Meta-learning Framework Research Articles

Personalized Federated Learning on Non-IID Data via Group-based Meta-learning

Personalized federated learning (PFL) has emerged as a paradigm to provide a personalized model that can fit the local data distribution of each client. One natural choice for PFL is to leverage the fast adaptation capability of meta-learning, where it first obtains a single global model, and each client achieves a personalized model by fine-tuning the global one with its local data. However, existing meta-learning-based approaches implicitly assume that the data distribution among different clients is similar, which may not be applicable due to the property of data heterogeneity in federated learning. In this work, we propose a Group-based Federated Meta-Learning framework, called G-FML , which adaptively divides the clients into groups based on the similarity of their data distribution, and the personalized models are obtained with meta-learning within each group. In particular, we develop a simple yet effective grouping mechanism to adaptively partition the clients into multiple groups. Our mechanism ensures that each group is formed by the clients with similar data distribution such that the group-wise meta-model can achieve “personalization” at large. By doing so, our framework can be generalized to a highly heterogeneous environment. We evaluate the effectiveness of our proposed G-FML framework on three heterogeneous benchmarking datasets. The experimental results show that our framework improves the model accuracy by up to 13.15% relative to the state-of-the-art federated meta-learning.

DDPG with Transfer Learning and Meta Learning Framework for Resource Allocation in Underlay Cognitive Radio Network

DDPG with Transfer Learning and Meta Learning Framework for Resource Allocation in Underlay Cognitive Radio Network

Distilled Meta-learning for Multi-Class Incremental Learning

Meta-learning approaches have recently achieved promising performance in multi-class incremental learning. However, meta-learners still suffer from catastrophic forgetting, i.e., they tend to forget the learned knowledge from the old tasks when they focus on rapidly adapting to the new classes of the current task. To solve this problem, we propose a novel distilled meta-learning (DML) framework for multi-class incremental learning that integrates seamlessly meta-learning with knowledge distillation in each incremental stage. Specifically, during inner-loop training, knowledge distillation is incorporated into the DML to overcome catastrophic forgetting. During outer-loop training, a meta-update rule is designed for the meta-learner to learn across tasks and quickly adapt to new tasks. By virtue of the bilevel optimization, our model is encouraged to reach a balance between the retention of old knowledge and the learning of new knowledge. Experimental results on four benchmark datasets demonstrate the effectiveness of our proposal and show that our method significantly outperforms other state-of-the-art incremental learning methods.

Few-shot learning via graph embeddings with convolutional networks for low-data molecular property prediction

Graph neural networks and convolutional architectures have proven to be pivotal in improving the prediction of molecular properties in drug discovery. However, this is fundamentally a low data problem that is incompatible with regular deep learning approaches. Contemporary deep networks require large amounts of training data, which severely limits the prediction of new molecular entities from limited available data. In this paper, we address the challenge of low data in molecular property prediction by: (1) defining a set of deep learning architectures that accept compound chemical structures in the form of molecular graphs, (2) creating a few-shot learning strategy across graph neural networks and convolutional neural networks to leverage the rich information of graph embeddings, and (3) proposing a two-module meta-learning framework to learn from task-transferable knowledge and predict molecular properties on few-shot data. Furthermore, we conduct multiple experiments on two benchmark multiproperty datasets to demonstrate a superior performance over conventional graph-based baselines. ROC-AUC results for 10-shot experiments show an average improvement of +11.37% on Tox21 and +0.53% on SIDER, which are representative small-sized biological datasets for molecular property prediction.

Pan-Peptide Meta Learning for T-cell receptor–antigen binding recognition

The identification of the mechanisms by which T-cell receptors (TCRs) interact with human antigens provides a crucial opportunity to develop new vaccines, diagnostics and immunotherapies. However, the accurate prediction and recognition of TCR–antigen pairing represents a substantial computational challenge in immunology. Existing tools only learn the binding patterns of antigens from many known TCR binding repertoires and fail to recognize antigens that have never been presented to the immune system or for which only a few TCR binding repertoires are known. However, the binding specificity for neoantigens or exogenous peptides is crucial for immune studies and immunotherapy. Therefore, we developed Pan-Peptide Meta Learning (PanPep), a general and robust framework to recognize TCR–antigen binding, by combining the concepts of meta-learning and the neural Turing machine. The neural Turing machine adds external memory to avoid forgetting previously learned tasks, which is used here to accurately predict TCR binding specificity with any peptide, particularly unseen ones. We applied PanPep to various challenging clinical tasks, including (1) qualitatively measuring the clonal expansion of T cells; (2) efficiently sorting responsive T cells in tumour neoantigen therapy; and (3) accurately identifying immune-responsive TCRs in a large cohort from a COVID-19 study. Our comprehensive tests show that PanPep outperforms existing tools. PanPep also offers interpretability, revealing the nature of peptide and TCR interactions in 3D crystal structures. We believe PanPep can be a useful tool to decipher TCR–antigen interactions and that it has broad clinical applications. Machine learning methods can predict and recognize binding patterns between T-cell receptors and human antigens, but they struggle with antigens for which no or little data exist regarding interactions with the immune system. A new method called PanPep based on meta-learning can learn quickly on new binding prediction tasks and accurately predicts pairing between T-cell receptors and new antigens.

Meta-Learning Initializations for Interactive Medical Image Registration.

We present a meta-learning framework for interactive medical image registration. Our proposed framework comprises three components: a learning-based medical image registration algorithm, a form of user interaction that refines registration at inference, and a meta-learning protocol that learns a rapidly adaptable network initialization. This paper describes a specific algorithm that implements the registration, interaction and meta-learning protocol for our exemplar clinical application: registration of magnetic resonance (MR) imaging to interactively acquired, sparsely-sampled transrectal ultrasound (TRUS) images. Our approach obtains comparable registration error (4.26 mm) to the best-performing non-interactive learning-based 3D-to-3D method (3.97 mm) while requiring only a fraction of the data, and occurring in real-time during acquisition. Applying sparsely sampled data to non-interactive methods yields higher registration errors (6.26 mm), demonstrating the effectiveness of interactive MR-TRUS registration, which may be applied intraoperatively given the real-time nature of the adaptation process.

A meta-learning framework for recommending CNN models for plant disease identification tasks

Plant diseases are a major threat to food security and economic prosperity around the globe. Deep learning models based on Convolution Neural Network (CNN) have shown promising results in dealing with plant disease detection tasks. However, according to the No Free Lunch Theorem, no single model is suitable for all cases. Moreover, the vast diversity of plant diseases makes the model selection process time and resource extensive, using exhaustive search. This work proposes a meta-learning-based framework that recommends top-n suitable models for an unseen plant disease detection dataset using the prior evaluations of benchmark models on plant disease detection tasks. Rank-Biased Overlap (RBO) is used to evaluate the efficacy of the proposed framework by evaluating actual rankings with respect to the predicted rankings. Extensive comparative experiments are carried out with different configurations of meta-extractors and meta-learners. The results obtained demonstrate that the probe network trained for 10 epochs (termed as “intermediate stage”) along with standard deviation as meta-extractor and Support Vector Regressor as the meta-learner outperforms the rest with average RBO scores of 0.76, 0.73 and 0.75 for Top-5, Top-3 and Top-1 recommendations, respectively. Overall, this paper presents a viable substitute for the exhaustive search process carried out for choosing the best deep learning model for plant disease detection scenario, leading to better resource utilization and faster implementation procedure.

Few-shot working condition recognition of a sucker-rod pumping system based on a 4-dimensional time-frequency signature and meta-learning convolutional shrinkage neural network

The accurate and intelligent identification of the working conditions of a sucker-rod pumping system is necessary. As onshore oil extraction gradually enters its mid-to late-stage, the cost required to train a deep learning working condition recognition model for pumping wells by obtaining enough new working condition samples is expensive. For the few-shot problem and large calculation issues of new working conditions of oil wells, a working condition recognition method for pumping unit wells based on a 4-dimensional time-frequency signature (4D-TFS) and meta-learning convolutional shrinkage neural network (ML-CSNN) is proposed. First, the measured pumping unit well workup data are converted into 4D-TFS data, and the initial feature extraction task is performed while compressing the data. Subsequently, a convolutional shrinkage neural network (CSNN) with a specific structure that can ablate low-frequency features is designed to extract working conditions features. Finally, a meta-learning fine-tuning framework for learning the network parameters that are susceptible to task changes is merged into the CSNN to solve the few-shot issue. The results of the experiments demonstrate that the trained ML-CSNN has good recognition accuracy and generalization ability for few-shot working condition recognition. More specifically, in the case of lower computational complexity, only few-shot samples are needed to fine-tune the network parameters, and the model can be quickly adapted to new classes of well conditions.

Parametric unsteady flow modeling by using meta learning

Parametric unsteady flow modeling plays a significant role in the fluid dynamics, since the unsteady flow problems are usually involved with complex physical phenomena. Currently, in the study of data-driven unsteady flow modeling, the convolutional neural network based autoencoder (CNN-AE) model has been widely used. While, its application in engineering field has been limited due to challenges that the CNNs are expensive to be trained and the generalization performance of CNN-AE model needs to be improved. Therefore, based on the CNN-AE model, a meta-learning framework is introduced in the present work for parametric unsteady flow modeling. Specifically, both the model-agnostic meta-learning (MAML) and the reptile methods are used to learn the correlations between flow fields under various physical parameters. First, a meta-initialization CNN-AE model is established, with the training data-set under various physical parameters. Then, for a new physical parameter, the meta-initialization CNN-AE model will be finetuned by using few snapshots in a new task with limited epochs. Finally, the proposed CNN-AE based meta-learning framework is validated in two canonical unsteady flow problems with moving boundaries, including oscillating cylinder and flapping airfoil. The results have shown that the CNN-AE based meta-leaning framework can greatly accelerate the adaptation of CNN-AE model in new physical parameters, which can greatly enhance the generalization performance of CNN-AE model.

A Time Series Forecasting Model Selection Framework using CNN and Data Augmentation for Small Sample Data

The key to the accuracy of time series forecasting is to find the most appropriate forecasting method. Therefore, the forecasting model selection of time series has become a new research hotspot in the data analysis field. However, most of the existing meta learning forecasting model selection methods rely on manual selection of features, which leads to low efficiency and lack of objectivity. Therefore, this paper proposes an improved meta learning framework for deep learning time series forecasting model selection. Inspired by computer vision, we transform one -dimensional time series into two-dimensional images, and use convolution neural network to train and classify time series images (model selection). Moreover, in order to deal with the over fitting problem caused by small sample datasets, the sliding window data augmentation method is used to improve the accuracy of small datasets model selection. The large-scale empirical study on M3 data sets shows that the framework has better model selection accuracy and smaller forecasting error than the recurrent neural network (RNN), back propagation neural network and traditional time series image algorithms. In addition, compared with the traditional time series image method, RNN and BP, the classification rate (model selection accuracy) of this algorithm is improved by 6.5%, 4.4% and 3.2%, respectively.

Few-shot Aspect Category Sentiment Analysis via Meta-learning

Existing aspect-based/category sentiment analysis methods have shown great success in detecting sentiment polarity toward a given aspect in a sentence with supervised learning, where the training and inference stages share the same pre-defined set of aspects. However, in practice, the aspect categories are changing rather than keeping fixed over time. Dealing with unseen aspect categories is under-explored in existing methods. In this article, we formulate a new few-shot aspect category sentiment analysis (FSACSA) task, which aims to effectively predict the sentiment polarity of previously unseen aspect categories. To this end, we propose a novel Aspect-Focused Meta-Learning (AFML) framework that constructs aspect-aware and aspect-contrastive representations from external knowledge to match the target aspect with aspects in the training set. Concretely, we first construct two auxiliary contrastive sentences for a given sentence with the incorporation of external knowledge, enabling the learning of sentence representations with a better generalization. Then, we devise an aspect-focused induction network to leverage the contextual sentiment toward a given aspect to refine the label vectors. Furthermore, we employ the episode-based meta-learning algorithm to train the whole network, so as to learn to generalize to novel aspects. Extensive experiments on multiple real-life datasets show that our proposed AFML framework achieves the state-of-the-art results for the FSACSA task.

MetaSEM: Gene Regulatory Network Inference from Single-Cell RNA Data by Meta-Learning.

Regulators in gene regulatory networks (GRNs) are crucial for identifying cell states. However, GRN inference based on scRNA-seq data has several problems, including high dimensionality and sparsity, and requires more label data. Therefore, we propose a meta-learning GRN inference framework to identify regulatory factors. Specifically, meta-learning solves the parameter optimization problem caused by high-dimensional sparse data features. In addition, a few-shot solution was used to solve the problem of lack of label data. A structural equation model (SEM) was embedded in the model to identify important regulators. We integrated the parameter optimization strategy into the bi-level optimization to extract the feature consistent with GRN reasoning. This unique design makes our model robust to small-scale data. By studying the GRN inference task, we confirmed that the selected regulators were closely related to gene expression specificity. We further analyzed the GRN inferred to find the important regulators in cell type identification. Extensive experimental results showed that our model effectively captured the regulator in single-cell GRN inference. Finally, the visualization results verified the importance of the selected regulators for cell type recognition.

Meta-Learning: A Nine-Layer Model Based on Metacognition and Smart Technologies

The international organizations of education have already pointed out that the way students learn, what they learn, and the skills needed, will be radically transformed in the coming years. Smart technologies are ready to come into play, changing the conditions of learning, providing opportunities for transformative learning experiences, and promising more conscious, self-directed and self-motivated learning. Meta-learning refers to a set of mental meta-processes by which learners consciously create and manage personal models of learning. Meta-learning entails a cluster of meta-skills that are progressively and hierarchically transformed, ensuring the transition to the highest levels of understanding termed meta-comprehension. The current article aims to investigate the concept of meta-learning and describe the meta-levels of learning through the lens of metacognition. In addition, the potential of smart technologies to provide fertile ground for the implementation of meta-learning training strategies is examined. The results of this article provide a new meta-learning theoretical framework supported by smart devices capable of supporting future meta-learners or, more accurately, meta-thinkers, to transcend the usual states of knowing and move to the next meta-levels of human intelligence.

Is an SV caller compatible with sequencing data? An online recommendation tool to automatically recommend the optimal caller based on data features.

A lot of bioinformatics tools were released to detect structural variants from the sequencing data during the past decade. For a data analyst, a natural question is about the selection of a tool fits for the data. Thus, this study presents an automatic tool recommendation method to facilitate data analysis. The optimal variant calling tool was recommended from a set of state-of-the-art bioinformatics tools by given a sequencing data. This recommendation method was implemented under a meta-learning framework, identifying the relationships between data features and the performance of tools. First, the meta-features were extracted to characterize the sequencing data and meta-targets were identified to pinpoint the optimal caller for the sequencing data. Second, a meta-model was constructed to bridge the meta-features and meta-targets. Finally, the recommendation was made according to the evaluation from the meta-model. A series of experiments were conducted to validate this recommendation method on both the simulated and real sequencing data. The results revealed that different SV callers often fit different sequencing data. The recommendation accuracy averaged more than 80% across all experimental configurations, outperforming the random- and fixed-pick strategy. To further facilitate the research community, we incorporated the recommendation method into an online cloud services for genomic data analysis, which is available at https://c.solargenomics.com/ via a simple registration. In addition, the source code and a pre-trained model is available at https://github.com/hello-json/CallerRecommendation for academic usages only.

Weakly Supervised Few-Shot Semantic Segmentation via Pseudo Mask Enhancement and Meta Learning

Few shot semantic segmentation has been proposed to enhance the generalization ability of traditional models with limited data. Previous works mainly focus on the supervised tasks, while limited amount of work is explored for the weakly supervised tasks. Weakly supervised semantic segmentation has become an active research area because weakly supervised labels effectively reduce the annotation cost of visual tasks. To this end, we propose a weakly supervised few-shot semantic segmentation model based on the meta learning framework, which utilizes prior knowledge and adjusts itself according to new tasks. Thereupon then, the proposed network is capable of both high efficiency and generalization ability to new tasks. In the pseudo mask generation stage, we develop a WRCAM method with the channel-spatial attention mechanism to refine the coverage size of targets in pseudo masks. In the few-shot semantic segmentation stage, the optimization based meta learning method is used to realize few-shot semantic segmentation by virtue of the refined pseudo masks. The experimental results show that the proposed method not only significantly outperforms weakly supervised SOTA methods, but also could be comparative to some supervised SOTA methods.

Meta-Learning Amidst Heterogeneity and Ambiguity

Meta-learning aims to learn a model that can handle multiple tasks generated from an unknown but shared distribution. However, typical meta-learning algorithms have assumed the tasks to be similar such that a single meta-learner is sufficient to aggregate the variations in all aspects. In addition, there has been less consideration of uncertainty when limited information is given as context. In this paper, we devise a novel meta-learning framework, called Meta-learning Amidst Heterogeneity and Ambiguity (MAHA), that outperforms previous works in prediction based on its ability to task identification. By extensively conducting several experiments in regression and classification, we demonstrate the validity of our model, which turns out to generalize to both task heterogeneity and ambiguity.

Meta-Self-Training Based on Teacher–Student Network for Industrial Label-Noise Fault Diagnosis

The success of deep-learning-based fault diagnosis methods usually relies on a large amount of high-quality data annotation, which is difficult or/and expensive to obtain in many practical industrial applications. When label noise exists, the performance of these methods will be severely affected. Self-training-based label correction is a possible approach to address the label-noise problem. It uses the pseudolabels produced by deep neural network (DNN) to gradually replace the original noisy labels, thereby reducing the impact of label noise. However, due to lack of guidance, self-training methods are often prone to confirmation bias, which can lead to many incorrect pseudolabels. In this work, we aim to address the confirmation bias issue of self-training in the label-noise scenario, so as to obtain better label correction results. A meta-self-training method is proposed, which adopts a self-training mechanism to train a teacher network and uses the pseudolabels generated by the teacher to train a student network. On this basis, the test loss of the student on a small amount of clean data is used to measure the quality of these pseudolabels, and a meta-learning framework is developed to feed this test loss back to the teacher to guide the generation of better pseudolabels. Experimental results show that the proposed method can effectively improve the label correction results of self-training, so that both teacher and student models can achieve good performance in label-noise fault diagnosis problems.

PARS-Push: Personalized, Asynchronous and Robust Decentralized Optimization

We study the multi-step Model-Agnostic Meta-Learning (MAML) framework where a group of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$n$ </tex-math></inline-formula> agents seeks to find a common point that enables “few-shot” learning (personalization) via local stochastic gradient steps on their local functions. We formulate the personalized optimization problem under the MAML framework and propose PARS-Push, a decentralized asynchronous algorithm robust to message failures, communication delays, and directed message sharing. We characterize the convergence rate of PARS-Push under arbitrary multi-step personalization for smooth strongly convex, and smooth non-convex functions. Moreover, we provide numerical experiments showing its performance under heterogeneous data setups.

SPEC: Summary Preference Decomposition for Low-Resource Abstractive Summarization

Neural abstractive summarization has been widely studied and achieved great success with large-scale corpora. However, the considerable cost of annotating data motivates the need for learning strategies under low-resource settings. In this paper, we investigate the problems of learning summarizers with only few examples and propose corresponding methods for improvements. First, typical transfer learning methods are prone to be affected by data properties and learning objectives in the pretext tasks. Therefore, based on pretrained language models, we further present a meta learning framework to transfer few-shot learning processes from source corpora to the target corpus. Second, previous methods learn from training examples without decomposing the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">content</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">preference</i> . The generated summaries could therefore be constrained by the preference bias in the training set, especially under low-resource settings. As such, we propose decomposing the contents and preferences during learning through the parameter modulation, which enables control over preferences during inference. Third, given a target application, specifying required preferences could be non-trivial because the preferences may be difficult to derive through observations. Therefore, we propose a novel decoding method to automatically estimate suitable preferences and generate corresponding summary candidates from the few training examples. Extensive experiments demonstrate that our methods achieve state-of-the-art performance on six diverse corpora with 30.11%/33.95%/27.51% and 26.74%/31.14%/24.48% average improvements on ROUGE-1/2/L under 10- and 100-example settings.

Randomized Spectrum Transformations for Adapting Object Detector in Unseen Domains.

We propose a Meta Learning on Randomized Transformations (MLRT) to learn domain invariant object detectors. Domain generalization is a problem about learning an invariant model from multiple source domains which can generalize well on unseen target domains. This problem is overlooked in object detection field, which is formally named as domain generalizable object detection (DGOD). Moreover, existing domain generalization methods have the problem of domain bias so that they can easily overfit to some specific domain (e.g., source domain). In order to alleviate the domain bias, in MLRT model, a novel randomized spectrum transformation (RST) module is proposed to increase the diversity of source domains. Specifically, RST randomizes the domain specific information of images in frequency-space, which can transform single or multiple source domains into various new domains. Besides, we observe a prior that the gradient imbalance degree among domains can also reflect the domain bias. Therefore, we further propose to alleviate the domain bias from the perspective of gradient balancing, and a novel gradient weighting (GW) module is proposed to balance the gradients over all domains via a hand-crafted weight. Finally we embed our RST and GW into a general meta learning framework and the proposed MLRT model is formalized for DGOD task. Extensive experiments are conducted on six benchmarks, and our method achieves the SOTA performance.