Few-shot Learning Setting Research Articles

Extraction of Substance Use Information From Clinical Notes: Generative Pretrained Transformer-Based Investigation.

Understanding the multifaceted nature of health outcomes requires a comprehensive examination of the social, economic, and environmental determinants that shape individual well-being. Among these determinants, behavioral factors play a crucial role, particularly the consumption patterns of psychoactive substances, which have important implications on public health. The Global Burden of Disease Study shows a growing impact in disability-adjusted life years due to substance use. The successful identification of patients' substance use information equips clinical care teams to address substance-related issues more effectively, enabling targeted support and ultimately improving patient outcomes. Traditional natural language processing methods face limitations in accurately parsing diverse clinical language associated with substance use. Large language models offer promise in overcoming these challenges by adapting to diverse language patterns. This study investigates the application of the generative pretrained transformer (GPT) model in specific GPT-3.5 for extracting tobacco, alcohol, and substance use information from patient discharge summaries in zero-shot and few-shot learning settings. This study contributes to the evolving landscape of health care informatics by showcasing the potential of advanced language models in extracting nuanced information critical for enhancing patient care. The main data source for analysis in this paper is Medical Information Mart for Intensive Care III data set. Among all notes in this data set, we focused on discharge summaries. Prompt engineering was undertaken, involving an iterative exploration of diverse prompts. Leveraging carefully curated examples and refined prompts, we investigate the model's proficiency through zero-shot as well as few-shot prompting strategies. Results show GPT's varying effectiveness in identifying mentions of tobacco, alcohol, and substance use across learning scenarios. Zero-shot learning showed high accuracy in identifying substance use, whereas few-shot learning reduced accuracy but improved in identifying substance use status, enhancing recall and F1-score at the expense of lower precision. Excellence of zero-shot learning in precisely extracting text span mentioning substance use demonstrates its effectiveness in situations in which comprehensive recall is important. Conversely, few-shot learning offers advantages when accurately determining the status of substance use is the primary focus, even if it involves a trade-off in precision. The results contribute to enhancement of early detection and intervention strategies, tailor treatment plans with greater precision, and ultimately, contribute to a holistic understanding of patient health profiles. By integrating these artificial intelligence-driven methods into electronic health record systems, clinicians can gain immediate, comprehensive insights into substance use that results in shaping interventions that are not only timely but also more personalized and effective.

On the Opportunities and Challenges of Foundation Models for GeoAI (Vision Paper)

Large pre-trained models, also known as foundation models (FMs), are trained in a task-agnostic manner on large-scale data and can be adapted to a wide range of downstream tasks by fine-tuning, few-shot, or even zero-shot learning. Despite their successes in language and vision tasks, we have not yet seen an attempt to develop foundation models for geospatial artificial intelligence (GeoAI). In this work, we explore the promises and challenges of developing multimodal foundation models for GeoAI. We first investigate the potential of many existing FMs by testing their performances on seven tasks across multiple geospatial domains, including Geospatial Semantics, Health Geography, Urban Geography, and Remote Sensing. Our results indicate that on several geospatial tasks that only involve text modality, such as toponym recognition, location description recognition, and US state-level/county-level dementia time series forecasting, the task-agnostic large learning models (LLMs) can outperform task-specific fully supervised models in a zero-shot or few-shot learning setting. However, on other geospatial tasks, especially tasks that involve multiple data modalities (e.g., POI-based urban function classification, street view image–based urban noise intensity classification, and remote sensing image scene classification), existing FMs still underperform task-specific models. Based on these observations, we propose that one of the major challenges of developing an FM for GeoAI is to address the multimodal nature of geospatial tasks. After discussing the distinct challenges of each geospatial data modality, we suggest the possibility of a multimodal FM that can reason over various types of geospatial data through geospatial alignments. We conclude this article by discussing the unique risks and challenges to developing such a model for GeoAI.

Task-Adaptive Multi-Source Representations for Few-Shot Image Recognition

Conventional few-shot learning (FSL) mainly focuses on knowledge transfer from a single source dataset to a recognition scenario with only a few training samples available but still similar to the source domain. In this paper, we consider a more practical FSL setting where multiple semantically different datasets are available to address a wide range of FSL tasks, especially for some recognition scenarios beyond natural images, such as remote sensing and medical imagery. It can be referred to as multi-source cross-domain FSL. To tackle the problem, we propose a two-stage learning scheme, termed learning and adapting multi-source representations (LAMR). In the first stage, we propose a multi-head network to obtain efficient multi-domain representations, where all source domains share the same backbone except for the last parallel projection layers for domain specialization. We train the representations in a multi-task setting where each in-domain classification task is taken by a cosine classifier. In the second stage, considering that instance discrimination and class discrimination are crucial for robust recognition, we propose two contrastive objectives for adapting the pre-trained representations to be task-specialized on the few-shot data. Careful ablation studies verify that LAMR significantly improves representation transferability, showing consistent performance boosts. We also extend LAMR to single-source FSL by introducing a dataset-splitting strategy that equally splits one source dataset into sub-domains. The empirical results show that LAMR can achieve SOTA performance on the BSCD-FSL benchmark and competitive performance on mini-ImageNet, highlighting its versatility and effectiveness for FSL of both natural and specific imaging.

Few-shot defect classification via feature aggregation based on graph neural network

The effectiveness of deep learning models is greatly dependent on the availability of a vast amount of labeled data. However, in the realm of surface defect classification, acquiring and annotating defect samples proves to be quite challenging. Consequently, accurately predicting defect types with only a limited number of labeled samples has emerged as a prominent research focus in recent years. Few-shot learning, which leverages a restricted sample set in the support set, can effectively predict the categories of unlabeled samples in the query set. This approach is particularly well-suited for defect classification scenarios. In this article, we propose a transductive few-shot surface defect classification method, which using both the instance-level relations and distribution-level relations in each few-shot learning task. Furthermore, we calculate class center features in transductive manner and incorporate them into the feature aggregation operation to rectify the positioning of edge samples in the mapping space. This adjustment aims to minimize the distance between samples of the same category, thereby mitigating the influence of unlabeled samples at category boundary on classification accuracy. Experimental results on the public dataset show the outstanding performance of our proposed approach compared to the state-of-the-art methods in the few-shot learning settings. Our code is available at https://github.com/Harry10459/CIDnet.

Few-shot satellite image classification for bringing deep learning on board OPS-SAT

Bringing artificial intelligence on board Earth observation satellites unlocks unprecedented possibilities to extract actionable items from various image modalities at the global scale in real time. This is of paramount importance nowadays, as downlinking large amounts of imagery is not only prohibitively expensive but also time-consuming. However, building deep learning solutions that could be deployed on board an edge device is challenging due to the limited manually-annotated satellite datasets and hardware constraints of an edge device. This paper addresses these challenges through harnessing a blend of data-centric and model-centric approaches to build a well-generalizing yet efficient and resource-frugal deep learning model for multi-class satellite image classification in the few-shot learning settings. This integrated strategy is formulated to enhance classification performance while accommodating the unique demands of an image analysis chain on board OPS-SAT, a nanosatellite operated by the European Space Agency. The experiments performed over a real-world dataset of OPS-SAT images delves into the interactions between data- and model-centric techniques, underscores the significance of synthesizing artificial training data and emphasizes the value of ensemble learning. However, they also caution against negative transfer in domain adaptation. This study sheds light on effective model training strategies and highlights the multifaceted challenges inherent in deep learning for practical Earth observation, contributing insights to the field of satellite image classification within the constraints of nanosatellite operations. To ensure reproducibility of our study, the implementation is available at https://github.com/ShendoxParadox/Few-shot-satellite-image-classification-OPS-SAT.

An empirical study of Multimodal Entity-Based Sentiment Analysis with ChatGPT: Improving in-context learning via entity-aware contrastive learning

Multimodal Entity-Based Sentiment Analysis (MEBSA) is an emerging task within sentiment analysis, with the objective of simultaneously detecting entity, sentiment, and entity category from multimodal inputs. Despite achieving promising results, most existing MEBSA studies requires a substantial quantity of annotated data. The acquisition of such data is both costly and time-intensive in practical applications. To alleviate the reliance on annotated data, this work explores the potential of in-context learning (ICL) with a representative large language model, ChatGPT, for the MEBSA task. Specifically, we develop a general ICL framework with task instructions for zero-shot learning, followed by extending it to few-shot learning by incorporating a few demonstration samples in the prompt. To enhance the performance of the ICL framework in the few-shot learning setting, we further develop an Entity-Aware Contrastive Learning model to effectively retrieve demonstration samples that are similar to each test sample. Experiments demonstrate that our developed ICL framework exhibits superior performance over other baseline ICL methods, and is comparable to or even outperforms many existing fine-tuned methods on four MEBSA subtasks.

Adapting Vision-Language Models via Learning to Inject Knowledge.

Pre-trained vision-language models (VLM) such as CLIP, have demonstrated impressive zero-shot performance on various vision tasks. Trained on millions or even billions of image-text pairs, the text encoder has memorized a substantial amount of appearance knowledge. Such knowledge in VLM is usually leveraged by learning specific task-oriented prompts, which may limit its performance in unseen tasks. This paper proposes a new knowledge injection framework to pursue a generalizable adaption of VLM to downstream vision tasks. Instead of learning task-specific prompts, we extract task-agnostic knowledge features, and insert them into features of input images or texts. The fused features hence gain better discriminative capability and robustness to intra-category variances. Those knowledge features are generated by inputting learnable prompt sentences into text encoder of VLM, and extracting its multi-layer features. A new knowledge injection module (KIM) is proposed to refine text features or visual features using knowledge features. This knowledge injection framework enables both modalities to benefit from the rich knowledge memorized in the text encoder. Experiments show that our method outperforms recently proposed methods under few-shot learning, base-to-new classes generalization, cross-dataset transfer, and domain generalization settings. For instance, it outperforms CoOp by 4.5% under the few-shot learning setting, and CoCoOp by 4.4% under the base-to-new classes generalization setting. Our code will be released.

QAmeleon: Multilingual QA with Only 5 Examples

Abstract The availability of large, high-quality datasets has been a major driver of recent progress in question answering (QA). Such annotated datasets, however, are difficult and costly to collect, and rarely exist in languages other than English, rendering QA technology inaccessible to underrepresented languages. An alternative to building large monolingual training datasets is to leverage pre-trained language models (PLMs) under a few-shot learning setting. Our approach, QAmeleon, uses a PLM to automatically generate multilingual data upon which QA models are fine-tuned, thus avoiding costly annotation. Prompt tuning the PLM with only five examples per language delivers accuracy superior to translation-based baselines; it bridges nearly 60% of the gap between an English-only baseline and a fully-supervised upper bound fine-tuned on almost 50,000 hand-labeled examples; and consistently leads to improvements compared to directly fine-tuning a QA model on labeled examples in low resource settings. Experiments on the TyDiqa-GoldP and MLQA benchmarks show that few-shot prompt tuning for data synthesis scales across languages and is a viable alternative to large-scale annotation.1

Few-shot classification via efficient meta-learning with hybrid optimization

Meta-learning is one of the important methods to solve the challenging few-shot learning setting by using previous knowledge and experience to guide the learning of new tasks. Model-agnostic meta-learning (MAML) is one of the most popular meta-learning algorithms, and many variants of MAML have appeared in recent years. However, the performance of this algorithm for few-shot classification falls behind some other algorithms working on this problem. Therefore, its generalization performance needs to be further explored and improved. In view of the generalization problem, we found that MAML always shares an initialization in the process of parameter update, ignoring the bias between different tasks, resulting in limited generalization performance. On the other hand, the sample diversity of meta-learning model is low, and shallow network training is generally used, so it is difficult to obtain good performance based on deep neural network models. Based on these problems, we propose a hybrid optimization meta-learning method based on data augmentation, initialization attenuation, and resolution increase, called Mix-MAML. Experimental results show that our method reaches 76.93% classification accuracy on mini-ImageNet with 100 × 100 resolution, and 83.62% classification accuracy on CIFAR-FS with 80 × 80 resolution in the 5-way 5-shot settings under ResNet12, which achieves comparable or even better performance than other algorithms in some standard few-shot learning benchmarks without changing MAML simplicity and model-agnostic.

Prototype Completion for Few-Shot Learning.

Few-shot learning (FSL) aims to recognize novel classes with few examples. Pre-training based methods effectively tackle the problem by pre-training a feature extractor and then fine-tuning it through the nearest centroid based meta-learning. However, results show that the fine-tuning step makes marginal improvements. In this paper, 1) we figure out the reason, i.e., in the pre-trained feature space, the base classes already form compact clusters while novel classes spread as groups with large variances, which implies that fine-tuning feature extractor is less meaningful; 2) instead of fine-tuning feature extractor, we focus on estimating more representative prototypes. Consequently, we propose a novel prototype completion based meta-learning framework. This framework first introduces primitive knowledge (i.e., class-level part or attribute annotations) and extracts representative features for seen attributes as priors. Second, a part/attribute transfer network is designed to learn to infer the representative features for unseen attributes as supplementary priors. Finally, a prototype completion network is devised to learn to complete prototypes with these priors. Moreover, to avoid the prototype completion error, we further develop a Gaussian based prototype fusion strategy that fuses the mean-based and completed prototypes by exploiting the unlabeled samples. At last, we also develop an economic prototype completion version for FSL, which does not need to collect primitive knowledge, for a fair comparison with existing FSL methods without external knowledge. Extensive experiments show that our method: (i) obtains more accurate prototypes; (ii) achieves superior performance on both inductive and transductive FSL settings. Our codes are open-sourced at https://github.com/zhangbq-research/Prototype_Completion_for_FSL.

Are LSTMs good few-shot learners?

Deep learning requires large amounts of data to learn new tasks well, limiting its applicability to domains where such data is available. Meta-learning overcomes this limitation by learning how to learn. Hochreiter et al. (International conference on artificial neural networks, Springer, 2001) showed that an LSTM trained with backpropagation across different tasks is capable of meta-learning. Despite promising results of this approach on small problems, and more recently, also on reinforcement learning problems, the approach has received little attention in the supervised few-shot learning setting. We revisit this approach and test it on modern few-shot learning benchmarks. We find that LSTM, surprisingly, outperform the popular meta-learning technique MAML on a simple few-shot sine wave regression benchmark, but that LSTM, expectedly, fall short on more complex few-shot image classification benchmarks. We identify two potential causes and propose a new method called Outer Product LSTM (OP-LSTM) that resolves these issues and displays substantial performance gains over the plain LSTM. Compared to popular meta-learning baselines, OP-LSTM yields competitive performance on within-domain few-shot image classification, and performs better in cross-domain settings by 0.5–1.9% in accuracy score. While these results alone do not set a new state-of-the-art, the advances of OP-LSTM are orthogonal to other advances in the field of meta-learning, yield new insights in how LSTM work in image classification, allowing for a whole range of new research directions. For reproducibility purposes, we publish all our research code publicly.

Few-shot entity linking of food names

Entity linking (EL), the task of automatically matching mentions in text to concepts in a target knowledge base, remains under-explored when it comes to the food domain, despite its many potential applications, e.g., finding the nutritional value of ingredients in databases. In this paper, we describe the creation of new resources supporting the development of EL methods applied to the food domain: the E.Care Knowledge Base (E.Care KB) which contains 664 food concepts and the E.Care dataset, a corpus of 468 cooking recipes where ingredient names have been manually linked to corresponding concepts in the E.Care KB. We developed and evaluated different methods for EL, namely, deep learning-based approaches underpinned by Siamese networks trained under a few-shot learning setting, traditional machine learning-based approaches underpinned by support vector machines (SVMs) and unsupervised approaches based on string matching algorithms. Combining the strengths of each of these approaches, we built a hybrid model for food EL that balances the trade-offs between performance and inference speed. Specifically, our hybrid model obtains 89.40% accuracy and links mentions at an average speed of 0.24 seconds per mention, whereas our best deep learning-based model, SVM model and unsupervised model obtain accuracies of 86.99%, 87.19% and 87.43% at inference speeds of 0.007, 0.66 and 0.02 seconds per mention, respectively.

Exploring lottery ticket hypothesis in few-shot learning

Lottery Ticket Hypothesis (LTH) [14] has gathered great focus since being proposed. Researchers then succeed in figuring out alternative ways to find the ”winning ticket” and extending the vanilla version to various kinds of situations ranging from image segmentation to language pretraining. However, these works all emphasize fully supervised learning with plenty of training instances, whilst they ignore the important scenario of learning from few examples, i.e., Few-Shot Learning (FSL). Different from classical many-shot learning tasks, the common FSL setting assumes the disjoint of source and target categories. To the best of our knowledge, the lottery ticket hypothesis has for the first time, systematically studied in few-shot learning scenarios. To validate the hypothesis, we conduct extensive experiments on several few-shot learning methods with three widely-used datasets, miniImageNet, CUB, CIFARFS. Results reveal that we can even find “winning tickets” for some high-performance methods. In addition, our experiments on Cross-Domain FSL further validates the transferability of the found “winning tickets”. Furthermore, the process of finding LTH can be costly. So we study the early-stage LTH for FSL via exploring the Inverse Scale Space(ISS). Empirical results validate the efficacy of early-stage LTH.

Multi-stage distribution correction: A promising data augmentation method for few-shot fault diagnosis

Benefiting from the excellent capability of data processing, deep learning-based methods have been well applied in fault diagnosis. However, these methods may perform poorly due to lack of labeled data for training in real-world applications In few-shot learning settings, these methods may be trapped in overfitting since the data distribution estimated from a small number of labeled samples may be overly biased, which cannot cover the ground-truth data distribution. Given this problem, we propose a data augmentation method named Multi-Stage Distribution Correction (MSDC) for few-shot fault diagnosis. The proposed method can be divided into four training stages. In the first stage, unlabeled query features are clustered into several groups via an unsupervised Fuzzy C-means algorithm. Secondly, according to cosine similarity, the few labeled support features are assigned to the cluster with the highest similarity. Then, the Gaussian statistics of each cluster are extracted to correct the distributions of support features. Specifically, Our proposed method assumes that each dimension of the feature representation from the same class follows Gaussian distribution. New labeled features are generated by sampling from the corrected distributions. These generated features, along with the labeled support features, are used to train the task-specific classifier in the last stage. By joint training of the four stages, the scale of the labeled training dataset is effectively expanded, and the classification performance of the proposed method can be enhanced. The experimental results reveal that the proposed method outperforms the selected comparative methods in two case studies.

A metric-learning method for few-shot cross-event rumor detection

With the rapid development of social media, quickly detecting rumors on social media has become vitally crucial. However, there exists the following challenges. 1) Rumors are always flooded with time-critical events, where large-scale labeled datasets are difficult to obtain. 2) Although historic events have sufficient labels, the performance of models tends to degrade on the newly emergent events due to event difference shift. Facing these challenges, in this study, we attempt to leverage the idea of few-shot learning which aims to quickly acquire knowledge on unseen events with a few labeled samples by sufficiently learning from old events with a number of verified samples. Different from few-shot learning tasks in the literature, rumors in the same class of different events (no matter historical events or new events) are likely to contain coincident features in their embedding space. Therefore, we validate two classical metric learning methods, the prototypical network (PN) and the relation network (RN) which are able to capture the class-level representations in few-shot learning settings, to explore the effectiveness of metric learning methods for cross-event rumor detection. Our proposed model contains two stages corresponding to Base-Classifier pre-trained module and Base-Meta training module. The Base-Classifier pre-trained module is a classification model trained on old events. Then its last layer for class prediction is removed and the remains of the module is viewed as an encoder. The Base-Meta training module fine-tunes the encoder, meanwhile trains a selected metric learning model with episodic training strategies on old events. Empirical tests on novel events have showed that our model can outperform the state-of-the-art baseline models on the benchmark cross-event rumor datasets PHEME5 and PHEME9. What’s more, the model can improve the performance of cross-data rumor detection, where the model is trained on Twitter15, but tested on Twitter16.

Towards well-generalizing meta-learning via adversarial task augmentation

Meta-learning aims to use the knowledge from previous tasks to facilitate the learning of novel tasks. Many meta-learning models elaborately design various task-shared inductive bias, and learn it from a large number of tasks, so the generalization capability of the learned inductive bias depends on the diversity of the training tasks. A common assumption in meta-learning is that the training tasks and the test tasks come from the same or similar task distributions. However, this is usually not strictly satisfied in practice, so meta-learning models need to cope with various novel in-domain or cross-domain tasks. To this end, we propose to use task augmentation to increase the diversity of training tasks, thereby improving the generalization capability of meta-learning models. Concretely, we consider the worst-case problem around the base task distribution, and derive the adversarial task augmentation method which can generate inductive bias-adaptive ‘challenging’ tasks. Our method can be used as a simple plug-and-play module for various meta-learning models, and improve their generalization capability. We conduct extensive experiments under in-domain and cross-domain few-shot learning and unsupervised few-shot learning settings, and evaluate our method on different types of data (images and text). Experimental results show that our method can effectively improve the generalization capability of various meta-learning models under different settings.

An Adversarial Meta-Training Framework for Cross-Domain Few-Shot Learning

Meta-learning provides a promising way for deep learning models to efficiently learn in few-shot learning. With this capacity, many deep learning systems can be applied in many real applications. However, many existing meta-learning based few-shot learning systems suffer from vulnerable generalization when new tasks are from unseen domains (a.k.a, cross-domain few-shot learning). In this work, we consider this problem from the perspective of designing a model-agnostic meta-training framework to improve the generalization of existing meta-learning methods in cross-domain few-shot learning. In this way, compared with focusing on elaborately designing modules for a specific meta-learning model, our method is endowed with the ability to be compatible with different meta-learning models in various few-shot problems. To achieve this goal, a novel adversarial meta-training framework is proposed. The proposed framework utilizes max-min episodic iteration. In the episode of maximization, our framework focuses on how to dynamically generate appropriate pseudo tasks which benefit learning cross-domain knowledge. In the episode of minimization, our method aims to solve how to help meta-learning model learn cross-task and robust meta-knowledge. To comprehensively evaluate our framework, experiments are conducted on two few-shot learning settings, three meta-learning models, and eight datasets. These results demonstrate that our method is applicable to various meta-learning models in different few-shot learning problems. The superiority of our method is verified compared with existing state-of-the-art methods.

Multi-Learner Based Deep Meta-Learning for Few-Shot Medical Image Classification.

Few-shot learning (FSL) is promising in the field of medical image analysis due to high cost of establishing high-quality medical datasets. Many FSL approaches have been proposed in natural image scenes. However, present FSL methods are rarely evaluated on medical images and the FSL technology applicable to medical scenarios need to be further developed. Meta-learning has supplied an optional framework to address the challenging FSL setting. In this paper, we propose a novel multi-learner based FSL method for multiple medical image classification tasks, combining meta-learning with transfer-learning and metric-learning. Our designed model is composed of three learners, including auto-encoder, metric-learner and task-learner. In transfer-learning, all the learners are trained on the base classes. In the ensuing meta-learning, we leverage multiple novel tasks to fine-tune the metric-learner and task-learner in order to fast adapt to unseen tasks. Moreover, to further boost the learning efficiency of our model, we devised real-time data augmentation and dynamic Gaussian disturbance soft label (GDSL) scheme as effective generalization strategies of few-shot classification tasks. We have conducted experiments for three-class few-shot classification tasks on three newly-built challenging medical benchmarks, BLOOD, PATH and CHEST. Extensive comparisons to related works validated that our method achieved top performance both on homogeneous medical datasets and cross-domain datasets.

Thermal to Visual Person Re-Identification Using Collaborative Metric Learning Based on Maximum Margin Matrix Factorization

Thermal to visual person re-identification (T2V-ReID) is a cross-domain image retrieval problem. In this problem, the matching of a person’s image takes place, where the image is taken by different cameras (thermal and visual) at different times. This problem has numerous applications in night-time security surveillance. It is challenging due to the large intra-class variations and cross-domain discrepancies. Recently, deep metric learning methods are proposed for this problem. Still, there is a scope to improve the metric learning by generalizing the metric. In this paper, we have proposed the collaborative metric learning using Maximum Margin Matrix Factorization. It uses the group-wise similarities and collaboratively predicts the similarities. We can learn a more generalized metric by utilizing the maximized margin in this method. The proposed method is tested on the RegDB and RGB-D-T data sets, and the method outperforms the existing works in the few-shot learning settings.

MDFM: Multi-Decision Fusing Model for Few-Shot Learning

In recent years, researchers pay growing attention to the few-shot learning (FSL) task to address the data-scarce problem. A standard FSL framework is composed of two components: i) Pre-train. Employ the base data to generate a CNN-based feature extraction model (FEM). ii) Meta-test. Apply the trained FEM to the novel data (category is different from base data) to acquire the feature embeddings and recognize them. Although researchers have made remarkable breakthroughs in FSL, there still exists a fundamental problem. Since the trained FEM with base data usually cannot adapt to the novel class flawlessly, the novel data’s feature may lead to the distribution shift problem. To address this challenge, we hypothesize that even if most of the decisions based on different FEMs are viewed as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">weak decisions</i> , which are not available for all classes, they still perform decent in some specific categories. Inspired by this assumption, we propose a novel method Multi-Decision Fusing Model (MDFM), which comprehensively considers the decisions based on multiple FEMs to enhance the efficacy and robustness of the model. MDFM is a simple, flexible, non-parametric method that can directly apply to the existing FEMs. Besides, we extend the proposed MDFM to two FSL settings (e.g., supervised and semi-supervised settings). We evaluate the proposed method on five benchmark datasets and achieve significant improvements of 3.4%-7.3% compared with state-of-the-arts.