Furong's Lab

ARMA-Networks

Dynamics-Aware-Robust-Training

ICLR 2023 paper "Exploring and Exploiting Decision Boundary Dynamics for Adversarial Robustness" by Yuancheng Xu, Yanchao Sun, Micah Goldblum, Tom Goldstein and Furong Huang

Mementos

paad_adv_rl

Code for ICLR 2022 publication: Who Is the Strongest Enemy? Towards Optimal and Efficient Evasion Attacks in Deep RL. https://openreview.net/forum?id=JM2kFbJvvI

perceptionCLIP

Code for our ICLR 2024 paper "PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts"

SWIFT

SWIFT: Shared WaIt Free Transmission

tuformer

VLM-Poisoning

Code for Neurips 2024 paper "Shadowcast: Stealthy Data Poisoning Attacks Against Vision-Language Models"

WAVES

Code for our paper "Benchmarking the Robustness of Image Watermarks"

WocaR-RL

Efficient Adversarial Training without Attacking: Worst-Case-Aware Robust Reinforcement Learning

umd-huang-lab/reinforcement-learning-via-spectral-methods

Model-based reinforcement learning algorithms make decisions by building and utilizing a model of the environment. However, none of the existing algorithms attempts to infer the dynamics of any state-action pair from known state-action pairs before meeting it for sufficient times. We propose a new model-based method called Greedy Inference Model (GIM) that infers the unknown dynamics from known dynamics based on the internal spectral properties of the environment. In other words, GIM can “learn by analogy”. We further introduce a new exploration strategy which ensures that the agent rapidly and evenly visits unknown state-action pairs. GIM is much more computationally efficient than state-of-the-art model-based algorithms, as the number of dynamic programming operations is independent of the environment size. Lower sample complexity could also be achieved under mild conditions compared against methods without inferring. Experimental results demon- strate the effectiveness and efficiency of GIM in a variety of real- world tasks.

umd-huang-lab/template-reinforcement-learning

umd-huang-lab/neural-net-generalization-via-tensor

Deep neural networks generalize well on unseen data though the number of parameters often far exceeds the number of training examples. Recently proposed complexity measures have provided insights to understanding the generalizability in neural networks from perspectives of PAC-Bayes, robustness, overparametrization, compression and so on. In this work, we advance the understanding of the relations between the network's architecture and its generalizability from the compression perspective. Using tensor analysis, we propose a series of intuitive, data-dependent and easily-measurable properties that tightly characterize the compressibility and generalizability of neural networks; thus, in practice, our generalization bound outperforms the previous compression-based ones, especially for neural networks using tensors as their weight kernels (e.g. CNNs). Moreover, these intuitive measurements provide further insights into designing neural network architectures with properties favorable for better/guaranteed generalizability. Our experimental results demonstrate that through the proposed measurable properties, our generalization error bound matches the trend of the test error well. Our theoretical analysis further provides justifications for the empirical success and limitations of some widely-used tensor-based compression approaches. We also discover the improvements to the compressibility and robustness of current neural networks when incorporating tensor operations via our proposed layer-wise structure.

umd-huang-lab/parallel-tnn

umd-huang-lab/Bayesian-Quantized-Networks

umd-huang-lab/intent-shift

Off-policy evaluation in contextual bandits, which evaluates the reward of a target policy given the history of a logged pol- icy, is a task of importance as it provides an estimate of the performance of a new policy without experimenting with it. Existing off-policy evaluation methods in contextual bandits make an oversimplified assumption that the distribution of contexts is stationary. In this paper, we consider a more prac- tical setting of a context/reward distribution shift between the logged data and the contexts observed for evaluating a target policy in the future. We propose an intent shift model which introduces a latent intent variable to capture the distribution shift on context and reward, avoiding the intractable prob- lem of density estimation of contexts in high-dimension. Un- der the intent shift model, we introduce a consistent spectral- based IPS estimator, characterize its finite-sample complexity and derive an MSE bound on the performance of the final re- ward estimation. Experiments demonstrate that the proposed spectral-based IPS estimator outperforms the existing estimators under distribution shift.

umd-huang-lab/ParallelTNNLayers

Parallel implementations of decomposed tensorial neural network layers