University of Michigan, Ann Arbor
Abstract:We propose and address a novel few-shot RL problem, where a task is characterized by a subtask graph which describes a set of subtasks and their dependencies that are unknown to the agent. The agent needs to quickly adapt to the task over few episodes during adaptation phase to maximize the return in the test phase. Instead of directly learning a meta-policy, we develop a Meta-learner with Subtask Graph Inference(MSGI), which infers the latent parameter of the task by interacting with the environment and maximizes the return given the latent parameter. To facilitate learning, we adopt an intrinsic reward inspired by upper confidence bound (UCB) that encourages efficient exploration. Our experiment results on two grid-world domains and StarCraft II environments show that the proposed method is able to accurately infer the latent task parameter, and to adapt more efficiently than existing meta RL and hierarchical RL methods.
Abstract:Adversarial training is an effective defense method to protect classification models against adversarial attacks. However, one limitation of this approach is that it can require orders of magnitude additional training time due to high cost of generating strong adversarial examples during training. In this paper, we first show that there is high transferability between models from neighboring epochs in the same training process, i.e., adversarial examples from one epoch continue to be adversarial in subsequent epochs. Leveraging this property, we propose a novel method, Adversarial Training with Transferable Adversarial Examples (ATTA), that can enhance the robustness of trained models and greatly improve the training efficiency by accumulating adversarial perturbations through epochs. Compared to state-of-the-art adversarial training methods, ATTA enhances adversarial accuracy by up to 7.2% on CIFAR10 and requires 12~14x less training time on MNIST and CIFAR10 datasets with comparable model robustness.
Abstract:We present a method for learning intrinsic reward functions to drive the learning of an agent during periods of practice in which extrinsic task rewards are not available. During practice, the environment may differ from the one available for training and evaluation with extrinsic rewards. We refer to this setup of alternating periods of practice and objective evaluation as practice-match, drawing an analogy to regimes of skill acquisition common for humans in sports and games. The agent must effectively use periods in the practice environment so that performance improves during matches. In the proposed method the intrinsic practice reward is learned through a meta-gradient approach that adapts the practice reward parameters to reduce the extrinsic match reward loss computed from matches. We illustrate the method on a simple grid world, and evaluate it in two games in which the practice environment differs from match: Pong with practice against a wall without an opponent, and PacMan with practice in a maze without ghosts. The results show gains from learning in practice in addition to match periods over learning in matches only.
Abstract:Predicting future video frames is extremely challenging, as there are many factors of variation that make up the dynamics of how frames change through time. Previously proposed solutions require complex inductive biases inside network architectures with highly specialized computation, including segmentation masks, optical flow, and foreground and background separation. In this work, we question if such handcrafted architectures are necessary and instead propose a different approach: finding minimal inductive bias for video prediction while maximizing network capacity. We investigate this question by performing the first large-scale empirical study and demonstrate state-of-the-art performance by learning large models on three different datasets: one for modeling object interactions, one for modeling human motion, and one for modeling car driving.
Abstract:Generative Adversarial Networks (GANs) are known to be difficult to train, despite considerable research effort. Several regularization techniques for stabilizing training have been proposed, but they introduce non-trivial computational overheads and interact poorly with existing techniques like spectral normalization. In this work, we propose a simple, effective training stabilizer based on the notion of consistency regularization---a popular technique in the semi-supervised learning literature. In particular, we augment data passing into the GAN discriminator and penalize the sensitivity of the discriminator to these augmentations. We conduct a series of experiments to demonstrate that consistency regularization works effectively with spectral normalization and various GAN architectures, loss functions and optimizer settings. Our method achieves the best FID scores for unconditional image generation compared to other regularization methods on CIFAR-10 and CelebA. Moreover, Our consistency regularized GAN (CR-GAN) improves state-of-the-art FID scores for conditional generation from 14.73 to 11.67 on CIFAR-10 and from 8.73 to 6.66 on ImageNet-2012.
Abstract:Collecting large-scale data with clean labels for supervised training of neural networks is practically challenging. Although noisy labels are usually cheap to acquire, existing methods suffer severely for training datasets with high noise ratios, making high-cost human labeling a necessity. Here we present a method to train neural networks in a way that is almost invulnerable to severe label noise by utilizing a tiny trusted set. Our method, named IEG, is based on three key insights: (i) Isolation of noisy labels, (ii) Escalation of useful supervision from mislabeled data, and (iii) Guidance from small trusted data. On CIFAR100 with a 40% uniform noise ratio and 10 trusted labeled data per class, our method achieves $80.2{\pm}0.3\%$ classification accuracy, only 1.4% higher error than a neural network trained without label noise. Moreover, increasing the noise ratio to 80%, our method still achieves a high accuracy of $75.5{\pm}0.2\%$, compared to the previous best 47.7%. Finally, our method sets new state of the art on various types of challenging label corruption types and levels and large-scale WebVision benchmarks.
Abstract:Deep reinforcement learning (RL) agents often fail to generalize to unseen environments (yet semantically similar to trained agents), particularly when they are trained on high-dimensional state spaces, such as images. In this paper, we propose a simple technique to improve a generalization ability of deep RL agents by introducing a randomized (convolutional) neural network that randomly perturbs input observations. It enables trained agents to adapt to new domains by learning robust features invariant across varied and randomized environments. Furthermore, we consider an inference method based on the Monte Carlo approximation to reduce the variance induced by this randomization. We demonstrate the superiority of our method across 2D CoinRun, 3D DeepMind Lab exploration and 3D robotics control tasks: it significantly outperforms various regularization and data augmentation methods for the same purpose.
Abstract:Hierarchical reinforcement learning has demonstrated significant success at solving difficult reinforcement learning (RL) tasks. Previous works have motivated the use of hierarchy by appealing to a number of intuitive benefits, including learning over temporally extended transitions, exploring over temporally extended periods, and training and exploring in a more semantically meaningful action space, among others. However, in fully observed, Markovian settings, it is not immediately clear why hierarchical RL should provide benefits over standard "shallow" RL architectures. In this work, we isolate and evaluate the claimed benefits of hierarchical RL on a suite of tasks encompassing locomotion, navigation, and manipulation. Surprisingly, we find that most of the observed benefits of hierarchy can be attributed to improved exploration, as opposed to easier policy learning or imposed hierarchical structures. Given this insight, we present exploration techniques inspired by hierarchy that achieve performance competitive with hierarchical RL while at the same time being much simpler to use and implement.
Abstract:This paper proposes a method for learning a trajectory-conditioned policy to imitate diverse demonstrations from the agent's own past experiences. We demonstrate that such self-imitation drives exploration in diverse directions and increases the chance of finding a globally optimal solution in reinforcement learning problems, especially when the reward is sparse and deceptive. Our method significantly outperforms existing self-imitation learning and count-based exploration methods on various sparse-reward reinforcement learning tasks with local optima. In particular, we report a state-of-the-art score of more than 25,000 points on Montezuma's Revenge without using expert demonstrations or resetting to arbitrary states.
Abstract:Training a deep network policy for robot manipulation is notoriously costly and time consuming as it depends on collecting a significant amount of real world data. To work well in the real world, the policy needs to see many instances of the task, including various object arrangements in the scene as well as variations in object geometry, texture, material, and environmental illumination. In this paper, we propose a method that learns to perform table-top instance grasping of a wide variety of objects while using no real world grasping data, outperforming the baseline using 2.5D shape by 10%. Our method learns 3D point cloud of object, and use that to train a domain-invariant grasping policy. We formulate the learning process as a two-step procedure: 1) Learning a domain-invariant 3D shape representation of objects from about 76K episodes in simulation and about 530 episodes in the real world, where each episode lasts less than a minute and 2) Learning a critic grasping policy in simulation only based on the 3D shape representation from step 1. Our real world data collection in step 1 is both cheaper and faster compared to existing approaches as it only requires taking multiple snapshots of the scene using a RGBD camera. Finally, the learned 3D representation is not specific to grasping, and can potentially be used in other interaction tasks.