Machine Learning (cs.LG)

Abstract: Federated Learning (FL) has emerged as an effective learning paradigm for distributed computation owing to its strong potential in capturing underlying data statistics while preserving data privacy. However, in cases of practical data heterogeneity among FL clients, existing FL frameworks still exhibit deficiency in capturing the overall feature properties of local client data that exhibit disparate distributions. In response, generative adversarial networks (GANs) have recently been exploited in FL to address data heterogeneity since GANs can be integrated for data regeneration without exposing original raw data. Despite some successes, existing GAN-related FL frameworks often incur heavy communication cost and also elicit other privacy concerns, which limit their applications in real scenarios. To this end, this work proposes a novel FL framework that requires only partial GAN model sharing. Named as PS-FedGAN, this new framework enhances the GAN releasing and training mechanism to address heterogeneous data distributions across clients and to strengthen privacy preservation at reduced communication cost, especially over wireless networks. Our analysis demonstrates the convergence and privacy benefits of the proposed PS-FEdGAN framework. Through experimental results based on several well-known benchmark datasets, our proposed PS-FedGAN shows great promise to tackle FL under non-IID client data distributions, while securing data privacy and lowering communication overhead.

Abstract: This study aims to solve the over-reliance on the rank estimation strategy in the standard tensor factorization-based tensor recovery and the problem of a large computational cost in the standard t-SVD-based tensor recovery. To this end, we proposes a new tensor norm with a dual low-rank constraint, which utilizes the low-rank prior and rank information at the same time. In the proposed tensor norm, a series of surrogate functions of the tensor tubal rank can be used to achieve better performance in harness low-rankness within tensor data. It is proven theoretically that the resulting tensor completion model can effectively avoid performance degradation caused by inaccurate rank estimation. Meanwhile, attributed to the proposed dual low-rank constraint, the t-SVD of a smaller tensor instead of the original big one is computed by using a sample trick. Based on this, the total cost at each iteration of the optimization algorithm is reduced to $\mathcal{O}(n^3\log n +kn^3)$ from $\mathcal{O}(n^4)$ achieved with standard methods, where $k$ is the estimation of the true tensor rank and far less than $n$. Our method was evaluated on synthetic and real-world data, and it demonstrated superior performance and efficiency over several existing state-of-the-art tensor completion methods.

Abstract: Maximum mean discrepancy (MMD) flows suffer from high computational costs in large scale computations. In this paper, we show that MMD flows with Riesz kernels $K(x,y) = - \|x-y\|^r$, $r \in (0,2)$ have exceptional properties which allow for their efficient computation. First, the MMD of Riesz kernels coincides with the MMD of their sliced version. As a consequence, the computation of gradients of MMDs can be performed in the one-dimensional setting. Here, for $r=1$, a simple sorting algorithm can be applied to reduce the complexity from $O(MN+N^2)$ to $O((M+N)\log(M+N))$ for two empirical measures with $M$ and $N$ support points. For the implementations we approximate the gradient of the sliced MMD by using only a finite number $P$ of slices. We show that the resulting error has complexity $O(\sqrt{d/P})$, where $d$ is the data dimension. These results enable us to train generative models by approximating MMD gradient flows by neural networks even for large scale applications. We demonstrate the efficiency of our model by image generation on MNIST, FashionMNIST and CIFAR10.

Abstract: Generalized Additive Models (GAMs) have recently experienced a resurgence in popularity due to their interpretability, which arises from expressing the target value as a sum of non-linear transformations of the features. Despite the current enthusiasm for GAMs, their susceptibility to concurvity - i.e., (possibly non-linear) dependencies between the features - has hitherto been largely overlooked. Here, we demonstrate how concurvity can severly impair the interpretability of GAMs and propose a remedy: a conceptually simple, yet effective regularizer which penalizes pairwise correlations of the non-linearly transformed feature variables. This procedure is applicable to any differentiable additive model, such as Neural Additive Models or NeuralProphet, and enhances interpretability by eliminating ambiguities due to self-canceling feature contributions. We validate the effectiveness of our regularizer in experiments on synthetic as well as real-world datasets for time-series and tabular data. Our experiments show that concurvity in GAMs can be reduced without significantly compromising prediction quality, improving interpretability and reducing variance in the feature importances.

Abstract: Among the great successes of Reinforcement Learning (RL), self-play algorithms play an essential role in solving competitive games. Current self-play algorithms optimize the agent to maximize expected win-rates against its current or historical copies, making it often stuck in the local optimum and its strategy style simple and homogeneous. A possible solution is to improve the diversity of policies, which helps the agent break the stalemate and enhances its robustness when facing different opponents. However, enhancing diversity in the self-play algorithms is not trivial. In this paper, we aim to introduce diversity from the perspective that agents could have diverse risk preferences in the face of uncertainty. Specifically, we design a novel reinforcement learning algorithm called Risk-sensitive Proximal Policy Optimization (RPPO), which smoothly interpolates between worst-case and best-case policy learning and allows for policy learning with desired risk preferences. Seamlessly integrating RPPO with population-based self-play, agents in the population optimize dynamic risk-sensitive objectives with experiences from playing against diverse opponents. Empirical results show that our method achieves comparable or superior performance in competitive games and that diverse modes of behaviors emerge. Our code is public online at \url{https://github.com/Jackory/RPBT}.

Abstract: Incomplete multi-view clustering is a challenging and non-trivial task to provide effective data analysis for large amounts of unlabeled data in the real world. All incomplete multi-view clustering methods need to address the problem of how to reduce the impact of missing views. To address this issue, we propose diffusion completion to recover the missing views integrated into an incomplete multi-view clustering framework. Based on the observable views information, the diffusion model is used to recover the missing views, and then the consistency information of the multi-view data is learned by contrastive learning to improve the performance of multi-view clustering. To the best of our knowledge, this may be the first work to incorporate diffusion models into an incomplete multi-view clustering framework. Experimental results show that the proposed method performs well in recovering the missing views while achieving superior clustering performance compared to state-of-the-art methods.

Abstract: Within the tensor singular value decomposition (T-SVD) framework, existing robust low-rank tensor completion approaches have made great achievements in various areas of science and engineering. Nevertheless, these methods involve the T-SVD based low-rank approximation, which suffers from high computational costs when dealing with large-scale tensor data. Moreover, most of them are only applicable to third-order tensors. Against these issues, in this article, two efficient low-rank tensor approximation approaches fusing randomized techniques are first devised under the order-d (d >= 3) T-SVD framework. On this basis, we then further investigate the robust high-order tensor completion (RHTC) problem, in which a double nonconvex model along with its corresponding fast optimization algorithms with convergence guarantees are developed. To the best of our knowledge, this is the first study to incorporate the randomized low-rank approximation into the RHTC problem. Empirical studies on large-scale synthetic and real tensor data illustrate that the proposed method outperforms other state-of-the-art approaches in terms of both computational efficiency and estimated precision.

Abstract: Random Search is one of the most widely-used method for Hyperparameter Optimization, and is critical to the success of deep learning models. Despite its astonishing performance, little non-heuristic theory has been developed to describe the underlying working mechanism. This paper gives a theoretical accounting of Random Search. We introduce the concept of \emph{scattering dimension} that describes the landscape of the underlying function, and quantifies the performance of random search. We show that, when the environment is noise-free, the output of random search converges to the optimal value in probability at rate $ \widetilde{\mathcal{O}} \left( \left( \frac{1}{T} \right)^{ \frac{1}{d_s} } \right) $, where $ d_s \ge 0 $ is the scattering dimension of the underlying function. When the observed function values are corrupted by bounded $iid$ noise, the output of random search converges to the optimal value in probability at rate $ \widetilde{\mathcal{O}} \left( \left( \frac{1}{T} \right)^{ \frac{1}{d_s + 1} } \right) $. In addition, based on the principles of random search, we introduce an algorithm, called BLiN-MOS, for Lipschitz bandits in doubling metric spaces that are also emdowed with a Borel measure, and show that BLiN-MOS achieves a regret rate of order $ \widetilde{\mathcal{O}} \left( T^{ \frac{d_z}{d_z + 1} } \right) $, where $d_z$ is the zooming dimension of the problem instance. Our results show that in metric spaces with a Borel measure, the classic theory of Lipschitz bandits can be improved. This result suggests an intrinsic axiomatic gap between metric spaces and metric measure spaces from an algorithmic perspective, since the upper bound in a metric measure space breaks the known information-theoretical lower bounds for Lipschitz bandits in a metric space with no measure structure.

Abstract: Disentangled representation learning is a challenging task that involves separating multiple factors of variation in complex data. Although various metrics for learning and evaluating disentangled representations have been proposed, it remains unclear what these metrics truly quantify and how to compare them. In this work, we study the definitions of disentanglement given by first-order equational predicates and introduce a systematic approach for transforming an equational definition into a compatible quantitative metric based on enriched category theory. Specifically, we show how to replace (i) equality with metric or divergence, (ii) logical connectives with order operations, (iii) universal quantifier with aggregation, and (iv) existential quantifier with the best approximation. Using this approach, we derive metrics for measuring the desired properties of a disentangled representation extractor and demonstrate their effectiveness on synthetic data. Our proposed approach provides practical guidance for researchers in selecting appropriate evaluation metrics and designing effective learning algorithms for disentangled representation learning.

Abstract: The safe and stable operation of power systems is greatly challenged by the high variability and randomness of wind power in large-scale wind-power-integrated grids. Wind power forecasting is an effective solution to tackle this issue, with wind speed forecasting being an essential aspect. In this paper, a Graph-attentive Frequency-enhanced Spatial-Temporal Wind Speed Forecasting model based on graph attention and frequency-enhanced mechanisms, i.e., GFST-WSF, is proposed to improve the accuracy of short-term wind speed forecasting. The GFST-WSF comprises a Transformer architecture for temporal feature extraction and a Graph Attention Network (GAT) for spatial feature extraction. The GAT is specifically designed to capture the complex spatial dependencies among wind speed stations to effectively aggregate information from neighboring nodes in the graph, thus enhancing the spatial representation of the data. To model the time lag in wind speed correlation between adjacent wind farms caused by geographical factors, a dynamic complex adjacency matrix is formulated and utilized by the GAT. Benefiting from the effective spatio-temporal feature extraction and the deep architecture of the Transformer, the GFST-WSF outperforms other baselines in wind speed forecasting for the 6-24 hours ahead forecast horizon in case studies.

Abstract: Temporally indexed data are essential in a wide range of fields and of interest to machine learning researchers. Time series data, however, are often scarce or highly sensitive, which precludes the sharing of data between researchers and industrial organizations and the application of existing and new data-intensive ML methods. A possible solution to this bottleneck is to generate synthetic data. In this work, we introduce Time Series Generative Modeling (TSGM), an open-source framework for the generative modeling of synthetic time series. TSGM includes a broad repertoire of machine learning methods: generative models, probabilistic, and simulator-based approaches. The framework enables users to evaluate the quality of the produced data from different angles: similarity, downstream effectiveness, predictive consistency, diversity, and privacy. The framework is extensible, which allows researchers to rapidly implement their own methods and compare them in a shareable environment. TSGM was tested on open datasets and in production and proved to be beneficial in both cases. Additionally to the library, the project allows users to employ command line interfaces for synthetic data generation which lowers the entry threshold for those without a programming background.

Abstract: In federated learning (FL), a cluster of local clients are chaired under the coordination of the global server and cooperatively train one model with privacy protection. Due to the multiple local updates and the isolated non-iid dataset, clients are prone to overfit into their own optima, which extremely deviates from the global objective and significantly undermines the performance. Most previous works only focus on enhancing the consistency between the local and global objectives to alleviate this prejudicial client drifts from the perspective of the optimization view, whose performance would be prominently deteriorated on the high heterogeneity. In this work, we propose a novel and general algorithm {\ttfamily FedSMOO} by jointly considering the optimization and generalization targets to efficiently improve the performance in FL. Concretely, {\ttfamily FedSMOO} adopts a dynamic regularizer to guarantee the local optima towards the global objective, which is meanwhile revised by the global Sharpness Aware Minimization (SAM) optimizer to search for the consistent flat minima. Our theoretical analysis indicates that {\ttfamily FedSMOO} achieves fast $\mathcal{O}(1/T)$ convergence rate with low generalization bound. Extensive numerical studies are conducted on the real-world dataset to verify its peerless efficiency and excellent generality.

Abstract: Conventional Gaussian process regression exclusively assumes the existence of noise in the output data of model observations. In many scientific and engineering applications, however, the input locations of observational data may also be compromised with uncertainties owing to modeling assumptions, measurement errors, etc. In this work, we propose a Bayesian method that integrates the variability of input data into Gaussian process regression. Considering two types of observables -- noise-corrupted outputs with fixed inputs and those with prior-distribution-defined uncertain inputs, a posterior distribution is estimated via a Bayesian framework to infer the uncertain data locations. Thereafter, such quantified uncertainties of inputs are incorporated into Gaussian process predictions by means of marginalization. The effectiveness of this new regression technique is demonstrated through several numerical examples, in which a consistently good performance of generalization is observed, while a substantial reduction in the predictive uncertainties is achieved by the Bayesian inference of uncertain inputs.

Abstract: Model substructure learning aims to find an invariant network substructure that can have better out-of-distribution (OOD) generalization than the original full structure. Existing works usually search the invariant substructure using modular risk minimization (MRM) with fully exposed out-domain data, which may bring about two drawbacks: 1) Unfairness, due to the dependence of the full exposure of out-domain data; and 2) Sub-optimal OOD generalization, due to the equally feature-untargeted pruning on the whole data distribution. Based on the idea that in-distribution (ID) data with spurious features may have a lower experience risk, in this paper, we propose a novel Spurious Feature-targeted model Pruning framework, dubbed SFP, to automatically explore invariant substructures without referring to the above drawbacks. Specifically, SFP identifies spurious features within ID instances during training using our theoretically verified task loss, upon which, SFP attenuates the corresponding feature projections in model space to achieve the so-called spurious feature-targeted pruning. This is typically done by removing network branches with strong dependencies on identified spurious features, thus SFP can push the model learning toward invariant features and pull that out of spurious features and devise optimal OOD generalization. Moreover, we also conduct detailed theoretical analysis to provide the rationality guarantee and a proof framework for OOD structures via model sparsity, and for the first time, reveal how a highly biased data distribution affects the model's OOD generalization. Experiments on various OOD datasets show that SFP can significantly outperform both structure-based and non-structure-based OOD generalization SOTAs, with accuracy improvement up to 4.72% and 23.35%, respectively

Abstract: This paper studies the open problem of conformalized entry prediction in a row/column-exchangeable matrix. The matrix setting presents novel and unique challenges, but there exists little work on this interesting topic. We meticulously define the problem, differentiate it from closely related problems, and rigorously delineate the boundary between achievable and impossible goals. We then propose two practical algorithms. The first method provides a fast emulation of the full conformal prediction, while the second method leverages the technique of algorithmic stability for acceleration. Both methods are computationally efficient and can effectively safeguard coverage validity in presence of arbitrary missing pattern. Further, we quantify the impact of missingness on prediction accuracy and establish fundamental limit results. Empirical evidence from synthetic and real-world data sets corroborates the superior performance of our proposed methods.

Abstract: Machine learning (ML) has revolutionized transportation systems, enabling autonomous driving and smart traffic services. Federated learning (FL) overcomes privacy constraints by training ML models in distributed systems, exchanging model parameters instead of raw data. However, the dynamic states of connected vehicles affect the network connection quality and influence the FL performance. To tackle this challenge, we propose a contextual client selection pipeline that uses Vehicle-to-Everything (V2X) messages to select clients based on the predicted communication latency. The pipeline includes: (i) fusing V2X messages, (ii) predicting future traffic topology, (iii) pre-clustering clients based on local data distribution similarity, and (iv) selecting clients with minimal latency for future model aggregation. Experiments show that our pipeline outperforms baselines on various datasets, particularly in non-iid settings.

Abstract: This commentary tests a methodology proposed by Munk et al. (2022) for using failed predictions in machine learning as a method to identify ambiguous and rich cases for qualitative analysis. Using a dataset describing actions performed by fictional characters interacting with machine vision technologies in 500 artworks, movies, novels and videogames, I trained a simple machine learning algorithm (using the kNN algorithm in R) to predict whether or not an action was active or passive using only information about the fictional characters. Predictable actions were generally unemotional and unambiguous activities where machine vision technologies were treated as simple tools. Unpredictable actions, that is, actions that the algorithm could not correctly predict, were more ambivalent and emotionally loaded, with more complex power relationships between characters and technologies. The results thus support Munk et al.'s theory that failed predictions can be productively used to identify rich cases for qualitative analysis. This test goes beyond simply replicating Munk et al.'s results by demonstrating that the method can be applied to a broader humanities domain, and that it does not require complex neural networks but can also work with a simpler machine learning algorithm. Further research is needed to develop an understanding of what kinds of data the method is useful for and which kinds of machine learning are most generative. To support this, the R code required to produce the results is included so the test can be replicated. The code can also be reused or adapted to test the method on other datasets.

Abstract: Apart from the high accuracy of machine learning models, what interests many researchers in real-life problems (e.g., fraud detection, credit scoring) is to find hidden patterns in data; particularly when dealing with their challenging imbalanced characteristics. Interpretability is also a key requirement that needs to accompany the used machine learning model. In this concern, often, intrinsically interpretable models are preferred to complex ones, which are in most cases black-box models. Also, linear models are used in some high-risk fields to handle tabular data, even if performance must be sacrificed. In this paper, we introduce Self-Reinforcement Attention (SRA), a novel attention mechanism that provides a relevance of features as a weight vector which is used to learn an intelligible representation. This weight is then used to reinforce or reduce some components of the raw input through element-wise vector multiplication. Our results on synthetic and real-world imbalanced data show that our proposed SRA block is effective in end-to-end combination with baseline models.

Abstract: Identifying molecules that exhibit some pre-specified properties is a difficult problem to solve. In the last few years, deep generative models have been used for molecule generation. Deep Graph Variational Autoencoders are among the most powerful machine learning tools with which it is possible to address this problem. However, existing methods struggle in capturing the true data distribution and tend to be computationally expensive. In this work, we propose RGCVAE, an efficient and effective Graph Variational Autoencoder based on: (i) an encoding network exploiting a new powerful Relational Graph Isomorphism Network; (ii) a novel probabilistic decoding component. Compared to several state-of-the-art VAE methods on two widely adopted datasets, RGCVAE shows state-of-the-art molecule generation performance while being significantly faster to train.

Abstract: Projection-free online learning has drawn increasing interest due to its efficiency in solving high-dimensional problems with complicated constraints. However, most existing projection-free online methods focus on minimizing the static regret, which unfortunately fails to capture the challenge of changing environments. In this paper, we investigate non-stationary projection-free online learning, and choose dynamic regret and adaptive regret to measure the performance. Specifically, we first provide a novel dynamic regret analysis for an existing projection-free method named $\text{BOGD}_\text{IP}$, and establish an $\mathcal{O}(T^{3/4}(1+P_T))$ dynamic regret bound, where $P_T$ denotes the path-length of the comparator sequence. Then, we improve the upper bound to $\mathcal{O}(T^{3/4}(1+P_T)^{1/4})$ by running multiple $\text{BOGD}_\text{IP}$ algorithms with different step sizes in parallel, and tracking the best one on the fly. Our results are the first general-case dynamic regret bounds for projection-free online learning, and can recover the existing $\mathcal{O}(T^{3/4})$ static regret by setting $P_T = 0$. Furthermore, we propose a projection-free method to attain an $\tilde{\mathcal{O}}(\tau^{3/4})$ adaptive regret bound for any interval with length $\tau$, which nearly matches the static regret over that interval. The essential idea is to maintain a set of $\text{BOGD}_\text{IP}$ algorithms dynamically, and combine them by a meta algorithm. Moreover, we demonstrate that it is also equipped with an $\mathcal{O}(T^{3/4}(1+P_T)^{1/4})$ dynamic regret bound. Finally, empirical studies verify our theoretical findings.

Abstract: Monitoring the health status of patients and predicting mortality in advance is vital for providing patients with timely care and treatment. Massive medical signs in electronic health records (EHR) are fitted into advanced machine learning models to make predictions. However, the data-quality problem of original clinical signs is less discussed in the literature. Based on an in-depth measurement of the missing rate and correlation score across various medical signs and a large amount of patient hospital admission records, we discovered the comprehensive missing rate is extremely high, and a large number of useless signs could hurt the performance of prediction models. Then we concluded that only improving data-quality could improve the baseline accuracy of different prediction algorithms. We designed MEDLENS, with an automatic vital medical signs selection approach via statistics and a flexible interpolation approach for high missing rate time series. After augmenting the data-quality of original medical signs, MEDLENS applies ensemble classifiers to boost the accuracy and reduce the computation overhead at the same time. It achieves a very high accuracy performance of 0.96% AUC-ROC and 0.81% AUC-PR, which exceeds the previous benchmark.

Abstract: Inferring a state sequence from a sequence of measurements is a fundamental problem in bioinformatics and natural language processing. The Viterbi and the Beam Search (BS) algorithms are popular inference methods, but they have limitations when applied to Hierarchical Hidden Markov Models (HHMMs), where the interest lies in the outer state sequence. The Viterbi algorithm can not infer outer states without inner states, while the BS algorithm requires marginalization over prohibitively large state spaces. We propose two new algorithms to overcome these limitations: the greedy marginalized BS algorithm and the local focus BS algorithm. We show that they approximate the most likely outer state sequence with higher performance than the Viterbi algorithm, and we evaluate the performance of these algorithms on an explicit duration HMM with simulation and nanopore base calling data.

Abstract: We give the first tester-learner for halfspaces that succeeds universally over a wide class of structured distributions. Our universal tester-learner runs in fully polynomial time and has the following guarantee: the learner achieves error $O(\mathrm{opt}) + \epsilon$ on any labeled distribution that the tester accepts, and moreover, the tester accepts whenever the marginal is any distribution that satisfies a Poincar\'e inequality. In contrast to prior work on testable learning, our tester is not tailored to any single target distribution but rather succeeds for an entire target class of distributions. The class of Poincar\'e distributions includes all strongly log-concave distributions, and, assuming the Kannan--L\'{o}vasz--Simonovits (KLS) conjecture, includes all log-concave distributions. In the special case where the label noise is known to be Massart, our tester-learner achieves error $\mathrm{opt} + \epsilon$ while accepting all log-concave distributions unconditionally (without assuming KLS). Our tests rely on checking hypercontractivity of the unknown distribution using a sum-of-squares (SOS) program, and crucially make use of the fact that Poincar\'e distributions are certifiably hypercontractive in the SOS framework.

Abstract: Recent research has observed that in machine learning optimization, gradient descent (GD) often operates at the edge of stability (EoS) [Cohen, et al., 2021], where the stepsizes are set to be large, resulting in non-monotonic losses induced by the GD iterates. This paper studies the convergence and implicit bias of constant-stepsize GD for logistic regression on linearly separable data in the EoS regime. Despite the presence of local oscillations, we prove that the logistic loss can be minimized by GD with any constant stepsize over a long time scale. Furthermore, we prove that with any constant stepsize, the GD iterates tend to infinity when projected to a max-margin direction (the hard-margin SVM direction) and converge to a fixed vector that minimizes a strongly convex potential when projected to the orthogonal complement of the max-margin direction. In contrast, we also show that in the EoS regime, GD iterates may diverge catastrophically under the exponential loss, highlighting the superiority of the logistic loss. These theoretical findings are in line with numerical simulations and complement existing theories on the convergence and implicit bias of GD, which are only applicable when the stepsizes are sufficiently small.

Abstract: We provide the first polynomial-time convergence guarantees for the probability flow ODE implementation (together with a corrector step) of score-based generative modeling. Our analysis is carried out in the wake of recent results obtaining such guarantees for the SDE-based implementation (i.e., denoising diffusion probabilistic modeling or DDPM), but requires the development of novel techniques for studying deterministic dynamics without contractivity. Through the use of a specially chosen corrector step based on the underdamped Langevin diffusion, we obtain better dimension dependence than prior works on DDPM ($O(\sqrt{d})$ vs. $O(d)$, assuming smoothness of the data distribution), highlighting potential advantages of the ODE framework.

Abstract: The Private Aggregation of Teacher Ensembles (PATE) is a machine learning framework that enables the creation of private models through the combination of multiple "teacher" models and a "student" model. The student model learns to predict an output based on the voting of the teachers, and the resulting model satisfies differential privacy. PATE has been shown to be effective in creating private models in semi-supervised settings or when protecting data labels is a priority. This paper explores whether the use of PATE can result in unfairness, and demonstrates that it can lead to accuracy disparities among groups of individuals. The paper also analyzes the algorithmic and data properties that contribute to these disproportionate impacts, why these aspects are affecting different groups disproportionately, and offers recommendations for mitigating these effects

Abstract: This work presents a comprehensive analysis to regularize the Soft Actor-Critic (SAC) algorithm with automatic temperature adjustment. The the policy evaluation, the policy improvement and the temperature adjustment are reformulated, addressing certain modification and enhancing the clarity of the original theory in a more explicit manner.

Abstract: The progress of autonomous web navigation has been hindered by the dependence on billions of exploratory interactions via online reinforcement learning, and domain-specific model designs that make it difficult to leverage generalization from rich out-of-domain data. In this work, we study data-driven offline training for web agents with vision-language foundation models. We propose an instruction-following multimodal agent, WebGUM, that observes both webpage screenshots and HTML pages and outputs web navigation actions, such as click and type. WebGUM is trained by jointly finetuning an instruction-finetuned language model and a vision transformer on a large corpus of demonstrations. We empirically demonstrate this recipe improves the agent's ability of grounded visual perception, HTML comprehension and multi-step reasoning, outperforming prior works by a significant margin. On the MiniWoB benchmark, we improve over the previous best offline methods by more than 31.9%, being close to reaching online-finetuned SoTA. On the WebShop benchmark, our 3-billion-parameter model achieves superior performance to the existing SoTA, PaLM-540B. We also collect 347K high-quality demonstrations using our trained models, 38 times larger than prior work, and make them available to promote future research in this direction.