Machine Learning (cs.LG)

Abstract: A unique feature of Recurrent Neural Networks (RNNs) is that it incrementally processes input sequences. In this research, we aim to uncover the inherent generalization properties, i.e., inductive bias, of RNNs with respect to how frequently RNNs switch the outputs through time steps in the sequence classification task, which we call output sequence frequency. Previous work analyzed inductive bias by training models with a few synthetic data and comparing the model's generalization with candidate generalization patterns. However, when examining the output sequence frequency, previous methods cannot be directly applied since enumerating candidate patterns is computationally difficult for longer sequences. To this end, we propose to directly calculate the output sequence frequency for each model by regarding the outputs of the model as discrete-time signals and applying frequency domain analysis. Experimental results showed that Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) have an inductive bias towards lower-frequency patterns, while Elman RNN tends to learn patterns in which the output changes at high frequencies. We also found that the inductive bias of LSTM and GRU varies with the number of layers and the size of hidden layers.

Abstract: Neural Ordinary Differential Equations (NODEs) probed the usage of numerical solvers to solve the differential equation characterized by a Neural Network (NN), therefore initiating a new paradigm of deep learning models with infinite depth. NODEs were designed to tackle the irregular time series problem. However, NODEs have demonstrated robustness against various noises and adversarial attacks. This paper is about the natural robustness of NODEs and examines the cause behind such surprising behaviour. We show that by controlling the Lipschitz constant of the ODE dynamics the robustness can be significantly improved. We derive our approach from Grownwall's inequality. Further, we draw parallels between contractivity theory and Grownwall's inequality. Experimentally we corroborate the enhanced robustness on numerous datasets - MNIST, CIFAR-10, and CIFAR 100. We also present the impact of adaptive and non-adaptive solvers on the robustness of NODEs.

Abstract: Accurate prediction of aerodynamic forces in real-time is crucial for autonomous navigation of unmanned aerial vehicles (UAVs). This paper presents a data-driven aerodynamic force prediction model based on a small number of pressure sensors located on the surface of UAV. The model is built on a linear term that can make a reasonably accurate prediction and a nonlinear correction for accuracy improvement. The linear term is based on a reduced basis reconstruction of the surface pressure distribution, where the basis is extracted from numerical simulation data and the basis coefficients are determined by solving linear pressure reconstruction equations at a set of sensor locations. Sensor placement is optimized using the discrete empirical interpolation method (DEIM). Aerodynamic forces are computed by integrating the reconstructed surface pressure distribution. The nonlinear term is an artificial neural network (NN) that is trained to bridge the gap between the ground truth and the DEIM prediction, especially in the scenario where the DEIM model is constructed from simulation data with limited fidelity. A large network is not necessary for accurate correction as the linear model already captures the main dynamics of the surface pressure field, thus yielding an efficient DEIM+NN aerodynamic force prediction model. The model is tested on numerical and experimental dynamic stall data of a 2D NACA0015 airfoil, and numerical simulation data of dynamic stall of a 3D drone. Numerical results demonstrate that the machine learning enhanced model can make fast and accurate predictions of aerodynamic forces using only a few pressure sensors, even for the NACA0015 case in which the simulations do not agree well with the wind tunnel experiments. Furthermore, the model is robust to noise.

Abstract: Feature attribution aims to explain the reasoning behind a black-box model's prediction by identifying the impact of each feature on the prediction. Recent work has extended feature attribution to interactions between multiple features. However, the lack of a unified framework has led to a proliferation of methods that are often not directly comparable. This paper introduces a parameterized attribution framework -- the Weighted M\"obius Score -- and (i) shows that many different attribution methods for both individual features and feature interactions are special cases and (ii) identifies some new methods. By studying the vector space of attribution methods, our framework utilizes standard linear algebra tools and provides interpretations in various fields, including cooperative game theory and causal mediation analysis. We empirically demonstrate the framework's versatility and effectiveness by applying these attribution methods to feature interactions in sentiment analysis and chain-of-thought prompting.

Abstract: Counterfactual outcome prediction in longitudinal data has recently gained attention due to its potential applications in healthcare and social sciences. In this paper, we explore the use of the state space model, a popular sequence model, for this task. Specifically, we compare the performance of two models: Treatment Effect Neural Controlled Differential Equation (TE-CDE) and structured state space model (S4Model). While TE-CDE uses controlled differential equations to address time-dependent confounding, it suffers from optimization issues and slow training. In contrast, S4Model is more efficient at modeling long-range dependencies and easier to train. We evaluate the models on a simulated lung tumor growth dataset and find that S4Model outperforms TE-CDE with 1.63x reduction in per epoch training time and 10x better normalized mean squared error. Additionally, S4Model is more stable during training and less sensitive to weight initialization than TE-CDE. Our results suggest that the state space model may be a promising approach for counterfactual outcome prediction in longitudinal data, with S4Model offering a more efficient and effective alternative to TE-CDE.

Abstract: This study presents a novel approach for touch sensing using semi-elastic textile surfaces that does not require the placement of additional sensors in the sensing area, instead relying on sensors located on the border of the textile. The proposed approach is demonstrated through experiments involving an elastic Jersey fabric and a variety of machine-learning models. The performance of one particular border-based sensor design is evaluated in depth. By using visual markers, the best-performing visual sensor arrangement predicts a single touch point with a mean squared error of 1.36 mm on an area of 125mm by 125mm. We built a textile only prototype that is able to classify touch at three indent levels (0, 15, and 20 mm) with an accuracy of 82.85%. Our results suggest that this approach has potential applications in wearable technology and smart textiles, making it a promising avenue for further exploration in these fields.

Abstract: Generating synthetic data through generative models is gaining interest in the ML community and beyond, promising a future where datasets can be tailored to individual needs. Unfortunately, synthetic data is usually not perfect, resulting in potential errors in downstream tasks. In this work we explore how the generative process affects the downstream ML task. We show that the naive synthetic data approach -- using synthetic data as if it is real -- leads to downstream models and analyses that do not generalize well to real data. As a first step towards better ML in the synthetic data regime, we introduce Deep Generative Ensemble (DGE) -- a framework inspired by Deep Ensembles that aims to implicitly approximate the posterior distribution over the generative process model parameters. DGE improves downstream model training, evaluation, and uncertainty quantification, vastly outperforming the naive approach on average. The largest improvements are achieved for minority classes and low-density regions of the original data, for which the generative uncertainty is largest.

Abstract: Safeguarding data from unauthorized exploitation is vital for privacy and security, especially in recent rampant research in security breach such as adversarial/membership attacks. To this end, \textit{unlearnable examples} (UEs) have been recently proposed as a compelling protection, by adding imperceptible perturbation to data so that models trained on them cannot classify them accurately on original clean distribution. Unfortunately, we find UEs provide a false sense of security, because they cannot stop unauthorized users from utilizing other unprotected data to remove the protection, by turning unlearnable data into learnable again. Motivated by this observation, we formally define a new threat by introducing \textit{learnable unauthorized examples} (LEs) which are UEs with their protection removed. The core of this approach is a novel purification process that projects UEs onto the manifold of LEs. This is realized by a new joint-conditional diffusion model which denoises UEs conditioned on the pixel and perceptual similarity between UEs and LEs. Extensive experiments demonstrate that LE delivers state-of-the-art countering performance against both supervised UEs and unsupervised UEs in various scenarios, which is the first generalizable countermeasure to UEs across supervised learning and unsupervised learning.

Abstract: We revisit the common practice of evaluating adaptation of Online Continual Learning (OCL) algorithms through the metric of online accuracy, which measures the accuracy of the model on the immediate next few samples. However, we show that this metric is unreliable, as even vacuous blind classifiers, which do not use input images for prediction, can achieve unrealistically high online accuracy by exploiting spurious label correlations in the data stream. Our study reveals that existing OCL algorithms can also achieve high online accuracy, but perform poorly in retaining useful information, suggesting that they unintentionally learn spurious label correlations. To address this issue, we propose a novel metric for measuring adaptation based on the accuracy on the near-future samples, where spurious correlations are removed. We benchmark existing OCL approaches using our proposed metric on large-scale datasets under various computational budgets and find that better generalization can be achieved by retaining and reusing past seen information. We believe that our proposed metric can aid in the development of truly adaptive OCL methods. We provide code to reproduce our results at https://github.com/drimpossible/EvalOCL.

Abstract: Time Series Classification (TSC) is an extensively researched field from which a broad range of real-world problems can be addressed obtaining excellent results. One sort of the approaches performing well are the so-called dictionary-based techniques. The Temporal Dictionary Ensemble (TDE) is the current state-of-the-art dictionary-based TSC approach. In many TSC problems we find a natural ordering in the labels associated with the time series. This characteristic is referred to as ordinality, and can be exploited to improve the methods performance. The area dealing with ordinal time series is the Time Series Ordinal Classification (TSOC) field, which is yet unexplored. In this work, we present an ordinal adaptation of the TDE algorithm, known as ordinal TDE (O-TDE). For this, a comprehensive comparison using a set of 18 TSOC problems is performed. Experiments conducted show the improvement achieved by the ordinal dictionary-based approach in comparison to four other existing nominal dictionary-based techniques.

Abstract: AI systems empowered by reinforcement learning (RL) algorithms harbor the immense potential to catalyze societal advancement, yet their deployment is often impeded by significant safety concerns. Particularly in safety-critical applications, researchers have raised concerns about unintended harms or unsafe behaviors of unaligned RL agents. The philosophy of safe reinforcement learning (SafeRL) is to align RL agents with harmless intentions and safe behavioral patterns. In SafeRL, agents learn to develop optimal policies by receiving feedback from the environment, while also fulfilling the requirement of minimizing the risk of unintended harm or unsafe behavior. However, due to the intricate nature of SafeRL algorithm implementation, combining methodologies across various domains presents a formidable challenge. This had led to an absence of a cohesive and efficacious learning framework within the contemporary SafeRL research milieu. In this work, we introduce a foundational framework designed to expedite SafeRL research endeavors. Our comprehensive framework encompasses an array of algorithms spanning different RL domains and places heavy emphasis on safety elements. Our efforts are to make the SafeRL-related research process more streamlined and efficient, therefore facilitating further research in AI safety. Our project is released at: https://github.com/PKU-Alignment/omnisafe.

Abstract: Neural networks (NNs) are capable of learning complex patterns and relationships in data to make predictions with high accuracy, making them useful for various tasks. However, NNs are both computation-intensive and memory-intensive methods, making them challenging for edge applications. To accelerate the most common operations (matrix-vector multiplication) in NNs, hardware accelerator architectures such as computation-in-memory (CiM) with non-volatile memristive crossbars are utilized. Although they offer benefits such as power efficiency, parallelism, and nonvolatility, they suffer from various faults and variations, both during manufacturing and lifetime operations. This can lead to faulty computations and, in turn, degradation of post-mapping inference accuracy, which is unacceptable for many applications, including safety-critical applications. Therefore, proper testing of NN hardware accelerators is required. In this paper, we propose a \emph{one-shot} testing approach that can test NNs accelerated on memristive crossbars with only one test vector, making it very suitable for online testing applications. Our approach can consistently achieve $100\%$ fault coverage across several large topologies with up to $201$ layers and challenging tasks like semantic segmentation. Nevertheless, compared to existing methods, the fault coverage is improved by up to $24\%$, the memory overhead is only $0.0123$ MB, a reduction of up to $19980\times$ and the number of test vectors is reduced by $10000\times$.

Abstract: The recently-developed infant wearable MAIJU provides a means to automatically evaluate infants' motor performance in an objective and scalable manner in out-of-hospital settings. This information could be used for developmental research and to support clinical decision-making, such as detection of developmental problems and guiding of their therapeutic interventions. MAIJU-based analyses rely fully on the classification of infant's posture and movement; it is hence essential to study ways to increase the accuracy of such classifications, aiming to increase the reliability and robustness of the automated analysis. Here, we investigated how self-supervised pre-training improves performance of the classifiers used for analyzing MAIJU recordings, and we studied whether performance of the classifier models is affected by context-selective quality-screening of pre-training data to exclude periods of little infant movement or with missing sensors. Our experiments show that i) pre-training the classifier with unlabeled data leads to a robust accuracy increase of subsequent classification models, and ii) selecting context-relevant pre-training data leads to substantial further improvements in the classifier performance.

Abstract: In this paper, we integrate the concepts of feature importance with implicit bias in the context of pattern classification. This is done by means of a three-step methodology that involves (i) building a classifier and tuning its hyperparameters, (ii) building a Fuzzy Cognitive Map model able to quantify implicit bias, and (iii) using the SHAP feature importance to active the neural concepts when performing simulations. The results using a real case study concerning fairness research support our two-fold hypothesis. On the one hand, it is illustrated the risks of using a feature importance method as an absolute tool to measure implicit bias. On the other hand, it is concluded that the amount of bias towards protected features might differ depending on whether the features are numerically or categorically encoded.

Abstract: Deep ReLU Networks can be decomposed into a collection of linear models, each defined in a region of a partition of the input space. This paper provides three results extending this theory. First, we extend this linear decompositions to Graph Neural networks and tensor convolutional networks, as well as networks with multiplicative interactions. Second, we provide proofs that neural networks can be understood as interpretable models such as Multivariate Decision trees and logical theories. Finally, we show how this model leads to computing cheap and exact SHAP values. We validate the theory through experiments with on Graph Neural Networks.

Abstract: Multi-task neural networks learn tasks simultaneously to improve individual task performance. There are three mechanisms of multi-task learning (MTL) which are explored here for the context of structural health monitoring (SHM): (i) the natural occurrence of multiple tasks; (ii) using outputs as inputs (both linked to the recent research in population-based SHM (PBSHM)); and, (iii) additional loss functions to provide different insights. Each of these problem settings for MTL is detailed and an example is given.

Abstract: The scores of distance-based outlier detection methods are difficult to interpret, making it challenging to determine a cut-off threshold between normal and outlier data points without additional context. We describe a generic transformation of distance-based outlier scores into interpretable, probabilistic estimates. The transformation is ranking-stable and increases the contrast between normal and outlier data points. Determining distance relationships between data points is necessary to identify the nearest-neighbor relationships in the data, yet, most of the computed distances are typically discarded. We show that the distances to other data points can be used to model distance probability distributions and, subsequently, use the distributions to turn distance-based outlier scores into outlier probabilities. Our experiments show that the probabilistic transformation does not impact detection performance over numerous tabular and image benchmark datasets but results in interpretable outlier scores with increased contrast between normal and outlier samples. Our work generalizes to a wide range of distance-based outlier detection methods, and because existing distance computations are used, it adds no significant computational overhead.

Abstract: Few multi-agent reinforcement learning (MARL) research on Google Research Football (GRF) focus on the 11v11 multi-agent full-game scenario and to the best of our knowledge, no open benchmark on this scenario has been released to the public. In this work, we fill the gap by providing a population-based MARL training pipeline and hyperparameter settings on multi-agent football scenario that outperforms the bot with difficulty 1.0 from scratch within 2 million steps. Our experiments serve as a reference for the expected performance of Independent Proximal Policy Optimization (IPPO), a state-of-the-art multi-agent reinforcement learning algorithm where each agent tries to maximize its own policy independently across various training configurations. Meanwhile, we open-source our training framework Light-MALib which extends the MALib codebase by distributed and asynchronized implementation with additional analytical tools for football games. Finally, we provide guidance for building strong football AI with population-based training and release diverse pretrained policies for benchmarking. The goal is to provide the community with a head start for whoever experiment their works on GRF and a simple-to-use population-based training framework for further improving their agents through self-play. The implementation is available at https://github.com/Shanghai-Digital-Brain-Laboratory/DB-Football.

Abstract: To reduce the complexity of the hardware implementation of neural network-based optical channel equalizers, we demonstrate that the performance of the biLSTM equalizer with approximated activation functions is close to that of the original model.

Abstract: Label distribution learning (LDL) is a new machine learning paradigm for solving label ambiguity. Since it is difficult to directly obtain label distributions, many studies are focusing on how to recover label distributions from logical labels, dubbed label enhancement (LE). Existing LE methods estimate label distributions by simply building a mapping relationship between features and label distributions under the supervision of logical labels. They typically overlook the fact that both features and logical labels are descriptions of the instance from different views. Therefore, we propose a novel method called Contrastive Label Enhancement (ConLE) which integrates features and logical labels into the unified projection space to generate high-level features by contrastive learning strategy. In this approach, features and logical labels belonging to the same sample are pulled closer, while those of different samples are projected farther away from each other in the projection space. Subsequently, we leverage the obtained high-level features to gain label distributions through a welldesigned training strategy that considers the consistency of label attributes. Extensive experiments on LDL benchmark datasets demonstrate the effectiveness and superiority of our method.

Abstract: Protecting user privacy is a major concern for many machine learning systems that are deployed at scale and collect from a diverse set of population. One way to address this concern is by collecting and releasing data labels in an aggregated manner so that the information about a single user is potentially combined with others. In this paper, we explore the possibility of training machine learning models with aggregated data labels, rather than individual labels. Specifically, we consider two natural aggregation procedures suggested by practitioners: curated bags where the data points are grouped based on common features and random bags where the data points are grouped randomly in bag of similar sizes. For the curated bag setting and for a broad range of loss functions, we show that we can perform gradient-based learning without any degradation in performance that may result from aggregating data. Our method is based on the observation that the sum of the gradients of the loss function on individual data examples in a curated bag can be computed from the aggregate label without the need for individual labels. For the random bag setting, we provide a generalization risk bound based on the Rademacher complexity of the hypothesis class and show how empirical risk minimization can be regularized to achieve the smallest risk bound. In fact, in the random bag setting, there is a trade-off between size of the bag and the achievable error rate as our bound indicates. Finally, we conduct a careful empirical study to confirm our theoretical findings. In particular, our results suggest that aggregate learning can be an effective method for preserving user privacy while maintaining model accuracy.

Abstract: A private learner is trained on a sample of labeled points and generates a hypothesis that can be used for predicting the labels of newly sampled points while protecting the privacy of the training set [Kasiviswannathan et al., FOCS 2008]. Research uncovered that private learners may need to exhibit significantly higher sample complexity than non-private learners as is the case with, e.g., learning of one-dimensional threshold functions [Bun et al., FOCS 2015, Alon et al., STOC 2019]. We explore prediction as an alternative to learning. Instead of putting forward a hypothesis, a predictor answers a stream of classification queries. Earlier work has considered a private prediction model with just a single classification query [Dwork and Feldman, COLT 2018]. We observe that when answering a stream of queries, a predictor must modify the hypothesis it uses over time, and, furthermore, that it must use the queries for this modification, hence introducing potential privacy risks with respect to the queries themselves. We introduce private everlasting prediction taking into account the privacy of both the training set and the (adaptively chosen) queries made to the predictor. We then present a generic construction of private everlasting predictors in the PAC model. The sample complexity of the initial training sample in our construction is quadratic (up to polylog factors) in the VC dimension of the concept class. Our construction allows prediction for all concept classes with finite VC dimension, and in particular threshold functions with constant size initial training sample, even when considered over infinite domains, whereas it is known that the sample complexity of privately learning threshold functions must grow as a function of the domain size and hence is impossible for infinite domains.

Abstract: A common pipeline in learning-based control is to iteratively estimate a model of system dynamics, and apply a trajectory optimization algorithm - e.g.~$\mathtt{iLQR}$ - on the learned model to minimize a target cost. This paper conducts a rigorous analysis of a simplified variant of this strategy for general nonlinear systems. We analyze an algorithm which iterates between estimating local linear models of nonlinear system dynamics and performing $\mathtt{iLQR}$-like policy updates. We demonstrate that this algorithm attains sample complexity polynomial in relevant problem parameters, and, by synthesizing locally stabilizing gains, overcomes exponential dependence in problem horizon. Experimental results validate the performance of our algorithm, and compare to natural deep-learning baselines.

Abstract: In this paper, we present a machine learning-based data generator framework tailored to aid researchers who utilize simulations to examine various physical systems or processes. High computational costs and the resulting limited data often pose significant challenges to gaining insights into these systems or processes. Our approach involves a two-step process: initially, we train a supervised predictive model using a limited simulated dataset to predict simulation outcomes. Subsequently, a reinforcement learning agent is trained to generate accurate, simulation-like data by leveraging the supervised model. With this framework, researchers can generate more accurate data and know the outcomes without running high computational simulations, which enables them to explore the parameter space more efficiently and gain deeper insights into physical systems or processes. We demonstrate the effectiveness of the proposed framework by applying it to two case studies, one focusing on earthquake rupture physics and the other on new material development.

Abstract: In Federated Learning (FL) client devices connected over the internet collaboratively train a machine learning model without sharing their private data with a central server or with other clients. The seminal Federated Averaging (FedAvg) algorithm trains a single global model by performing rounds of local training on clients followed by model averaging. FedAvg can improve the communication-efficiency of training by performing more steps of Stochastic Gradient Descent (SGD) on clients in each round. However, client data in real-world FL is highly heterogeneous, which has been extensively shown to slow model convergence and harm final performance when $K > 1$ steps of SGD are performed on clients per round. In this work we propose decaying $K$ as training progresses, which can jointly improve the final performance of the FL model whilst reducing the wall-clock time and the total computational cost of training compared to using a fixed $K$. We analyse the convergence of FedAvg with decaying $K$ for strongly-convex objectives, providing novel insights into the convergence properties, and derive three theoretically-motivated decay schedules for $K$. We then perform thorough experiments on four benchmark FL datasets (FEMNIST, CIFAR100, Sentiment140, Shakespeare) to show the real-world benefit of our approaches in terms of real-world convergence time, computational cost, and generalisation performance.

Abstract: The article presents the torchosr package - a Python package compatible with PyTorch library - offering tools and methods dedicated to Open Set Recognition in Deep Neural Networks. The package offers two state-of-the-art methods in the field, a set of functions for handling base sets and generation of derived sets for the Open Set Recognition task (where some classes are considered unknown and used only in the testing process) and additional tools to handle datasets and methods. The main goal of the package proposal is to simplify and promote the correct experimental evaluation, where experiments are carried out on a large number of derivative sets with various Openness and class-to-category assignments. The authors hope that state-of-the-art methods available in the package will become a source of a correct and open-source implementation of the relevant solutions in the domain.

Abstract: Prompt-tuning has emerged as a promising method for adapting pre-trained models to downstream tasks or aligning with human preferences. Prompt learning is widely used in NLP but has limited applicability to RL due to the complex physical meaning and environment-specific information contained within RL prompts. These factors require supervised learning to imitate the demonstrations and may result in a loss of meaning after learning. Additionally, directly extending prompt-tuning approaches to RL is challenging because RL prompts guide agent behavior based on environmental modeling and analysis, rather than filling in missing information, making it unlikely that adjustments to the prompt format for downstream tasks, as in NLP, can yield significant improvements. In this work, we propose the Prompt-Tuning DT algorithm to address these challenges by using trajectory segments as prompts to guide RL agents in acquiring environmental information and optimizing prompts via black-box tuning to enhance their ability to contain more relevant information, thereby enabling agents to make better decisions. Our approach involves randomly sampling a Gaussian distribution to fine-tune the elements of the prompt trajectory and using preference ranking function to find the optimization direction, thereby providing more informative prompts and guiding the agent towards specific preferences in the target environment. Extensive experiments show that with only 0.03% of the parameters learned, Prompt-Tuning DT achieves comparable or even better performance than full-model fine-tuning in low-data scenarios. Our work contributes to the advancement of prompt-tuning approaches in RL, providing a promising direction for optimizing large RL agents for specific preference tasks.

Abstract: This research paper presents an experimental approach to using the Reptile algorithm for reinforcement learning to train a neural network to play Super Mario Bros. We implement the Reptile algorithm using the Super Mario Bros Gym library and TensorFlow in Python, creating a neural network model with a single convolutional layer, a flatten layer, and a dense layer. We define the optimizer and use the Reptile class to create an instance of the Reptile meta-learning algorithm. We train the model using multiple tasks and episodes, choosing actions using the current weights of the neural network model, taking those actions in the environment, and updating the model weights using the Reptile algorithm. We evaluate the performance of the algorithm by printing the total reward for each episode. In addition, we compare the performance of the Reptile algorithm approach to two other popular reinforcement learning algorithms, Proximal Policy Optimization (PPO) and Deep Q-Network (DQN), applied to the same Super Mario Bros task. Our results demonstrate that the Reptile algorithm provides a promising approach to few-shot learning in video game AI, with comparable or even better performance than the other two algorithms, particularly in terms of moves vs distance that agent performs for 1M episodes of training. The results shows that best total distance for world 1-2 in the game environment were ~1732 (PPO), ~1840 (DQN) and ~2300 (RAMario). Full code is available at https://github.com/s4nyam/RAMario.

Abstract: We study distributionally robust offline reinforcement learning (robust offline RL), which seeks to find an optimal robust policy purely from an offline dataset that can perform well in perturbed environments. We propose a generic algorithm framework \underline{D}oubly \underline{P}essimistic \underline{M}odel-based \underline{P}olicy \underline{O}ptimization ($\texttt{P}^2\texttt{MPO}$) for robust offline RL, which features a novel combination of a flexible model estimation subroutine and a doubly pessimistic policy optimization step. The \emph{double pessimism} principle is crucial to overcome the distributional shift incurred by i) the mismatch between behavior policy and the family of target policies; and ii) the perturbation of the nominal model. Under certain accuracy assumptions on the model estimation subroutine, we show that $\texttt{P}^2\texttt{MPO}$ is provably efficient with \emph{robust partial coverage data}, which means that the offline dataset has good coverage of the distributions induced by the optimal robust policy and perturbed models around the nominal model. By tailoring specific model estimation subroutines for concrete examples including tabular Robust Markov Decision Process (RMDP), factored RMDP, and RMDP with kernel and neural function approximations, we show that $\texttt{P}^2\texttt{MPO}$ enjoys a $\tilde{\mathcal{O}}(n^{-1/2})$ convergence rate, where $n$ is the number of trajectories in the offline dataset. Notably, these models, except for the tabular case, are first identified and proven tractable by this paper. To the best of our knowledge, we first propose a general learning principle -- double pessimism -- for robust offline RL and show that it is provably efficient in the context of general function approximations.