ICML 2023 Papers

Aligning language models with preferences can be posed as approximating a target distribution representing some desired behavior. Existing approaches differ both in the functional form of the target distribution and the algorithm used to approximate it. For instance, Reinforcement Learning from Human Feedback (RLHF) corresponds to minimizing a reverse KL from an implicit target distribution arising from a KL penalty in the objective. On the other hand, Generative Distributional Control (GDC) has an explicit target distribution and minimizes a forward KL from it using the Distributional Policy Gradient (DPG) algorithm. In this paper, we propose a new approach, $f$-DPG, which allows the use of any $f$-divergence to approximate any target distribution that can be evaluated. $f$-DPG unifies both frameworks (RLHF, GDC) and the approximation methods (DPG, RL with KL penalties). We show the practical benefits of various choices of divergence objectives and demonstrate that there is no universally optimal objective but that different divergences present different alignment and diversity trade-offs. We show that Jensen-Shannon divergence strikes a good balance between these objectives, and frequently outperforms forward KL divergence by a wide margin, leading to significant improvements over prior work. These distinguishing characteristics between divergences persist as the model size increases, highlighting the importance of selecting appropriate divergence objectives.

Out-of-distribution (OOD) detection is a critical requirement for the deployment of deep neural networks. This paper introduces the HEAT model, a new post-hoc OOD detection method estimating the density of in-distribution (ID) samples using hybrid energy-based models (EBM) in the feature space of a pre-trained backbone. HEAT complements prior density estimators of the ID density, e.g. parametric models like the Gaussian Mixture Model (GMM), to provide an accurate yet robust density estimation. A second contribution is to leverage the EBM framework to provide a unified density estimation and to compose several energy terms. Extensive experiments demonstrate the significance of the two contributions. HEAT sets new state-of-the-art OOD detection results on the CIFAR-10 / CIFAR-100 benchmark as well as on the large-scale Imagenet benchmark. The code is available at: https://github.com/MarcLafon/heatood.

Understanding when the noise in stochastic gradient descent (SGD) affects generalization of deep neural networks remains a challenge, complicated by the fact that networks can operate in distinct training regimes. Here we study how the magnitude of this noise $T$ affects performance as the size of the training set $P$ and the scale of initialization $\alpha$ are varied. For gradient descent, $\alpha$ is a key parameter that controls if the network is `lazy' ($\alpha\gg1$) or instead learns features ($\alpha\ll1$). For classification of MNIST and CIFAR10 images, our central results are: *(i)* obtaining phase diagrams for performance in the $(\alpha,T)$ plane. They show that SGD noise can be detrimental or instead useful depending on the training regime. Moreover, although increasing $T$ or decreasing $\alpha$ both allow the net to escape the lazy regime, these changes can have opposite effects on performance. *(ii)* Most importantly, we find that the characteristic temperature $T_c$ where the noise of SGD starts affecting the trained model (and eventually performance) is a power law of $P$. We relate this finding with the observation that key dynamical quantities, such as the total variation of weights during training, depend on both $T$ and $P$ as power laws. These results indicate that a key effect of SGD noise occurs late in training, by affecting the stopping process whereby all data are fitted. Indeed, we argue that due to SGD noise, nets must develop a stronger `signal', i.e. larger informative weights, to fit the data, leading to a longer training time. A stronger signal and a longer training time are also required when the size of the training set $P$ increases. We confirm these views in the perceptron model, where signal and noise can be precisely measured. Interestingly, exponents characterizing the effect of SGD depend on the density of data near the decision boundary, as we explain.

Foundation models have shown impressive adaptation and scalability in supervised and self-supervised learning problems, but so far these successes have not fully translated to reinforcement learning (RL). In this work, we demonstrate that training an RL agent at scale leads to a general in-context learning algorithm that can adapt to open-ended novel embodied 3D problems as quickly as humans. In a vast space of held-out environment dynamics, our adaptive agent (AdA) displays on-the-fly hypothesis-driven exploration, efficient exploitation of acquired knowledge, and can successfully be prompted with first-person demonstrations. Adaptation emerges from three ingredients: (1) meta-reinforcement learning across a vast, smooth and diverse task distribution, (2) a policy parameterised as a large-scale attention-based memory architecture, and (3) an effective automated curriculum that prioritises tasks at the frontier of an agent's capabilities. We demonstrate characteristic scaling laws with respect to network size, memory length, and richness of the training task distribution. We believe our results lay the foundation for increasingly general and adaptive RL agents that perform well across ever-larger open-ended domains.

Spatial-temporal graph learning has emerged as the state-of-the-art solution for modeling structured spatial-temporal data in learning region representations for various urban sensing tasks (e.g., crime forecasting, traffic flow prediction). However, most existing models are vulnerable to the quality of the generated region graph due to the inartistic graph-structured information aggregation schema. The ubiquitous spatial-temporal data noise and incompleteness in real-life scenarios bring difficulties to generate high-quality region representations. In this paper, we propose a Spatial-Temporal Adversarial Graph contrastive learning model (STAG) to tackle this challenge for adaptive self-supervised graph augmentation. Specifically, we propose a learnable contrastive learning function that enables the automated distillation of important multi-view self-supervised signals for adaptive spatial-temporal graph augmentation. To enhance the representation discrimination ability and robustness, the designed adversarial contrastive learning mechanism empowers STAG to adaptively identify hard samples for better self-supervision. Finally, a cross-view contrastive learning paradigm is introduced to model the inter-dependencies across view-specific region representations and preserve the underlying relation heterogeneity. We verify the superiority of our STAG method in various spatial-temporal prediction tasks on several benchmark datasets.

Automated decision support systems promise to help human experts solve multiclass classification tasks more efficiently and accurately. However, existing systems typically require experts to understand when to cede agency to the system or when to exercise their own agency. Otherwise, the experts may be better off solving the classification tasks on their own. In this work, we develop an automated decision support system that, by design, does not require experts to understand when to trust the system to improve performance. Rather than providing (single) label predictions and letting experts decide when to trust these predictions, our system provides sets of label predictions constructed using conformal prediction---prediction sets---and forcefully asks experts to predict labels from these sets. By using conformal prediction, our system can precisely trade-off the probability that the true label is not in the prediction set, which determines how frequently our system will mislead the experts, and the size of the prediction set, which determines the difficulty of the classification task the experts need to solve using our system. In addition, we develop an efficient and near-optimal search method to find the conformal predictor under which the experts benefit the most from using our system. Simulation experiments using synthetic and real expert predictions demonstrate that our system may help experts make more accurate predictions and is robust to the accuracy of the classifier the conformal predictor relies on.

Sampling diverse programs from a code language model and reranking with model likelihood is a popular method for code generation but it is prone to preferring degenerate solutions. Inspired by collaborative programming, we propose Coder-Reviewer reranking. We augment Coder language models from past work, which generate programs given language instructions, with Reviewer models, which evaluate the likelihood of the instruction given the generated programs. We perform an extensive study across six datasets with eight models from three model families. Experimental results show that Coder-Reviewer reranking leads to consistent and significant improvement (up to 17% absolute accuracy gain) over reranking with the Coder model only. When combined with executability filtering, Coder-Reviewer reranking can often outperform the minimum Bayes risk method. Coder-Reviewer reranking is easy to implement by prompting, can generalize to different programming languages, and works well with off-the-shelf hyperparameters.

Algorithms for safely improving policies are important to deploy reinforcement learning approaches in real-world scenarios. In this work, we propose an algorithm, called MCTS-SPIBB, that computes safe policy improvement online using a Monte Carlo Tree Search based strategy. We theoretically prove that the policy generated by MCTS-SPIBB converges, as the number of simulations grows, to the optimal safely improved policy generated by Safe Policy Improvement with Baseline Bootstrapping (SPIBB), a popular algorithm based on policy iteration. Moreover, our empirical analysis performed on three standard benchmark domains shows that MCTS-SPIBB scales to significantly larger problems than SPIBB because it computes the policy online and locally, i.e., only in the states actually visited by the agent.

Nearly all practical neural models for classification are trained using the cross-entropy loss. Yet this ubiquitous choice is supported by little theoretical or empirical evidence. Recent work (Hui & Belkin, 2020) suggests that training using the (rescaled) square loss is often superior in terms of the classification accuracy. In this paper we propose the "squentropy" loss, which is the sum of two terms: the cross-entropy loss and the average square loss over the incorrect classes. We provide an extensive set of experiment on multi-class classification problems showing that the squentropy loss outperforms both the pure cross-entropy and rescaled square losses in terms of the classification accuracy. We also demonstrate that it provides significantly better model calibration than either of these alternative losses and, furthermore, has less variance with respect to the random initialization. Additionally, in contrast to the square loss, squentropy loss can frequently be trained using exactly the same optimization parameters, including the learning rate, as the standard cross-entropy loss, making it a true ''plug-and-play'' replacement. Finally, unlike the rescaled square loss, multiclass squentropy contains no parameters that need to be adjusted.

Many problems in science and engineering involve optimizing an expensive black-box function over a high-dimensional space. In the offline model-based optimization (MBO) setting, we assume access to a fixed, offline dataset for pretraining and a small budget for online function evaluations. Prior approaches seek to utilize the offline data to approximate the function or its inverse but are not sufficiently accurate far from the data distribution. We propose BONET, a generative framework for pretraining a novel model-based optimizer using offline datasets. In BONET, we train an autoregressive model on fixed-length trajectories derived from an offline dataset. We design a sampling strategy to synthesize trajectories from offline data using a simple heuristic of rolling out monotonic transitions from low-fidelity to high-fidelity samples. Empirically, we instantiate BONET using a causally masked Transformer (Radford et al., 2019) and evaluate it on Design-Bench (Trabucco et al., 2022), where we rank the best on average, outperforming state-of-the-art baselines.

Regret minimization in streaming multi-armed bandits (MABs) has been studied extensively, and recent work has shown that algorithms with $o(K)$ memory have to incur $\Omega(T^{2/3})$ regret, where $K$ and $T$ are the numbers of arms and trials. However, the previous best regret upper bound is still $O(K^{1/3} T^{2/3}\log^{1/3}(T))$, which is achieved by the simple uniform exploration algorithm. In this paper, we close this gap and complete the picture of regret minimization in single-pass streaming MABs. We first improve the regret lower bound to $\Omega(K^{1/3}T^{2/3})$ for algorithms with $o(K)$ memory. We then show that the $\log^{1/3}(T)$ factor is not necessary by designing algorithms with at most $O(\log^*(K))$-arm memory and achieve $O(K^{1/3}T^{2/3})$ expected regret based on streaming $\varepsilon$-best arm algorithms. We further tested the empirical performances of our algorithms on simulated MABs instances, where the proposed algorithms outperform the benchmark uniform exploration algorithm by a large margin and, on occasion, reduce the regret by up to 70%.

Calibrating deep learning models to yield uncertainty-aware predictions is crucial as deep neural networks get increasingly deployed in safety-critical applications. While existing post-hoc calibration methods achieve impressive results on in-domain test datasets, they are limited by their inability to yield reliable uncertainty estimates in domain-shift and out-of-domain (OOD) scenarios. We aim to bridge this gap by proposing DAC, an accuracy-preserving as well as Density-Aware Calibration method based on k-nearest-neighbors (KNN). In contrast to existing post-hoc methods, we utilize hidden layers of classifiers as a source for uncertainty-related information and study their importance. We show that DAC is a generic method that can readily be combined with state-of-the-art post-hoc methods. DAC boosts the robustness of calibration performance in domain-shift and OOD, while maintaining excellent in-domain predictive uncertainty estimates. We demonstrate that DAC leads to consistently better calibration across a large number of model architectures, datasets, and metrics. Additionally, we show that DAC improves calibration substantially on recent large-scale neural networks pre-trained on vast amounts of data.

The latest two-stage optimization framework based on graph neural network (GNN) and large neighborhood search (LNS) is the most popular framework in solving large-scale integer programs (IPs). However, the framework can not effectively use the embedding spatial information in GNN and still highly relies on large-scale solvers in LNS, resulting in the scale of IP being limited by the ability of the current solver and performance bottlenecks. To handle these issues, this paper presents a GNN&GBDT-guided fast optimizing framework for large-scale IPs that only uses a small-scale optimizer to solve large-scale IPs efficiently. Specifically, the proposed framework can be divided into three stages: Multi-task GNN Embedding to generate the embedding space, GBDT Prediction to effectively use the embedding spatial information, and Neighborhood Optimization to solve large-scale problems fast using the small-scale optimizer. Extensive experiments show that the proposed framework can solve IPs with millions of scales and surpass SCIP and Gurobi in the specified wall-clock time using only a small-scale optimizer with 30% of the problem size. It also shows that the proposed framework can save 99% of running time in achieving the same solution quality as SCIP, which verifies the effectiveness and efficiency of the proposed framework in solving large-scale IPs.

Neural network approaches for meta-learning distributions over functions have desirable properties such as increased flexibility and a reduced complexity of inference. Building on the successes of denoising diffusion models for generative modelling, we propose Neural Diffusion Processes (NDPs), a novel approach that learns to sample from a rich distribution over functions through its finite marginals. By introducing a custom attention block we are able to incorporate properties of stochastic processes, such as exchangeability, directly into the NDP's architecture. We empirically show that NDPs can capture functional distributions close to the true Bayesian posterior, demonstrating that they can successfully emulate the behaviour of Gaussian processes and surpass the performance of neural processes. NDPs enable a variety of downstream tasks, including regression, implicit hyperparameter marginalisation, non-Gaussian posterior prediction and global optimisation.

We develop a general theory to optimize the frequentist regret for sequential learning problems, where efficient bandit and reinforcement learning algorithms can be derived from unified Bayesian principles. We propose a novel optimization approach to create "algorithmic beliefs" at each round, and use Bayesian posteriors to make decisions. This is the first approach to make Bayesian-type algorithms prior-free and applicable to adversarial settings, in a generic and optimal manner. Moreover, the algorithms are simple and often efficient to implement. As a major application, we present a novel algorithm for multi-armed bandits that achieves the "best-of-all-worlds" empirical performance in the stochastic, adversarial, and non-stationary environments. And we illustrate how these principles can be used in linear bandits, convex bandits, and reinforcement learning.

Two main challenges in Reinforcement Learning (RL) are designing appropriate reward functions and ensuring the safety of the learned policy. To address these challenges, we present a theoretical framework for Inverse Reinforcement Learning (IRL) in constrained Markov decision processes. From a convex-analytic perspective, we extend prior results on reward identifiability and generalizability to both the constrained setting and a more general class of regularizations. In particular, we show that identifiability up to potential shaping (Cao et al., 2021) is a consequence of entropy regularization and may generally no longer hold for other regularizations or in the presence of safety constraints. We also show that to ensure generalizability to new transition laws and constraints, the true reward must be identified up to a constant. Additionally, we derive a finite sample guarantee for the suboptimality of the learned rewards, and validate our results in a gridworld environment.

We study linear contextual bandits in the misspecified setting, where the expected reward function can be approximated by a linear function class up to a bounded misspecification level $\zeta>0$. We propose an algorithm based on a novel data selection scheme, which only selects the contextual vectors with large uncertainty for online regression. We show that, when the misspecification level $\zeta$ is dominated by $\tilde O(\Delta / \sqrt{d})$ with $\Delta$ being the minimal sub-optimality gap and $d$ being the dimension of the contextual vectors, our algorithm enjoys the same gap-dependent regret bound $\tilde O ({d^2} /{\Delta})$ as in the well-specified setting up to logarithmic factors. Given this result, we show that the existing SupLinUCB algorithm (Chu et al., 2011) can also achieve a gap-dependent constant regret bound without the knowledge of sub-optimality gap $\Delta$. Together with a lower bound adapted from Lattimore et al. (2020), our result suggests an interplay between the misspecification level and the sub-optimality gap: (1) the linear contextual bandit model is efficiently learnable when $\zeta \leq \tilde O({\Delta} / \sqrt{d})$; and (2) it is not efficiently learnable when $\zeta \geq \tilde \Omega({\Delta} / {\sqrt{d}})$. Experiments on both synthetic and real-world datasets corroborate our theoretical results.

Domain randomization (DR) is widely used in reinforcement learning (RL) to bridge the gap between simulation and reality by maximizing its average returns under the perturbation of environmental parameters. However, even the most complex simulators cannot capture all details in reality due to finite domain parameters and simplified physical models. Additionally, the existing methods often assume that the distribution of domain parameters belongs to a specific family of probability functions, such as normal distributions, which may not be correct. To overcome these limitations, we propose a new approach to DR by rethinking it from the perspective of adversarial state perturbation, without the need for reconfiguring the simulator or relying on prior knowledge about the environment. We also address the issue of over-conservatism that can occur when perturbing agents to the worst states during training by introducing a Relaxed State-Adversarial Algorithm that simultaneously maximizes the average-case and worst-case returns. We evaluate our method by comparing it to state-of-the-art methods, providing experimental results and theoretical proofs to verify its effectiveness. Our source code and appendix are available at https://github.com/sophialien/RAPPO.

Collections of probability distributions arise in a variety of applications ranging from user activity pattern analysis to brain connectomics. In practice these distributions can be defined over diverse domain types including finite intervals, circles, cylinders, spheres, other manifolds, and graphs. This paper introduces an approach for detecting differences between two collections of distributions over such general domains. To this end, we propose the intrinsic slicing construction that yields a novel class of Wasserstein distances on manifolds and graphs. These distances are Hilbert embeddable, allowing us to reduce the distribution collection comparison problem to a more familiar mean testing problem in a Hilbert space. We provide two testing procedures one based on resampling and another on combining p-values from coordinate-wise tests. Our experiments in various synthetic and real data settings show that the resulting tests are powerful and the p-values are well-calibrated.