In online reinforcement learning (RL), efficient exploration remains particularly challenging in high-dimensional environments with sparse rewards. In low-dimensional environments, where tabular parameterization is possible, count-based upper confidence bound (UCB) exploration methods achieve minimax near-optimal rates. However, it remains unclear how to efficiently implement UCB in realistic RL tasks that involve non-linear function approximation. To address this, we propose a new exploration approach via \textit{maximizing} the deviation of the occupancy of the next policy from the explored regions. We add this term as an adaptive regularizer to the standard RL objective to balance exploration vs. exploitation. We pair the new objective with a provably convergent algorithm, giving rise to a new intrinsic reward that adjusts existing bonuses. The proposed intrinsic reward is easy to implement and combine with other existing RL algorithms to conduct exploration. As a proof of concept, we evaluate the new intrinsic reward on tabular examples across a variety of model-based and model-free algorithms, showing improvements over count-only exploration strategies. When tested on navigation and locomotion tasks from MiniGrid and DeepMind Control Suite benchmarks, our approach significantly improves sample efficiency over state-of-the-art methods. Our code is available at https://github.com/tianjunz/MADE.
Adversarial examples have appeared as a ubiquitous property of machine learning models where bounded adversarial perturbation could mislead the models to make arbitrarily incorrect predictions. Such examples provide a way to assess the robustness of machine learning models as well as a proxy for understanding the model training process. Extensive studies try to explain the existence of adversarial examples and provide ways to improve model robustness (e.g. adversarial training). While they mostly focus on models trained on datasets with predefined labels, we leverage the teacher-student framework and assume a teacher model, or oracle, to provide the labels for given instances. We extend Tian (2019) in the case of low-rank input data and show that student specialization (trained student neuron is highly correlated with certain teacher neuron at the same layer) still happens within the input subspace, but the teacher and student nodes could differ wildly out of the data subspace, which we conjecture leads to adversarial examples. Extensive experiments show that student specialization correlates strongly with model robustness in different scenarios, including student trained via standard training, adversarial training, confidence-calibrated adversarial training, and training with robust feature dataset. Our studies could shed light on the future exploration about adversarial examples, and enhancing model robustness via principled data augmentation.
Policy gradient methods are widely used in solving two-player zero-sum games to achieve superhuman performance in practice. However, it remains elusive when they can provably find a near-optimal solution and how many samples and iterations are needed. The current paper studies natural extensions of Natural Policy Gradient algorithm for solving two-player zero-sum games where function approximation is used for generalization across states. We thoroughly characterize the algorithms' performance in terms of the number of samples, number of iterations, concentrability coefficients, and approximation error. To our knowledge, this is the first quantitative analysis of policy gradient methods with function approximation for two-player zero-sum Markov games.
Contrastive approaches to self-supervised learning (SSL) learn representations by minimizing the distance between two augmented views of the same data point (positive pairs) and maximizing the same from different data points (negative pairs). However, recent approaches like BYOL and SimSiam, show remarkable performance {\it without} negative pairs, raising a fundamental theoretical question: how can SSL with only positive pairs avoid representational collapse? We study the nonlinear learning dynamics of non-contrastive SSL in simple linear networks. Our analysis yields conceptual insights into how non-contrastive SSL methods learn, how they avoid representational collapse, and how multiple factors, like predictor networks, stop-gradients, exponential moving averages, and weight decay all come into play. Our simple theory recapitulates the results of real-world ablation studies in both STL-10 and ImageNet. Furthermore, motivated by our theory we propose a novel approach that \emph{directly} sets the predictor based on the statistics of its inputs. In the case of linear predictors, our approach outperforms gradient training of the predictor by $5\%$ and on ImageNet it performs comparably with more complex two-layer non-linear predictors that employ BatchNorm. Code is released in https://github.com/facebookresearch/luckmatters/tree/master/ssl.
Efficient exploration under sparse rewards remains a key challenge in deep reinforcement learning. To guide exploration, previous work makes extensive use of intrinsic reward (IR). There are many heuristics for IR, including visitation counts, curiosity, and state-difference. In this paper, we analyze the pros and cons of each method and propose the regulated difference of inverse visitation counts as a simple but effective criterion for IR. The criterion helps the agent explore Beyond the Boundary of explored regions and mitigates common issues in count-based methods, such as short-sightedness and detachment. The resulting method, BeBold, solves the 12 most challenging procedurally-generated tasks in MiniGrid with just 120M environment steps, without any curriculum learning. In comparison, the previous SoTA only solves 50% of the tasks. BeBold also achieves SoTA on multiple tasks in NetHack, a popular rogue-like game that contains more challenging procedurally-generated environments.
Differential Neural Architecture Search (NAS) requires all layer choices to be held in memory simultaneously; this limits the size of both search space and final architecture. In contrast, Probabilistic NAS, such as PARSEC, learns a distribution over high-performing architectures, and uses only as much memory as needed to train a single model. Nevertheless, it needs to sample many architectures, making it computationally expensive for searching in an extensive space. To solve these problems, we propose a sampling method adaptive to the distribution entropy, drawing more samples to encourage explorations at the beginning, and reducing samples as learning proceeds. Furthermore, to search fast in the multi-variate space, we propose a coarse-to-fine strategy by using a factorized distribution at the beginning which can reduce the number of architecture parameters by over an order of magnitude.We call this method Fast Probabilistic NAS (FP-NAS). Compared with PARSEC, it can sample 64% fewer architectures and search 2.1x faster. Compared with FBNetV2, FP-NAS is 1.9x - 3.6x faster, and the searched models outperform FBNetV2 models on ImageNet. FP-NAS allows us to expand the giant FBNetV2 space to be wider (i.e. larger channel choices) and deeper (i.e. more blocks), while adding Split-Attention block and enabling the search over the number of splits. When searching a model of size 0.4G FLOPS, FP-NAS is 132x faster than EfficientNet, and the searched FP-NAS-L0 model outperforms EfficientNet-B0 by 0.6% accuracy. Without using any architecture surrogate or scaling tricks, we directly search large models up to 1.0G FLOPS. Our FP-NAS-L2 model with simple distillation outperforms BigNAS-XL with advanced inplace distillation by 0.7% accuracy with less FLOPS.
We propose a novel theoretical framework to understand self-supervised learning methods that employ dual pairs of deep ReLU networks (e.g., SimCLR, BYOL). First, we prove that in each SGD update of SimCLR, the weights at each layer are updated by a \emph{covariance operator} that specifically amplifies initial random selectivities that vary across data samples but survive averages over data augmentations, which we show leads to the emergence of hierarchical features, if the input data are generated from a hierarchical latent tree model. With the same framework, we also show analytically that in BYOL, the usage of BatchNorm and a predictor creates an implicit contrastive term, acting as an approximate covariance operator. The term is formed by the inter-play between the zero-mean operation of BatchNorm and the extra predictor in the online network. Extensive ablation studies justify our theoretical findings.
Recent advances in multi-agent reinforcement learning (MARL) have achieved super-human performance in games like Quake 3 and Dota 2. Unfortunately, these techniques require orders-of-magnitude more training rounds than humans and don't generalize to new agent configurations even on the same game. In this work, we propose Collaborative Q-learning (CollaQ) that achieves state-of-the-art performance in the StarCraft multi-agent challenge and supports ad hoc team play. We first formulate multi-agent collaboration as a joint optimization on reward assignment and show that each agent has an approximately optimal policy that decomposes into two parts: one part that only relies on the agent's own state, and the other part that is related to states of nearby agents. Following this novel finding, CollaQ decomposes the Q-function of each agent into a self term and an interactive term, with a Multi-Agent Reward Attribution (MARA) loss that regularizes the training. CollaQ is evaluated on various StarCraft maps and shows that it outperforms existing state-of-the-art techniques (i.e., QMIX, QTRAN, and VDN) by improving the win rate by 40% with the same number of samples. In the more challenging ad hoc team play setting (i.e., reweight/add/remove units without re-training or finetuning), CollaQ outperforms previous SoTA by over 30%.