building on the recently released VLN-CE setting for instruction following in continuous environments, we develop a class of language-conditioned waypoint prediction networks to examine this question. We vary the expressivity of these models to explore a spectrum between low-level actions and continuous waypoint prediction. We measure task performance and estimated execution time on a profiled LoCoBot robot. We find more expressive models result in simpler, faster to execute trajectories, but lower-level actions can achieve better navigation metrics by approximating shortest paths better. Further, our models outperform prior work in VLN-CE and set a new state-of-the-art on the public leaderboard -- increasing success rate by 4% with our best model on this challenging task.

Yilin Yang, Akiko Eriguchi, Alexandre Muzio et al.

Multilingual Neural Machine Translation (NMT) enables one model to serve all translation directions, including ones that are unseen during training, i.e. zero-shot translation. Despite being theoretically attractive, current models often produce low quality translations -- commonly failing to even produce outputs in the right target language. In this work, we observe that off-target translation is dominant even in strong multilingual systems, trained on massive multilingual corpora. To address this issue, we propose a joint approach to regularize NMT models at both representation-level and gradient-level. At the representation level, we leverage an auxiliary target language prediction task to regularize decoder outputs to retain information about the target language. At the gradient level, we leverage a small amount of direct data (in thousands of sentence pairs) to regularize model gradients. Our results demonstrate that our approach is highly effective in both reducing off-target translation occurrences and improving zero-shot translation performance by +5.59 and +10.38 BLEU on WMT and OPUS datasets respectively. Moreover, experiments show that our method also works well when the small amount of direct data is not available.

Sam Greydanus, Stefan Lee, Alan Fern

Neural networks are a popular tool for modeling sequential data but they generally do not treat time as a continuous variable. Neural ODEs represent an important exception: they parameterize the time derivative of a hidden state with a neural network and then integrate over arbitrary amounts of time. But these parameterizations, which have arbitrary curvature, can be hard to integrate and thus train and evaluate. In this paper, we propose making a piecewise-constant approximation to Neural ODEs to mitigate these issues. Our model can be integrated exactly via Euler integration and can generate autoregressive samples in 3-20 times fewer steps than comparable RNN and ODE-RNN models. We evaluate our model on several synthetic physics tasks and a planning task inspired by the game of billiards. We find that it matches the performance of baseline approaches while requiring less time to train and evaluate.

Erich Merrill, Stefan Lee, Li Fuxin et al.

We consider the problem of modeling the dynamics of continuous spatial-temporal processes represented by irregular samples through both space and time. Such processes occur in sensor networks, citizen science, multi-robot systems, and many others. We propose a new deep model that is able to directly learn and predict over this irregularly sampled data, without voxelization, by leveraging a recent convolutional architecture for static point clouds. The model also easily incorporates the notion of multiple entities in the process. In particular, the model can flexibly answer prediction queries about arbitrary space-time points for different entities regardless of the distribution of the training or test-time data. We present experiments on real-world weather station data and battles between large armies in StarCraft II. The results demonstrate the model's flexibility in answering a variety of query types and demonstrate improved performance and efficiency compared to state-of-the-art baselines.

Meera Hahn, Jacob Krantz, Dhruv Batra et al.

We present Where Are You? (WAY), a dataset of ~6k dialogs in which two humans -- an Observer and a Locator -- complete a cooperative localization task. The Observer is spawned at random in a 3D environment and can navigate from first-person views while answering questions from the Locator. The Locator must localize the Observer in a detailed top-down map by asking questions and giving instructions. Based on this dataset, we define three challenging tasks: Localization from Embodied Dialog or LED (localizing the Observer from dialog history), Embodied Visual Dialog (modeling the Observer), and Cooperative Localization (modeling both agents). In this paper, we focus on the LED task -- providing a strong baseline model with detailed ablations characterizing both dataset biases and the importance of various modeling choices. Our best model achieves 32.7% success at identifying the Observer's location within 3m in unseen buildings, vs. 70.4% for human Locators.

Peter Anderson, Ayush Shrivastava, Joanne Truong et al.

We study the challenging problem of releasing a robot in a previously unseen environment, and having it follow unconstrained natural language navigation instructions. Recent work on the task of Vision-and-Language Navigation (VLN) has achieved significant progress in simulation. To assess the implications of this work for robotics, we transfer a VLN agent trained in simulation to a physical robot. To bridge the gap between the high-level discrete action space learned by the VLN agent, and the robot's low-level continuous action space, we propose a subgoal model to identify nearby waypoints, and use domain randomization to mitigate visual domain differences. For accurate sim and real comparisons in parallel environments, we annotate a 325m2 office space with 1.3km of navigation instructions, and create a digitized replica in simulation. We find that sim-to-real transfer to an environment not seen in training is successful if an occupancy map and navigation graph can be collected and annotated in advance (success rate of 46.8% vs. 55.9% in sim), but much more challenging in the hardest setting with no prior mapping at all (success rate of 22.5%).

Simon Stepputtis, Joseph Campbell, Mariano Phielipp et al.

Imitation learning is a popular approach for teaching motor skills to robots. However, most approaches focus on extracting policy parameters from execution traces alone (i.e., motion trajectories and perceptual data). No adequate communication channel exists between the human expert and the robot to describe critical aspects of the task, such as the properties of the target object or the intended shape of the motion. Motivated by insights into the human teaching process, we introduce a method for incorporating unstructured natural language into imitation learning. At training time, the expert can provide demonstrations along with verbal descriptions in order to describe the underlying intent (e.g., "go to the large green bowl"). The training process then interrelates these two modalities to encode the correlations between language, perception, and motion. The resulting language-conditioned visuomotor policies can be conditioned at runtime on new human commands and instructions, which allows for more fine-grained control over the trained policies while also reducing situational ambiguity. We demonstrate in a set of simulation experiments how our approach can learn language-conditioned manipulation policies for a seven-degree-of-freedom robot arm and compare the results to a variety of alternative methods.

Aayam Shrestha, Stefan Lee, Prasad Tadepalli et al.

We study an approach to offline reinforcement learning (RL) based on optimally solving finitely-represented MDPs derived from a static dataset of experience. This approach can be applied on top of any learned representation and has the potential to easily support multiple solution objectives as well as zero-shot adjustment to changing environments and goals. Our main contribution is to introduce the Deep Averagers with Costs MDP (DAC-MDP) and to investigate its solutions for offline RL. DAC-MDPs are a non-parametric model that can leverage deep representations and account for limited data by introducing costs for exploiting under-represented parts of the model. In theory, we show conditions that allow for lower-bounding the performance of DAC-MDP solutions. We also investigate the empirical behavior in a number of environments, including those with image-based observations. Overall, the experiments demonstrate that the framework can work in practice and scale to large complex offline RL problems.

Yilin Yang, Longyue Wang, Shuming Shi et al.

There have been significant efforts to interpret the encoder of Transformer-based encoder-decoder architectures for neural machine translation (NMT); meanwhile, the decoder remains largely unexamined despite its critical role. During translation, the decoder must predict output tokens by considering both the source-language text from the encoder and the target-language prefix produced in previous steps. In this work, we study how Transformer-based decoders leverage information from the source and target languages -- developing a universal probe task to assess how information is propagated through each module of each decoder layer. We perform extensive experiments on three major translation datasets (WMT En-De, En-Fr, and En-Zh). Our analysis provides insight on when and where decoders leverage different sources. Based on these insights, we demonstrate that the residual feed-forward module in each Transformer decoder layer can be dropped with minimal loss of performance -- a significant reduction in computation and number of parameters, and consequently a significant boost to both training and inference speed.

Vincent Cartillier, Zhile Ren, Neha Jain et al.

We study the task of semantic mapping - specifically, an embodied agent (a robot or an egocentric AI assistant) is given a tour of a new environment and asked to build an allocentric top-down semantic map ("what is where?") from egocentric observations of an RGB-D camera with known pose (via localization sensors). Towards this goal, we present SemanticMapNet (SMNet), which consists of: (1) an Egocentric Visual Encoder that encodes each egocentric RGB-D frame, (2) a Feature Projector that projects egocentric features to appropriate locations on a floor-plan, (3) a Spatial Memory Tensor of size floor-plan length x width x feature-dims that learns to accumulate projected egocentric features, and (4) a Map Decoder that uses the memory tensor to produce semantic top-down maps. SMNet combines the strengths of (known) projective camera geometry and neural representation learning. On the task of semantic mapping in the Matterport3D dataset, SMNet significantly outperforms competitive baselines by 4.01-16.81% (absolute) on mean-IoU and 3.81-19.69% (absolute) on Boundary-F1 metrics. Moreover, we show how to use the neural episodic memories and spatio-semantic allocentric representations build by SMNet for subsequent tasks in the same space - navigating to objects seen during the tour("Find chair") or answering questions about the space ("How many chairs did you see in the house?"). Project page: https://vincentcartillier.github.io/smnet.html.

Samyak Datta, Oleksandr Maksymets, Judy Hoffman et al.

Recent work has presented embodied agents that can navigate to point-goal targets in novel indoor environments with near-perfect accuracy. However, these agents are equipped with idealized sensors for localization and take deterministic actions. This setting is practically sterile by comparison to the dirty reality of noisy sensors and actuations in the real world -- wheels can slip, motion sensors have error, actuations can rebound. In this work, we take a step towards this noisy reality, developing point-goal navigation agents that rely on visual estimates of egomotion under noisy action dynamics. We find these agents outperform naive adaptions of current point-goal agents to this setting as well as those incorporating classic localization baselines. Further, our model conceptually divides learning agent dynamics or odometry (where am I?) from task-specific navigation policy (where do I want to go?). This enables a seamless adaption to changing dynamics (a different robot or floor type) by simply re-calibrating the visual odometry model -- circumventing the expense of re-training of the navigation policy. Our agent was the runner-up in the PointNav track of CVPR 2020 Habitat Challenge.

Arjun Majumdar, Ayush Shrivastava, Stefan Lee et al.

Following a navigation instruction such as 'Walk down the stairs and stop at the brown sofa' requires embodied AI agents to ground scene elements referenced via language (e.g. 'stairs') to visual content in the environment (pixels corresponding to 'stairs'). We ask the following question -- can we leverage abundant 'disembodied' web-scraped vision-and-language corpora (e.g. Conceptual Captions) to learn visual groundings (what do 'stairs' look like?) that improve performance on a relatively data-starved embodied perception task (Vision-and-Language Navigation)? Specifically, we develop VLN-BERT, a visiolinguistic transformer-based model for scoring the compatibility between an instruction ('...stop at the brown sofa') and a sequence of panoramic RGB images captured by the agent. We demonstrate that pretraining VLN-BERT on image-text pairs from the web before fine-tuning on embodied path-instruction data significantly improves performance on VLN -- outperforming the prior state-of-the-art in the fully-observed setting by 4 absolute percentage points on success rate. Ablations of our pretraining curriculum show each stage to be impactful -- with their combination resulting in further positive synergistic effects.

Jacob Krantz, Erik Wijmans, Arjun Majumdar et al.

We develop a language-guided navigation task set in a continuous 3D environment where agents must execute low-level actions to follow natural language navigation directions. By being situated in continuous environments, this setting lifts a number of assumptions implicit in prior work that represents environments as a sparse graph of panoramas with edges corresponding to navigability. Specifically, our setting drops the presumptions of known environment topologies, short-range oracle navigation, and perfect agent localization. To contextualize this new task, we develop models that mirror many of the advances made in prior settings as well as single-modality baselines. While some of these techniques transfer, we find significantly lower absolute performance in the continuous setting -- suggesting that performance in prior `navigation-graph' settings may be inflated by the strong implicit assumptions.

Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach et al.

Much of vision-and-language research focuses on a small but diverse set of independent tasks and supporting datasets often studied in isolation; however, the visually-grounded language understanding skills required for success at these tasks overlap significantly. In this work, we investigate these relationships between vision-and-language tasks by developing a large-scale, multi-task training regime. Our approach culminates in a single model on 12 datasets from four broad categories of task including visual question answering, caption-based image retrieval, grounding referring expressions, and multi-modal verification. Compared to independently trained single-task models, this represents a reduction from approximately 3 billion parameters to 270 million while simultaneously improving performance by 2.05 points on average across tasks. We use our multi-task framework to perform in-depth analysis of the effect of joint training diverse tasks. Further, we show that finetuning task-specific models from our single multi-task model can lead to further improvements, achieving performance at or above the state-of-the-art.

Jingjing Pan, Yash Goyal, Stefan Lee

While Visual Question Answering (VQA) models continue to push the state-of-the-art forward, they largely remain black-boxes - failing to provide insight into how or why an answer is generated. In this ongoing work, we propose addressing this shortcoming by learning to generate counterfactual images for a VQA model - i.e. given a question-image pair, we wish to generate a new image such that i) the VQA model outputs a different answer, ii) the new image is minimally different from the original, and iii) the new image is realistic. Our hope is that providing such counterfactual examples allows users to investigate and understand the VQA model's internal mechanisms.

Erik Wijmans, Abhishek Kadian, Ari Morcos et al.

We present Decentralized Distributed Proximal Policy Optimization (DD-PPO), a method for distributed reinforcement learning in resource-intensive simulated environments. DD-PPO is distributed (uses multiple machines), decentralized (lacks a centralized server), and synchronous (no computation is ever stale), making it conceptually simple and easy to implement. In our experiments on training virtual robots to navigate in Habitat-Sim, DD-PPO exhibits near-linear scaling -- achieving a speedup of 107x on 128 GPUs over a serial implementation. We leverage this scaling to train an agent for 2.5 Billion steps of experience (the equivalent of 80 years of human experience) -- over 6 months of GPU-time training in under 3 days of wall-clock time with 64 GPUs. This massive-scale training not only sets the state of art on Habitat Autonomous Navigation Challenge 2019, but essentially solves the task --near-perfect autonomous navigation in an unseen environment without access to a map, directly from an RGB-D camera and a GPS+Compass sensor. Fortuitously, error vs computation exhibits a power-law-like distribution; thus, 90% of peak performance is obtained relatively early (at 100 million steps) and relatively cheaply (under 1 day with 8 GPUs). Finally, we show that the scene understanding and navigation policies learned can be transferred to other navigation tasks -- the analog of ImageNet pre-training + task-specific fine-tuning for embodied AI. Our model outperforms ImageNet pre-trained CNNs on these transfer tasks and can serve as a universal resource (all models and code are publicly available).

Arijit Ray, Karan Sikka, Ajay Divakaran et al.

While models for Visual Question Answering (VQA) have steadily improved over the years, interacting with one quickly reveals that these models lack consistency. For instance, if a model answers "red" to "What color is the balloon?", it might answer "no" if asked, "Is the balloon red?". These responses violate simple notions of entailment and raise questions about how effectively VQA models ground language. In this work, we introduce a dataset, ConVQA, and metrics that enable quantitative evaluation of consistency in VQA. For a given observable fact in an image (e.g. the balloon's color), we generate a set of logically consistent question-answer (QA) pairs (e.g. Is the balloon red?) and also collect a human-annotated set of common-sense based consistent QA pairs (e.g. Is the balloon the same color as tomato sauce?). Further, we propose a consistency-improving data augmentation module, a Consistency Teacher Module (CTM). CTM automatically generates entailed (or similar-intent) questions for a source QA pair and fine-tunes the VQA model if the VQA's answer to the entailed question is consistent with the source QA pair. We demonstrate that our CTM-based training improves the consistency of VQA models on the ConVQA datasets and is a strong baseline for further research.

Jiasen Lu, Dhruv Batra, Devi Parikh et al.

We present ViLBERT (short for Vision-and-Language BERT), a model for learning task-agnostic joint representations of image content and natural language. We extend the popular BERT architecture to a multi-modal two-stream model, pro-cessing both visual and textual inputs in separate streams that interact through co-attentional transformer layers. We pretrain our model through two proxy tasks on the large, automatically collected Conceptual Captions dataset and then transfer it to multiple established vision-and-language tasks -- visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval -- by making only minor additions to the base architecture. We observe significant improvements across tasks compared to existing task-specific models -- achieving state-of-the-art on all four tasks. Our work represents a shift away from learning groundings between vision and language only as part of task training and towards treating visual grounding as a pretrainable and transferable capability.

Peter Anderson, Ayush Shrivastava, Devi Parikh et al.

A visually-grounded navigation instruction can be interpreted as a sequence of expected observations and actions an agent following the correct trajectory would encounter and perform. Based on this intuition, we formulate the problem of finding the goal location in Vision-and-Language Navigation (VLN) within the framework of Bayesian state tracking - learning observation and motion models conditioned on these expectable events. Together with a mapper that constructs a semantic spatial map on-the-fly during navigation, we formulate an end-to-end differentiable Bayes filter and train it to identify the goal by predicting the most likely trajectory through the map according to the instructions. The resulting navigation policy constitutes a new approach to instruction following that explicitly models a probability distribution over states, encoding strong geometric and algorithmic priors while enabling greater explainability. Our experiments show that our approach outperforms a strong LingUNet baseline when predicting the goal location on the map. On the full VLN task, i.e. navigating to the goal location, our approach achieves promising results with less reliance on navigation constraints.

Michael Cogswell, Jiasen Lu, Stefan Lee et al.

Recent work has studied the emergence of language among deep reinforcement learning agents that must collaborate to solve a task. Of particular interest are the factors that cause language to be compositional -- i.e., express meaning by combining words which themselves have meaning. Evolutionary linguists have found that in addition to structural priors like those already studied in deep learning, the dynamics of transmitting language from generation to generation contribute significantly to the emergence of compositionality. In this paper, we introduce these cultural evolutionary dynamics into language emergence by periodically replacing agents in a population to create a knowledge gap, implicitly inducing cultural transmission of language. We show that this implicit cultural transmission encourages the resulting languages to exhibit better compositional generalization.

Yash Goyal, Ziyan Wu, Jan Ernst et al.

In this work, we develop a technique to produce counterfactual visual explanations. Given a 'query' image $I$ for which a vision system predicts class $c$, a counterfactual visual explanation identifies how $I$ could change such that the system would output a different specified class $c'$. To do this, we select a 'distractor' image $I'$ that the system predicts as class $c'$ and identify spatial regions in $I$ and $I'$ such that replacing the identified region in $I$ with the identified region in $I'$ would push the system towards classifying $I$ as $c'$. We apply our approach to multiple image classification datasets generating qualitative results showcasing the interpretability and discriminativeness of our counterfactual explanations. To explore the effectiveness of our explanations in teaching humans, we present machine teaching experiments for the task of fine-grained bird classification. We find that users trained to distinguish bird species fare better when given access to counterfactual explanations in addition to training examples.

Erik Wijmans, Samyak Datta, Oleksandr Maksymets et al.

To help bridge the gap between internet vision-style problems and the goal of vision for embodied perception we instantiate a large-scale navigation task -- Embodied Question Answering [1] in photo-realistic environments (Matterport 3D). We thoroughly study navigation policies that utilize 3D point clouds, RGB images, or their combination. Our analysis of these models reveals several key findings. We find that two seemingly naive navigation baselines, forward-only and random, are strong navigators and challenging to outperform, due to the specific choice of the evaluation setting presented by [1]. We find a novel loss-weighting scheme we call Inflection Weighting to be important when training recurrent models for navigation with behavior cloning and are able to out perform the baselines with this technique. We find that point clouds provide a richer signal than RGB images for learning obstacle avoidance, motivating the use (and continued study) of 3D deep learning models for embodied navigation.

Ramakrishna Vedantam, Karan Desai, Stefan Lee et al.

We propose a new class of probabilistic neural-symbolic models, that have symbolic functional programs as a latent, stochastic variable. Instantiated in the context of visual question answering, our probabilistic formulation offers two key conceptual advantages over prior neural-symbolic models for VQA. Firstly, the programs generated by our model are more understandable while requiring lesser number of teaching examples. Secondly, we show that one can pose counterfactual scenarios to the model, to probe its beliefs on the programs that could lead to a specified answer given an image. Our results on the CLEVR and SHAPES datasets verify our hypotheses, showing that the model gets better program (and answer) prediction accuracy even in the low data regime, and allows one to probe the coherence and consistency of reasoning performed.

Ramprasaath R. Selvaraju, Stefan Lee, Yilin Shen et al.

Many vision and language models suffer from poor visual grounding - often falling back on easy-to-learn language priors rather than basing their decisions on visual concepts in the image. In this work, we propose a generic approach called Human Importance-aware Network Tuning (HINT) that effectively leverages human demonstrations to improve visual grounding. HINT encourages deep networks to be sensitive to the same input regions as humans. Our approach optimizes the alignment between human attention maps and gradient-based network importances - ensuring that models learn not just to look at but rather rely on visual concepts that humans found relevant for a task when making predictions. We apply HINT to Visual Question Answering and Image Captioning tasks, outperforming top approaches on splits that penalize over-reliance on language priors (VQA-CP and robust captioning) using human attention demonstrations for just 6% of the training data.

Huda Alamri, Vincent Cartillier, Abhishek Das et al.

We introduce the task of scene-aware dialog. Our goal is to generate a complete and natural response to a question about a scene, given video and audio of the scene and the history of previous turns in the dialog. To answer successfully, agents must ground concepts from the question in the video while leveraging contextual cues from the dialog history. To benchmark this task, we introduce the Audio Visual Scene-Aware Dialog (AVSD) Dataset. For each of more than 11,000 videos of human actions from the Charades dataset, our dataset contains a dialog about the video, plus a final summary of the video by one of the dialog participants. We train several baseline systems for this task and evaluate the performance of the trained models using both qualitative and quantitative metrics. Our results indicate that models must utilize all the available inputs (video, audio, question, and dialog history) to perform best on this dataset.

Abhishek Das, Georgia Gkioxari, Stefan Lee et al.

We present a modular approach for learning policies for navigation over long planning horizons from language input. Our hierarchical policy operates at multiple timescales, where the higher-level master policy proposes subgoals to be executed by specialized sub-policies. Our choice of subgoals is compositional and semantic, i.e. they can be sequentially combined in arbitrary orderings, and assume human-interpretable descriptions (e.g. 'exit room', 'find kitchen', 'find refrigerator', etc.). We use imitation learning to warm-start policies at each level of the hierarchy, dramatically increasing sample efficiency, followed by reinforcement learning. Independent reinforcement learning at each level of hierarchy enables sub-policies to adapt to consequences of their actions and recover from errors. Subsequent joint hierarchical training enables the master policy to adapt to the sub-policies. On the challenging EQA (Das et al., 2018) benchmark in House3D (Wu et al., 2018), requiring navigating diverse realistic indoor environments, our approach outperforms prior work by a significant margin, both in terms of navigation and question answering.

Sainandan Ramakrishnan, Aishwarya Agrawal, Stefan Lee

Modern Visual Question Answering (VQA) models have been shown to rely heavily on superficial correlations between question and answer words learned during training such as overwhelmingly reporting the type of room as kitchen or the sport being played as tennis, irrespective of the image. Most alarmingly, this shortcoming is often not well reflected during evaluation because the same strong priors exist in test distributions; however, a VQA system that fails to ground questions in image content would likely perform poorly in real-world settings. In this work, we present a novel regularization scheme for VQA that reduces this effect. We introduce a question-only model that takes as input the question encoding from the VQA model and must leverage language biases in order to succeed. We then pose training as an adversarial game between the VQA model and this question-only adversary -- discouraging the VQA model from capturing language biases in its question encoding. Further,we leverage this question-only model to estimate the increase in model confidence after considering the image, which we maximize explicitly to encourage visual grounding. Our approach is a model agnostic training procedure and simple to implement. We show empirically that it can improve performance significantly on a bias-sensitive split of the VQA dataset for multiple base models -- achieving state-of-the-art on this task. Further, on standard VQA tasks, our approach shows significantly less drop in accuracy compared to existing bias-reducing VQA models.

Jianwei Yang, Jiasen Lu, Stefan Lee et al.

In an open-world setting, it is inevitable that an intelligent agent (e.g., a robot) will encounter visual objects, attributes or relationships it does not recognize. In this work, we develop an agent empowered with visual curiosity, i.e. the ability to ask questions to an Oracle (e.g., human) about the contents in images (e.g., What is the object on the left side of the red cube?) and build visual recognition model based on the answers received (e.g., Cylinder). In order to do this, the agent must (1) understand what it recognizes and what it does not, (2) formulate a valid, unambiguous and informative language query (a question) to ask the Oracle, (3) derive the parameters of visual classifiers from the Oracle response and (4) leverage the updated visual classifiers to ask more clarified questions. Specifically, we propose a novel framework and formulate the learning of visual curiosity as a reinforcement learning problem. In this framework, all components of our agent, visual recognition module (to see), question generation policy (to ask), answer digestion module (to understand) and graph memory module (to memorize), are learned entirely end-to-end to maximize the reward derived from the scene graph obtained by the agent as a consequence of the dialog with the Oracle. Importantly, the question generation policy is disentangled from the visual recognition system and specifics of the environment. Consequently, we demonstrate a sort of double generalization. Our question generation policy generalizes to new environments and a new pair of eyes, i.e., new visual system. Trained on a synthetic dataset, our results show that our agent learns new visual concepts significantly faster than several heuristic baselines, even when tested on synthetic environments with novel objects, as well as in a realistic environment.

Jianwei Yang, Jiasen Lu, Stefan Lee et al.

We propose a novel scene graph generation model called Graph R-CNN, that is both effective and efficient at detecting objects and their relations in images. Our model contains a Relation Proposal Network (RePN) that efficiently deals with the quadratic number of potential relations between objects in an image. We also propose an attentional Graph Convolutional Network (aGCN) that effectively captures contextual information between objects and relations. Finally, we introduce a new evaluation metric that is more holistic and realistic than existing metrics. We report state-of-the-art performance on scene graph generation as evaluated using both existing and our proposed metrics.

Ashwin Kalyan, Stefan Lee, Anitha Kannan et al.

Many structured prediction problems (particularly in vision and language domains) are ambiguous, with multiple outputs being correct for an input - e.g. there are many ways of describing an image, multiple ways of translating a sentence; however, exhaustively annotating the applicability of all possible outputs is intractable due to exponentially large output spaces (e.g. all English sentences). In practice, these problems are cast as multi-class prediction, with the likelihood of only a sparse set of annotations being maximized - unfortunately penalizing for placing beliefs on plausible but unannotated outputs. We make and test the following hypothesis - for a given input, the annotations of its neighbors may serve as an additional supervisory signal. Specifically, we propose an objective that transfers supervision from neighboring examples. We first study the properties of our developed method in a controlled toy setup before reporting results on multi-label classification and two image-grounded sequence modeling tasks - captioning and question generation. We evaluate using standard task-specific metrics and measures of output diversity, finding consistent improvements over standard maximum likelihood training and other baselines.

Abhishek Das, Samyak Datta, Georgia Gkioxari et al.

We present a new AI task -- Embodied Question Answering (EmbodiedQA) -- where an agent is spawned at a random location in a 3D environment and asked a question ("What color is the car?"). In order to answer, the agent must first intelligently navigate to explore the environment, gather information through first-person (egocentric) vision, and then answer the question ("orange"). This challenging task requires a range of AI skills -- active perception, language understanding, goal-driven navigation, commonsense reasoning, and grounding of language into actions. In this work, we develop the environments, end-to-end-trained reinforcement learning agents, and evaluation protocols for EmbodiedQA.

Prithvijit Chattopadhyay, Deshraj Yadav, Viraj Prabhu et al.

As AI continues to advance, human-AI teams are inevitable. However, progress in AI is routinely measured in isolation, without a human in the loop. It is crucial to benchmark progress in AI, not just in isolation, but also in terms of how it translates to helping humans perform certain tasks, i.e., the performance of human-AI teams. In this work, we design a cooperative game - GuessWhich - to measure human-AI team performance in the specific context of the AI being a visual conversational agent. GuessWhich involves live interaction between the human and the AI. The AI, which we call ALICE, is provided an image which is unseen by the human. Following a brief description of the image, the human questions ALICE about this secret image to identify it from a fixed pool of images. We measure performance of the human-ALICE team by the number of guesses it takes the human to correctly identify the secret image after a fixed number of dialog rounds with ALICE. We compare performance of the human-ALICE teams for two versions of ALICE. Our human studies suggest a counterintuitive trend - that while AI literature shows that one version outperforms the other when paired with an AI questioner bot, we find that this improvement in AI-AI performance does not translate to improved human-AI performance. This suggests a mismatch between benchmarking of AI in isolation and in the context of human-AI teams.

Satwik Kottur, José M. F. Moura, Stefan Lee et al.

A number of recent works have proposed techniques for end-to-end learning of communication protocols among cooperative multi-agent populations, and have simultaneously found the emergence of grounded human-interpretable language in the protocols developed by the agents, all learned without any human supervision! In this paper, using a Task and Tell reference game between two agents as a testbed, we present a sequence of 'negative' results culminating in a 'positive' one -- showing that while most agent-invented languages are effective (i.e. achieve near-perfect task rewards), they are decidedly not interpretable or compositional. In essence, we find that natural language does not emerge 'naturally', despite the semblance of ease of natural-language-emergence that one may gather from recent literature. We discuss how it is possible to coax the invented languages to become more and more human-like and compositional by increasing restrictions on how two agents may communicate.