PyTorchKR Papers

최신 논문

총 10,241편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2408.08295공식 코드 · Papers with Code

SLCA++: Unleash the Power of Sequential Fine-tuning for Continual Learning with Pre-training

In recent years, continual learning with pre-training (CLPT) has received widespread interest, instead of its traditional focus of training from scratch. The use of strong pre-trained models (PTMs) can greatly facilitate knowledge transfer and alleviate catastrophic forgetting, but also suffers from progressive overfitting of pre-trained knowledge into specific downstream tasks. A majority of current efforts often keep the PTMs frozen and incorporate task-specific prompts to instruct representation learning, coupled with a prompt selection process for inference. However, due to the limited capacity of prompt parameters, this strategy demonstrates only sub-optimal performance in continual learning. In comparison, tuning all parameters of PTMs often provides the greatest potential for representation learning, making sequential fine-tuning (Seq FT) a fundamental baseline that has been overlooked in CLPT. To this end, we present an in-depth analysis of the progressive overfitting problem from the lens of Seq FT. Considering that the overly fast representation learning and the biased classification layer constitute this particular problem, we introduce the advanced Slow Learner with Classifier Alignment (SLCA++) framework to unleash the power of Seq FT, serving as a strong baseline approach for CLPT. Our approach involves a Slow Learner to selectively reduce the learning rate of backbone parameters, and a Classifier Alignment to align the disjoint classification layers in a post-hoc fashion. We further enhance the efficacy of SL with a symmetric cross-entropy loss, as well as employ a parameter-efficient strategy to implement Seq FT with SLCA++. Across a variety of continual learning scenarios on image classification benchmarks, our approach provides substantial improvements and outperforms state-of-the-art methods by a large margin. Code: https://github.com/GengDavid/SLCA.

arXiv:2408.07362공식 코드 · Papers with Code

BadMerging: Backdoor Attacks Against Model Merging

Fine-tuning pre-trained models for downstream tasks has led to a proliferation of open-sourced task-specific models. Recently, Model Merging (MM) has emerged as an effective approach to facilitate knowledge transfer among these independently fine-tuned models. MM directly combines multiple fine-tuned task-specific models into a merged model without additional training, and the resulting model shows enhanced capabilities in multiple tasks. Although MM provides great utility, it may come with security risks because an adversary can exploit MM to affect multiple downstream tasks. However, the security risks of MM have barely been studied. In this paper, we first find that MM, as a new learning paradigm, introduces unique challenges for existing backdoor attacks due to the merging process. To address these challenges, we introduce BadMerging, the first backdoor attack specifically designed for MM. Notably, BadMerging allows an adversary to compromise the entire merged model by contributing as few as one backdoored task-specific model. BadMerging comprises a two-stage attack mechanism and a novel feature-interpolation-based loss to enhance the robustness of embedded backdoors against the changes of different merging parameters. Considering that a merged model may incorporate tasks from different domains, BadMerging can jointly compromise the tasks provided by the adversary (on-task attack) and other contributors (off-task attack) and solve the corresponding unique challenges with novel attack designs. Extensive experiments show that BadMerging achieves remarkable attacks against various MM algorithms. Our ablation study demonstrates that the proposed attack designs can progressively contribute to the attack performance. Finally, we show that prior defense mechanisms fail to defend against our attacks, highlighting the need for more advanced defense.

arXiv:2408.07176공식 코드 · Papers with Code

Surrogate-Assisted Search with Competitive Knowledge Transfer for Expensive Optimization

Expensive optimization problems (EOPs) have attracted increasing research attention over the decades due to their ubiquity in a variety of practical applications. Despite many sophisticated surrogate-assisted evolutionary algorithms (SAEAs) that have been developed for solving such problems, most of them lack the ability to transfer knowledge from previously-solved tasks and always start their search from scratch, making them troubled by the notorious cold-start issue. A few preliminary studies that integrate transfer learning into SAEAs still face some issues, such as defective similarity quantification that is prone to underestimate promising knowledge, surrogate-dependency that makes the transfer methods not coherent with the state-of-the-art in SAEAs, etc. In light of the above, a plug and play competitive knowledge transfer method is proposed to boost various SAEAs in this paper. Specifically, both the optimized solutions from the source tasks and the promising solutions acquired by the target surrogate are treated as task-solving knowledge, enabling them to compete with each other to elect the winner for expensive evaluation, thus boosting the search speed on the target task. Moreover, the lower bound of the convergence gain brought by the knowledge competition is mathematically analyzed, which is expected to strengthen the theoretical foundation of sequential transfer optimization. Experimental studies conducted on a series of benchmark problems and a practical application from the petroleum industry verify the efficacy of the proposed method. The source code of the competitive knowledge transfer is available at https://github.com/XmingHsueh/SAS-CKT.

arXiv:2408.07556공식 코드 · Papers with Code

PolyCL: Contrastive Learning for Polymer Representation Learning via Explicit and Implicit Augmentations

Polymers play a crucial role in a wide array of applications due to their diverse and tunable properties. Establishing the relationship between polymer representations and their properties is crucial to the computational design and screening of potential polymers via machine learning. The quality of the representation significantly influences the effectiveness of these computational methods. Here, we present a self-supervised contrastive learning paradigm, PolyCL, for learning high-quality polymer representation without the need for labels. Our model combines explicit and implicit augmentation strategies for improved learning performance. The results demonstrate that our model achieves either better, or highly competitive, performances on transfer learning tasks as a feature extractor without an overcomplicated training strategy or hyperparameter optimisation. Further enhancing the efficacy of our model, we conducted extensive analyses on various augmentation combinations used in contrastive learning. This led to identifying the most effective combination to maximise PolyCL's performance.

arXiv:2408.06870공식 코드 · Papers with Code

Spectrum Prediction With Deep 3D Pyramid Vision Transformer Learning

In this paper, we propose a deep learning (DL)-based task-driven spectrum prediction framework, named DeepSPred. The DeepSPred comprises a feature encoder and a task predictor, where the encoder extracts spectrum usage pattern features, and the predictor configures different networks according to the task requirements to predict future spectrum. Based on the Deep- SPred, we first propose a novel 3D spectrum prediction method combining a flow processing strategy with 3D vision Transformer (ViT, i.e., Swin) and a pyramid to serve possible applications such as spectrum monitoring task, named 3D-SwinSTB. 3D-SwinSTB unique 3D Patch Merging ViT-to-3D ViT Patch Expanding and pyramid designs help the model accurately learn the potential correlation of the evolution of the spectrogram over time. Then, we propose a novel spectrum occupancy rate (SOR) method by redesigning a predictor consisting exclusively of 3D convolutional and linear layers to serve possible applications such as dynamic spectrum access (DSA) task, named 3D-SwinLinear. Unlike the 3D-SwinSTB output spectrogram, 3D-SwinLinear projects the spectrogram directly as the SOR. Finally, we employ transfer learning (TL) to ensure the applicability of our two methods to diverse spectrum services. The results show that our 3D-SwinSTB outperforms recent benchmarks by more than 5%, while our 3D-SwinLinear achieves a 90% accuracy, with a performance improvement exceeding 10%.

arXiv:2408.06452

Wireless Channel Aware Data Augmentation Methods for Deep Learning-Based Indoor Localization

Indoor localization is a challenging problem that - unlike outdoor localization - lacks a universal and robust solution. Machine Learning (ML), particularly Deep Learning (DL), methods have been investigated as a promising approach. Although such methods bring remarkable localization accuracy, they heavily depend on the training data collected from the environment. The data collection is usually a laborious and time-consuming task, but Data Augmentation (DA) can be used to alleviate this issue. In this paper, different from previously used DA, we propose methods that utilize the domain knowledge about wireless propagation channels and devices. The methods exploit the typical hardware component drift in the transceivers and/or the statistical behavior of the channel, in combination with the measured Power Delay Profile (PDP). We comprehensively evaluate the proposed methods to demonstrate their effectiveness. This investigation mainly focuses on the impact of factors such as the number of measurements, augmentation proportion, and the environment of interest impact the effectiveness of the different DA methods. We show that in the low-data regime (few actual measurements available), localization accuracy increases up to 50%, matching non-augmented results in the high-data regime. In addition, the proposed methods may outperform the measurement-only high-data performance by up to 33% using only 1/4 of the amount of measured data. We also exhibit the effect of different training data distribution and quality on the effectiveness of DA. Finally, we demonstrate the power of the proposed methods when employed along with Transfer Learning (TL) to address the data scarcity in target and/or source environments.

arXiv:2408.07071

Approaches for enhancing extrapolability in process-based and data-driven models in hydrology

The application of process-based and data-driven hydrological models is crucial in modern hydrological research, especially for predicting key water cycle variables such as runoff, evapotranspiration (ET), and soil moisture. These models provide a scientific basis for water resource management, flood forecasting, and ecological protection. Process-based models simulate the physical mechanisms of watershed hydrological processes, while data-driven models leverage large datasets and advanced machine learning algorithms. This paper reviewed and compared methods for assessing and enhancing the extrapolability of both model types, discussing their prospects and limitations. Key strategies include the use of leave-one-out cross-validation and similarity-based methods to evaluate model performance in ungauged regions. Deep learning, transfer learning, and domain adaptation techniques are also promising in their potential to improve model predictions in data-sparse and extreme conditions. Interdisciplinary collaboration and continuous algorithmic advancements are also important to strengthen the global applicability and reliability of hydrological models.

arXiv:2408.06567

AquilaMoE: Efficient Training for MoE Models with Scale-Up and Scale-Out Strategies

In recent years, with the rapid application of large language models across various fields, the scale of these models has gradually increased, and the resources required for their pre-training have grown exponentially. Training an LLM from scratch will cost a lot of computation resources while scaling up from a smaller model is a more efficient approach and has thus attracted significant attention. In this paper, we present AquilaMoE, a cutting-edge bilingual 8*16B Mixture of Experts (MoE) language model that has 8 experts with 16 billion parameters each and is developed using an innovative training methodology called EfficientScale. This approach optimizes performance while minimizing data requirements through a two-stage process. The first stage, termed Scale-Up, initializes the larger model with weights from a pre-trained smaller model, enabling substantial knowledge transfer and continuous pretraining with significantly less data. The second stage, Scale-Out, uses a pre-trained dense model to initialize the MoE experts, further enhancing knowledge transfer and performance. Extensive validation experiments on 1.8B and 7B models compared various initialization schemes, achieving models that maintain and reduce loss during continuous pretraining. Utilizing the optimal scheme, we successfully trained a 16B model and subsequently the 8*16B AquilaMoE model, demonstrating significant improvements in performance and training efficiency.

arXiv:2408.06459

InfLocNet: Enhanced Lung Infection Localization and Disease Detection from Chest X-Ray Images Using Lightweight Deep Learning

In recent years, the integration of deep learning techniques into medical imaging has revolutionized the diagnosis and treatment of lung diseases, particularly in the context of COVID-19 and pneumonia. This paper presents a novel, lightweight deep learning based segmentation-classification network designed to enhance the detection and localization of lung infections using chest X-ray images. By leveraging the power of transfer learning with pre-trained VGG-16 weights, our model achieves robust performance even with limited training data. The architecture incorporates refined skip connections within the UNet++ framework, reducing semantic gaps and improving precision in segmentation tasks. Additionally, a classification module is integrated at the end of the encoder block, enabling simultaneous classification and segmentation. This dual functionality enhances the model's versatility, providing comprehensive diagnostic insights while optimizing computational efficiency. Experimental results demonstrate that our proposed lightweight network outperforms existing methods in terms of accuracy and computational requirements, making it a viable solution for real-time and resource constrained medical imaging applications. Furthermore, the streamlined design facilitates easier hyperparameter tuning and deployment on edge devices. This work underscores the potential of advanced deep learning architectures in improving clinical outcomes through precise and efficient medical image analysis. Our model achieved remarkable results with an Intersection over Union (IoU) of 93.59% and a Dice Similarity Coefficient (DSC) of 97.61% in lung area segmentation, and an IoU of 97.67% and a DSC of 87.61% for infection region localization. Additionally, it demonstrated high accuracy of 93.86% and sensitivity of 89.55% in detecting chest diseases, highlighting its efficacy and reliability.

arXiv:2408.06150

LipidBERT: A Lipid Language Model Pre-trained on METiS de novo Lipid Library

In this study, we generate and maintain a database of 10 million virtual lipids through METiS's in-house de novo lipid generation algorithms and lipid virtual screening techniques. These virtual lipids serve as a corpus for pre-training, lipid representation learning, and downstream task knowledge transfer, culminating in state-of-the-art LNP property prediction performance. We propose LipidBERT, a BERT-like model pre-trained with the Masked Language Model (MLM) and various secondary tasks. Additionally, we compare the performance of embeddings generated by LipidBERT and PhatGPT, our GPT-like lipid generation model, on downstream tasks. The proposed bilingual LipidBERT model operates in two languages: the language of ionizable lipid pre-training, using in-house dry-lab lipid structures, and the language of LNP fine-tuning, utilizing in-house LNP wet-lab data. This dual capability positions LipidBERT as a key AI-based filter for future screening tasks, including new versions of METiS de novo lipid libraries and, more importantly, candidates for in vivo testing for orgran-targeting LNPs. To the best of our knowledge, this is the first successful demonstration of the capability of a pre-trained language model on virtual lipids and its effectiveness in downstream tasks using web-lab data. This work showcases the clever utilization of METiS's in-house de novo lipid library as well as the power of dry-wet lab integration.

arXiv:2408.05307

Audio-visual cross-modality knowledge transfer for machine learning-based in-situ monitoring in laser additive manufacturing

Various machine learning (ML)-based in-situ monitoring systems have been developed to detect anomalies and defects in laser additive manufacturing (LAM) processes. While multimodal fusion, which integrates data from visual, audio, and other modalities, can improve monitoring performance, it also increases hardware, computational, and operational costs. This paper introduces a cross-modality knowledge transfer (CMKT) methodology for LAM in-situ monitoring, which transfers knowledge from a source modality to a target modality. CMKT enhances the representativeness of the features extracted from the target modality, allowing the removal of source modality sensors during prediction. This paper proposes three CMKT methods: semantic alignment, fully supervised mapping, and semi-supervised mapping. The semantic alignment method establishes a shared encoded space between modalities to facilitate knowledge transfer. It employs a semantic alignment loss to align the distributions of identical groups (e.g., visual and audio defective groups) and a separation loss to distinguish different groups (e.g., visual defective and audio defect-free groups). The two mapping methods transfer knowledge by deriving features from one modality to another using fully supervised and semi-supervised learning approaches. In a case study for LAM in-situ defect detection, the proposed CMKT methods were compared with multimodal audio-visual fusion. The semantic alignment method achieved an accuracy of 98.6% while removing the audio modality during the prediction phase, which is comparable to the 98.2% accuracy obtained through multimodal fusion. Using explainable artificial intelligence, we discovered that semantic alignment CMKT can extract more representative features while reducing noise by leveraging the inherent correlations between modalities.

arXiv:2408.05952

Optimizing Vision Transformers with Data-Free Knowledge Transfer

The groundbreaking performance of transformers in Natural Language Processing (NLP) tasks has led to their replacement of traditional Convolutional Neural Networks (CNNs), owing to the efficiency and accuracy achieved through the self-attention mechanism. This success has inspired researchers to explore the use of transformers in computer vision tasks to attain enhanced long-term semantic awareness. Vision transformers (ViTs) have excelled in various computer vision tasks due to their superior ability to capture long-distance dependencies using the self-attention mechanism. Contemporary ViTs like Data Efficient Transformers (DeiT) can effectively learn both global semantic information and local texture information from images, achieving performance comparable to traditional CNNs. However, their impressive performance comes with a high computational cost due to very large number of parameters, hindering their deployment on devices with limited resources like smartphones, cameras, drones etc. Additionally, ViTs require a large amount of data for training to achieve performance comparable to benchmark CNN models. Therefore, we identified two key challenges in deploying ViTs on smaller form factor devices: the high computational requirements of large models and the need for extensive training data. As a solution to these challenges, we propose compressing large ViT models using Knowledge Distillation (KD), which is implemented data-free to circumvent limitations related to data availability. Additionally, we conducted experiments on object detection within the same environment in addition to classification tasks. Based on our analysis, we found that datafree knowledge distillation is an effective method to overcome both issues, enabling the deployment of ViTs on less resourceconstrained devices.

arXiv:2408.05992

Transfer learning of state-based potential games for process optimization in decentralized manufacturing systems

This paper presents a novel transfer learning approach in state-based potential games (TL-SbPGs) for enhancing distributed self-optimization in manufacturing systems. The approach focuses on the practical relevant industrial setting where sharing and transferring gained knowledge among similar-behaved players improves the self-learning mechanism in large-scale systems. With TL-SbPGs, the gained knowledge can be reused by other players to optimize their policies, thereby improving the learning outcomes of the players and accelerating the learning process. To accomplish this goal, we develop transfer learning concepts and similarity criteria for players, which offer two distinct settings: (a) predefined similarities between players and (b) dynamically inferred similarities between players during training. We formally prove the applicability of the SbPG framework in transfer learning. Additionally, we introduce an efficient method to determine the optimal timing and weighting of the transfer learning procedure during the training phase. Through experiments on a laboratory-scale testbed, we demonstrate that TL-SbPGs significantly boost production efficiency while reducing power consumption of the production schedules while also outperforming native SbPGs.

arXiv:2408.04816공식 코드 · Papers with Code

FUSE-ing Language Models: Zero-Shot Adapter Discovery for Prompt Optimization Across Tokenizers

The widespread use of large language models has resulted in a multitude of tokenizers and embedding spaces, making knowledge transfer in prompt discovery tasks difficult. In this work, we propose FUSE (Flexible Unification of Semantic Embeddings), an inexpensive approach to approximating an adapter layer that maps from one model's textual embedding space to another, even across different tokenizers. We introduce a third-order tensor-based representation of a model's embedding space that aligns semantic embeddings that have been split apart by different tokenizers, and use this representation to derive an approximation of the gradient of one model's outputs with respect to another model's embedding space. We show the efficacy of our approach via multi-objective optimization over vision-language and causal language models for image captioning and sentiment-based image captioning.

arXiv:2408.05178공식 코드 · Papers with Code

ECG-FM: An Open Electrocardiogram Foundation Model

The electrocardiogram (ECG) is a ubiquitous diagnostic test. Conventional task-specific ECG analysis models require large numbers of expensive ECG annotations or associated labels to train. Transfer learning techniques have been shown to improve generalization and reduce reliance on labeled data. We present ECG-FM, an open foundation model for ECG analysis, and conduct a comprehensive study performed on a dataset of 1.66 million ECGs sourced from both publicly available and private institutional sources. ECG-FM adopts a transformer-based architecture and is pretrained on 2.5 million samples using ECG-specific augmentations and contrastive learning, as well as a continuous signal masking objective. Our transparent evaluation includes a diverse range of downstream tasks, where we predict ECG interpretation labels, reduced left ventricular ejection fraction, and abnormal cardiac troponin. Affirming ECG-FM's effectiveness as a foundation model, we demonstrate how its command of contextual information results in strong performance, rich pretrained embeddings, and reliable interpretability. Due to a lack of open-weight practices, we highlight how ECG analysis is lagging behind other medical machine learning subfields in terms of foundation model adoption. Our code is available at https://github.com/bowang-lab/ECG-FM/.

arXiv:2408.04914공식 코드 · Papers with Code

GuidedNet: Semi-Supervised Multi-Organ Segmentation via Labeled Data Guide Unlabeled Data

Semi-supervised multi-organ medical image segmentation aids physicians in improving disease diagnosis and treatment planning and reduces the time and effort required for organ annotation.Existing state-of-the-art methods train the labeled data with ground truths and train the unlabeled data with pseudo-labels. However, the two training flows are separate, which does not reflect the interrelationship between labeled and unlabeled data.To address this issue, we propose a semi-supervised multi-organ segmentation method called GuidedNet, which leverages the knowledge from labeled data to guide the training of unlabeled data. The primary goals of this study are to improve the quality of pseudo-labels for unlabeled data and to enhance the network's learning capability for both small and complex organs.A key concept is that voxel features from labeled and unlabeled data that are close to each other in the feature space are more likely to belong to the same class.On this basis, a 3D Consistent Gaussian Mixture Model (3D-CGMM) is designed to leverage the feature distributions from labeled data to rectify the generated pseudo-labels.Furthermore, we introduce a Knowledge Transfer Cross Pseudo Supervision (KT-CPS) strategy, which leverages the prior knowledge obtained from the labeled data to guide the training of the unlabeled data, thereby improving the segmentation accuracy for both small and complex organs. Extensive experiments on two public datasets, FLARE22 and AMOS, demonstrated that GuidedNet is capable of achieving state-of-the-art performance. The source code with our proposed model are available at https://github.com/kimjisoo12/GuidedNet.

arXiv:2408.04763

Segmentation of Mental Foramen in Orthopantomographs: A Deep Learning Approach

Precise identification and detection of the Mental Foramen are crucial in dentistry, impacting procedures such as impacted tooth removal, cyst surgeries, and implants. Accurately identifying this anatomical feature facilitates post-surgery issues and improves patient outcomes. Moreover, this study aims to accelerate dental procedures, elevating patient care and healthcare efficiency in dentistry. This research used Deep Learning methods to accurately detect and segment the Mental Foramen from panoramic radiograph images. Two mask types, circular and square, were used during model training. Multiple segmentation models were employed to identify and segment the Mental Foramen, and their effectiveness was evaluated using diverse metrics. An in-house dataset comprising 1000 panoramic radiographs was created for this study. Our experiments demonstrated that the Classical UNet model performed exceptionally well on the test data, achieving a Dice Coefficient of 0.79 and an Intersection over Union (IoU) of 0.67. Moreover, ResUNet++ and UNet Attention models showed competitive performance, with Dice scores of 0.675 and 0.676, and IoU values of 0.683 and 0.671, respectively. We also investigated transfer learning models with varied backbone architectures, finding LinkNet to produce the best outcomes. In conclusion, our research highlights the efficacy of the classical Unet model in accurately identifying and outlining the Mental Foramen in panoramic radiographs. While vital, this task is comparatively simpler than segmenting complex medical datasets such as brain tumours or skin cancer, given their diverse sizes and shapes. This research also holds value in optimizing dental practice, benefiting practitioners and patients.

arXiv:2408.05200공식 코드 · Papers with Code

KIF: Knowledge Identification and Fusion for Language Model Continual Learning

Language model continual learning (CL) has recently attracted significant interest for its ability to adapt large language models (LLMs) to dynamic real-world scenarios without retraining. A major challenge in this domain is catastrophic forgetting, where models lose previously acquired knowledge upon learning new tasks. Existing approaches commonly utilize multiple parameter-efficient fine-tuning (PEFT) blocks to acquire task-specific knowledge, yet these methods are inefficient and fail to leverage potential knowledge transfer across tasks. In this paper, we introduce a novel CL framework for language models, named Knowledge Identification and Fusion (KIF), which boosts knowledge transfer without depending on memory replay. KIF initially segregates the model into 'skill units' based on parameter dependencies, allowing for more precise control. Subsequently, it employs a novel group-wise knowledge identification technique to ascertain the importance distribution of skill units for a new task. By comparing this importance distribution with those from previous tasks, we implement a fine-grained knowledge fusion strategy that retains task-specific knowledge, thereby preventing forgetting, and updates task-shared knowledge, which facilitates bi-directional knowledge transfer. As a result, KIF achieves an optimal balance between retaining prior knowledge and excelling in new tasks. KIF also demonstrates strong generalizability, making it suitable for various base models and adaptable to PEFT methods like LoRA. Furthermore, it offers notable extensibility, supporting enhancements through integration with memory replay techniques. Comprehensive experiments conducted on two CL benchmarks, involving models ranging from 220M to 7B parameters, affirm the effectiveness of KIF and its variants across different settings.

arXiv:2408.04042

Scaling Law of Sim2Real Transfer Learning in Expanding Computational Materials Databases for Real-World Predictions

To address the challenge of limited experimental materials data, extensive physical property databases are being developed based on high-throughput computational experiments, such as molecular dynamics simulations. Previous studies have shown that fine-tuning a predictor pretrained on a computational database to a real system can result in models with outstanding generalization capabilities compared to learning from scratch. This study demonstrates the scaling law of simulation-to-real (Sim2Real) transfer learning for several machine learning tasks in materials science. Case studies of three prediction tasks for polymers and inorganic materials reveal that the prediction error on real systems decreases according to a power-law as the size of the computational data increases. Observing the scaling behavior offers various insights for database development, such as determining the sample size necessary to achieve a desired performance, identifying equivalent sample sizes for physical and computational experiments, and guiding the design of data production protocols for downstream real-world tasks.

arXiv:2408.04498공식 코드 · Papers with Code

Model-Based Transfer Learning for Contextual Reinforcement Learning

Deep reinforcement learning (RL) is a powerful approach to complex decision making. However, one issue that limits its practical application is its brittleness, sometimes failing to train in the presence of small changes in the environment. Motivated by the success of zero-shot transfer-where pre-trained models perform well on related tasks-we consider the problem of selecting a good set of training tasks to maximize generalization performance across a range of tasks. Given the high cost of training, it is critical to select training tasks strategically, but not well understood how to do so. We hence introduce Model-Based Transfer Learning (MBTL), which layers on top of existing RL methods to effectively solve contextual RL problems. MBTL models the generalization performance in two parts: 1) the performance set point, modeled using Gaussian processes, and 2) performance loss (generalization gap), modeled as a linear function of contextual similarity. MBTL combines these two pieces of information within a Bayesian optimization (BO) framework to strategically select training tasks. We show theoretically that the method exhibits sublinear regret in the number of training tasks and discuss conditions to further tighten regret bounds. We experimentally validate our methods using urban traffic and standard continuous control benchmarks. The experimental results suggest that MBTL can achieve up to 43x improved sample efficiency compared with canonical independent training and multi-task training. Further experiments demonstrate the efficacy of BO and the insensitivity to the underlying RL algorithm and hyperparameters. This work lays the foundations for investigating explicit modeling of generalization, thereby enabling principled yet effective methods for contextual RL.