While diffusion models have achieved great success in generating continuous signals such as images and audio, it remains elusive for diffusion models in learning discrete sequence data like natural languages. Although recent advances circumvent this challenge of discreteness by embedding discrete tokens as continuous surrogates, they still fall short of satisfactory generation quality. To understand this, we first dive deep into the denoised training protocol of diffusion-based sequence generative models and determine their three severe problems, i.e., 1) failing to learn, 2) lack of scalability, and 3) neglecting source conditions. We argue that these problems can be boiled down to the pitfall of the not completely eliminated discreteness in the embedding space, and the scale of noises is decisive herein. In this paper, we introduce DINOISER to facilitate diffusion models for sequence generation by manipulating noises. We propose to adaptively determine the range of sampled noise scales for counter-discreteness training; and encourage the proposed diffused sequence learner to leverage source conditions with amplified noise scales during inference. Experiments show that DINOISER enables consistent improvement over the baselines of previous diffusion-based sequence generative models on several conditional sequence modeling benchmarks thanks to both effective training and inference strategies. Analyses further verify that DINOISER can make better use of source conditions to govern its generative process.

We investigate the benefit of treating all the parameters in a Bayesian neural network stochastically and find compelling theoretical and empirical evidence that this standard construction may be unnecessary. To this end, we prove that expressive predictive distributions require only small amounts of stochasticity. In particular, partially stochastic networks with only $n$ stochastic biases are universal probabilistic predictors for $n$-dimensional predictive problems. In empirical investigations, we find no systematic benefit of full stochasticity across four different inference modalities and eight datasets; partially stochastic networks can match and sometimes even outperform fully stochastic networks, despite their reduced memory costs.

Solving continuous Partially Observable Markov Decision Processes (POMDPs) is challenging, particularly for high-dimensional continuous action spaces. To alleviate this difficulty, we propose a new sampling-based online POMDP solver, called Adaptive Discretization using Voronoi Trees (ADVT). It uses Monte Carlo Tree Search in combination with an adaptive discretization of the action space as well as optimistic optimization to efficiently sample high-dimensional continuous action spaces and compute the best action to perform. Specifically, we adaptively discretize the action space for each sampled belief using a hierarchical partition called Voronoi tree, which is a Binary Space Partitioning that implicitly maintains the partition of a cell as the Voronoi diagram of two points sampled from the cell. ADVT uses the estimated diameters of the cells to form an upper-confidence bound on the action value function within the cell, guiding the Monte Carlo Tree Search expansion and further discretization of the action space. This enables ADVT to better exploit local information with respect to the action value function, allowing faster identification of the most promising regions in the action space, compared to existing solvers. Voronoi trees keep the cost of partitioning and estimating the diameter of each cell low, even in high-dimensional spaces where many sampled points are required to cover the space well. ADVT additionally handles continuous observation spaces, by adopting an observation progressive widening strategy, along with a weighted particle representation of beliefs. Experimental results indicate that ADVT scales substantially better to high-dimensional continuous action spaces, compared to state-of-the-art methods.

Understanding the complexity of sampling from a strongly log-concave and log-smooth distribution $\pi$ on $\mathbb{R}^d$ to high accuracy is a fundamental problem, both from a practical and theoretical standpoint. In practice, high-accuracy samplers such as the classical Metropolis-adjusted Langevin algorithm (MALA) remain the de facto gold standard; and in theory, via the proximal sampler reduction, it is understood that such samplers are key for sampling even beyond log-concavity (in particular, for distributions satisfying isoperimetric assumptions). In this work, we improve the dimension dependence of this sampling problem to $\tilde{O}(d^{1/2})$, whereas the previous best result for MALA was $\tilde{O}(d)$. This closes the long line of work on the complexity of MALA, and moreover leads to state-of-the-art guarantees for high-accuracy sampling under strong log-concavity and beyond (thanks to the aforementioned reduction). Our starting point is that the complexity of MALA improves to $\tilde{O}(d^{1/2})$, but only under a warm start (an initialization with constant R\'enyi divergence w.r.t. $\pi$). Previous algorithms took much longer to find a warm start than to use it, and closing this gap has remained an important open problem in the field. Our main technical contribution settles this problem by establishing the first $\tilde{O}(d^{1/2})$ R\'enyi mixing rates for the discretized underdamped Langevin diffusion. For this, we develop new differential-privacy-inspired techniques based on R\'enyi divergences with Orlicz--Wasserstein shifts, which allow us to sidestep longstanding challenges for proving fast convergence of hypocoercive differential equations.

When learning task-oriented dialogue (ToD) agents, reinforcement learning (RL) techniques can naturally be utilized to train dialogue strategies to achieve user-specific goals. Prior works mainly focus on adopting advanced RL techniques to train the ToD agents, while the design of the reward function is not well studied. This paper aims at answering the question of how to efficiently learn and leverage a reward function for training end-to-end (E2E) ToD agents. Specifically, we introduce two generalized objectives for reward-function learning, inspired by the classical learning-to-rank literature. Further, we utilize the learned reward function to guide the training of the E2E ToD agent. With the proposed techniques, we achieve competitive results on the E2E response-generation task on the Multiwoz 2.0 dataset. Source code and checkpoints are publicly released at https://github.com/Shentao-YANG/Fantastic_Reward_ICLR2023.

Generalization in partially observed markov decision processes (POMDPs) is critical for successful applications of visual reinforcement learning (VRL) in real scenarios. A widely used idea is to learn task-relevant representations that encode task-relevant information of common features in POMDPs, i.e., rewards and transition dynamics. As transition dynamics in the latent state space -- which are task-relevant and invariant to visual distractions -- are unknown to the agents, existing methods alternatively use transition dynamics in the observation space to extract task-relevant information in transition dynamics. However, such transition dynamics in the observation space involve task-irrelevant visual distractions, degrading the generalization performance of VRL methods. To tackle this problem, we propose the reward sequence distribution conditioned on the starting observation and the predefined subsequent action sequence (RSD-OA). The appealing features of RSD-OA include that: (1) RSD-OA is invariant to visual distractions, as it is conditioned on the predefined subsequent action sequence without task-irrelevant information from transition dynamics, and (2) the reward sequence captures long-term task-relevant information in both rewards and transition dynamics. Experiments demonstrate that our representation learning approach based on RSD-OA significantly improves the generalization performance on unseen environments, outperforming several state-of-the-arts on DeepMind Control tasks with visual distractions.

Best-of-both-worlds algorithms for online learning which achieve near-optimal regret in both the adversarial and the stochastic regimes have received growing attention recently. Existing techniques often require careful adaptation to every new problem setup, including specialised potentials and careful tuning of algorithm parameters. Yet, in domains such as linear bandits, it is still unknown if there exists an algorithm that can simultaneously obtain $O(\log(T))$ regret in the stochastic regime and $\tilde{O}(\sqrt{T})$ regret in the adversarial regime. In this work, we resolve this question positively and present a general reduction from best of both worlds to a wide family of follow-the-regularized-leader (FTRL) and online-mirror-descent (OMD) algorithms. We showcase the capability of this reduction by transforming existing algorithms that are only known to achieve worst-case guarantees into new algorithms with best-of-both-worlds guarantees in contextual bandits, graph bandits and tabular Markov decision processes.

Utilizing messages from teammates can improve coordination in cooperative Multi-agent Reinforcement Learning (MARL). Previous works typically combine raw messages of teammates with local information as inputs for policy. However, neglecting message aggregation poses significant inefficiency for policy learning. Motivated by recent advances in representation learning, we argue that efficient message aggregation is essential for good coordination in cooperative MARL. In this paper, we propose Multi-Agent communication via Self-supervised Information Aggregation (MASIA), where agents can aggregate the received messages into compact representations with high relevance to augment the local policy. Specifically, we design a permutation invariant message encoder to generate common information-aggregated representation from messages and optimize it via reconstructing and shooting future information in a self-supervised manner. Hence, each agent would utilize the most relevant parts of the aggregated representation for decision-making by a novel message extraction mechanism. Furthermore, considering the potential of offline learning for real-world applications, we build offline benchmarks for multi-agent communication, which is the first as we know. Empirical results demonstrate the superiority of our method in both online and offline settings. We also release the built offline benchmarks in this paper as a testbed for communication ability validation to facilitate further future research.

Deducing the contribution of each agent and assigning the corresponding reward to them is a crucial problem in cooperative Multi-Agent Reinforcement Learning (MARL). Previous studies try to resolve the issue through designing an intrinsic reward function, but the intrinsic reward is simply combined with the environment reward by summation in these studies, which makes the performance of their MARL framework unsatisfactory. We propose a novel method named Attention Individual Intrinsic Reward Mixing Network (AIIR-MIX) in MARL, and the contributions of AIIR-MIX are listed as follows:(a) we construct a novel intrinsic reward network based on the attention mechanism to make teamwork more effective. (b) we propose a Mixing network that is able to combine intrinsic and extrinsic rewards non-linearly and dynamically in response to changing conditions of the environment. We compare AIIR-MIX with many State-Of-The-Art (SOTA) MARL methods on battle games in StarCraft II. And the results demonstrate that AIIR-MIX performs admirably and can defeat the current advanced methods on average test win rate. To validate the effectiveness of AIIR-MIX, we conduct additional ablation studies. The results show that AIIR-MIX can dynamically assign each agent a real-time intrinsic reward in accordance with their actual contribution.

While machine learning models rapidly advance the state-of-the-art on various real-world tasks, out-of-domain (OOD) generalization remains a challenging problem given the vulnerability of these models to spurious correlations. We propose a balanced mini-batch sampling strategy to transform a biased data distribution into a spurious-free balanced distribution, based on the invariance of the underlying causal mechanisms for the data generation process. We argue that the Bayes optimal classifiers trained on such balanced distribution are minimax optimal across a diverse enough environment space. We also provide an identifiability guarantee of the latent variable model of the proposed data generation process, when utilizing enough train environments. Experiments are conducted on DomainBed, demonstrating empirically that our method obtains the best performance across 20 baselines reported on the benchmark.