MAD-TD: Model-Augmented Data stabilizes High Update Ratio RL

Building deep reinforcement learning (RL) agents that find a good policy with few samples has proven notoriously challenging. To achieve sample efficiency, recent work has explored updating neural networks with large numbers of gradient steps for every new sample. While such high update-to-data (UTD) ratios have shown strong empirical performance, they also introduce instability to the training process. Previous approaches need to rely on periodic neural network parameter resets to address this instability, but restarting the training process is infeasible in many real-world applications and requires tuning the resetting interval. In this paper, we focus on one of the core difficulties of stable training with limited samples: the inability of learned value functions to generalize to unobserved on-policy actions. We mitigate this issue directly by augmenting the off-policy RL training process with a small amount of data generated from a learned world model. Our method, Model-Augmented Data for Temporal Difference learning (MAD-TD) uses small amounts of generated data to stabilize high UTD training and achieve competitive performance on the most challenging tasks in the DeepMind control suite. Our experiments further highlight the importance of employing a good model to generate data, MAD-TD's ability to combat value overestimation, and its practical stability gains for continued learning.

Scene Flow as a Partial Differential Equation

We reframe scene flow as the problem of estimating a continuous space and time PDE that describes motion for an entire observation sequence, represented with a neural prior. Our resulting unsupervised method, EulerFlow, produces high quality scene flow on real-world data across multiple domains, including large-scale autonomous driving scenes and dynamic tabletop settings. Notably, EulerFlow produces high quality flow on small, fast moving objects like birds and tennis balls, and exhibits emergent 3D point tracking behavior by solving its estimated PDE over long time horizons. On the Argoverse 2 2024 Scene Flow Challenge, EulerFlow outperforms all prior art, beating the next best unsupervised method by over 2.5x and the next best supervised method by over 10%.

Assessing Modality Bias in Video Question Answering Benchmarks with Multimodal Large Language Models

Multimodal large language models (MLLMs) can simultaneously process visual, textual, and auditory data, capturing insights that complement human analysis. However, existing video question-answering (VidQA) benchmarks and datasets often exhibit a bias toward a single modality, despite the goal of requiring advanced reasoning skills that integrate diverse modalities to answer the queries. In this work, we introduce the modality importance score (MIS) to identify such bias. It is designed to assess which modality embeds the necessary information to answer the question. Additionally, we propose an innovative method using state-of-the-art MLLMs to estimate the modality importance, which can serve as a proxy for human judgments of modality perception. With this MIS, we demonstrate the presence of unimodal bias and the scarcity of genuinely multimodal questions in existing datasets. We further validate the modality importance score with multiple ablation studies to evaluate the performance of MLLMs on permuted feature sets. Our results indicate that current models do not effectively integrate information due to modality imbalance in existing datasets. Our proposed MLLM-derived MIS can guide the curation of modality-balanced datasets that advance multimodal learning and enhance MLLMs' capabilities to understand and utilize synergistic relations across modalities.

Disentangling spatio-temporal knowledge for weakly supervised object detection and segmentation in surgical video

Weakly supervised video object segmentation (WSVOS) enables the identification of segmentation maps without requiring an extensive training dataset of object masks, relying instead on coarse video labels indicating object presence. Current state-of-the-art methods either require multiple independent stages of processing that employ motion cues or, in the case of end-to-end trainable networks, lack in segmentation accuracy, in part due to the difficulty of learning segmentation maps from videos with transient object presence. This limits the application of WSVOS for semantic annotation of surgical videos where multiple surgical tools frequently move in and out of the field of view, a problem that is more difficult than typically encountered in WSVOS. This paper introduces Video Spatio-Temporal Disentanglement Networks (VDST-Net), a framework to disentangle spatiotemporal information using semi-decoupled knowledge distillation to predict high-quality class activation maps (CAMs). A teacher network designed to resolve temporal conflicts when specifics about object location and timing in the video are not provided works with a student network that integrates information over time by leveraging temporal dependencies. We demonstrate the efficacy of our framework on a public reference dataset and on a more challenging surgical video dataset where objects are, on average, present in less than 60\% of annotated frames. Our method outperforms state-of-the-art techniques and generates superior segmentation masks under video-level weak supervision.

Distributed Continual Learning

This work studies the intersection of continual and federated learning, in which independent agents face unique tasks in their environments and incrementally develop and share knowledge. We introduce a mathematical framework capturing the essential aspects of distributed continual learning, including agent model and statistical heterogeneity, continual distribution shift, network topology, and communication constraints. Operating on the thesis that distributed continual learning enhances individual agent performance over single-agent learning, we identify three modes of information exchange: data instances, full model parameters, and modular (partial) model parameters. We develop algorithms for each sharing mode and conduct extensive empirical investigations across various datasets, topology structures, and communication limits. Our findings reveal three key insights: sharing parameters is more efficient than sharing data as tasks become more complex; modular parameter sharing yields the best performance while minimizing communication costs; and combining sharing modes can cumulatively improve performance.

Dissecting Deep RL with High Update Ratios: Combatting Value Overestimation and Divergence

We show that deep reinforcement learning can maintain its ability to learn without resetting network parameters in settings where the number of gradient updates greatly exceeds the number of environment samples. Under such large update-to-data ratios, a recent study by Nikishin et al. (2022) suggested the emergence of a primacy bias, in which agents overfit early interactions and downplay later experience, impairing their ability to learn. In this work, we dissect the phenomena underlying the primacy bias. We inspect the early stages of training that ought to cause the failure to learn and find that a fundamental challenge is a long-standing acquaintance: value overestimation. Overinflated Q-values are found not only on out-of-distribution but also in-distribution data and can be traced to unseen action prediction propelled by optimizer momentum. We employ a simple unit-ball normalization that enables learning under large update ratios, show its efficacy on the widely used dm_control suite, and obtain strong performance on the challenging dog tasks, competitive with model-based approaches. Our results question, in parts, the prior explanation for sub-optimal learning due to overfitting on early data.

IBCL: Zero-shot Model Generation for Task Trade-offs in Continual Learning

Like generic multi-task learning, continual learning has the nature of multi-objective optimization, and therefore faces a trade-off between the performance of different tasks. That is, to optimize for the current task distribution, it may need to compromise performance on some previous tasks. This means that there exist multiple models that are Pareto-optimal at different times, each addressing a distinct task performance trade-off. Researchers have discussed how to train particular models to address specific trade-off preferences. However, existing algorithms require training overheads proportional to the number of preferences -- a large burden when there are multiple, possibly infinitely many, preferences. As a response, we propose Imprecise Bayesian Continual Learning (IBCL). Upon a new task, IBCL (1) updates a knowledge base in the form of a convex hull of model parameter distributions and (2) obtains particular models to address task trade-off preferences with zero-shot. That is, IBCL does not require any additional training overhead to generate preference-addressing models from its knowledge base. We show that models obtained by IBCL have guarantees in identifying the Pareto optimal parameters. Moreover, experiments on standard image classification and NLP tasks support this guarantee. Statistically, IBCL improves average per-task accuracy by at most 23\% and peak per-task accuracy by at most 15\% with respect to the baseline methods, with steadily near-zero or positive backward transfer. Most importantly, IBCL significantly reduces the training overhead from training 1 model per preference to at most 3 models for all preferences.

Robotic Manipulation Datasets for Offline Compositional Reinforcement Learning

Offline reinforcement learning (RL) is a promising direction that allows RL agents to pre-train on large datasets, avoiding the recurrence of expensive data collection. To advance the field, it is crucial to generate large-scale datasets. Compositional RL is particularly appealing for generating such large datasets, since 1) it permits creating many tasks from few components, 2) the task structure may enable trained agents to solve new tasks by combining relevant learned components, and 3) the compositional dimensions provide a notion of task relatedness. This paper provides four offline RL datasets for simulated robotic manipulation created using the 256 tasks from CompoSuite [Mendez et al., 2022a]. Each dataset is collected from an agent with a different degree of performance, and consists of 256 million transitions. We provide training and evaluation settings for assessing an agent's ability to learn compositional task policies. Our benchmarking experiments on each setting show that current offline RL methods can learn the training tasks to some extent and that compositional methods significantly outperform non-compositional methods. However, current methods are still unable to extract the tasks' compositional structure to generalize to unseen tasks, showing a need for further research in offline compositional RL.

CAROM Air -- Vehicle Localization and Traffic Scene Reconstruction from Aerial Videos

Road traffic scene reconstruction from videos has been desirable by road safety regulators, city planners, researchers, and autonomous driving technology developers. However, it is expensive and unnecessary to cover every mile of the road with cameras mounted on the road infrastructure. This paper presents a method that can process aerial videos to vehicle trajectory data so that a traffic scene can be automatically reconstructed and accurately re-simulated using computers. On average, the vehicle localization error is about 0.1 m to 0.3 m using a consumer-grade drone flying at 120 meters. This project also compiles a dataset of 50 reconstructed road traffic scenes from about 100 hours of aerial videos to enable various downstream traffic analysis applications and facilitate further road traffic related research. The dataset is available at https://github.com/duolu/CAROM.

Replicable Reinforcement Learning

The replicability crisis in the social, behavioral, and data sciences has led to the formulation of algorithm frameworks for replicability -- i.e., a requirement that an algorithm produce identical outputs (with high probability) when run on two different samples from the same underlying distribution. While still in its infancy, provably replicable algorithms have been developed for many fundamental tasks in machine learning and statistics, including statistical query learning, the heavy hitters problem, and distribution testing. In this work we initiate the study of replicable reinforcement learning, providing a provably replicable algorithm for parallel value iteration, and a provably replicable version of R-max in the episodic setting. These are the first formal replicability results for control problems, which present different challenges for replication than batch learning settings.