Learning with limited data is a key challenge for visual recognition. Many few-shot learning methods address this by an instance embedding function from seen classes and apply the to instances unseen labels. This style of transfer task-agnostic: not learned optimally discriminative respect classes, where discerning among them leads target task. In paper, we propose novel approach adapt embeddings classification task set-to-set function, yielding that are task-specific discriminative. We...

Traditional classifiers are deployed under closed-set setting, with both training and test classes belong to the same set. However, real-world applications probably face input of unknown categories, model will recognize them as known ones. Under such circumstances, open-set recognition is proposed maintain classification performance on reject unknowns. The models make overconfident predictions over familiar class instances, so that calibration thresholding across categories become essential...

Novel classes frequently arise in our dynamically changing world, e.g., new users the authentication system, and a machine learning model should recognize without forgetting old ones. This scenario becomes more challenging when class instances are insufficient, which is called few-shot class-incremental (FSCIL). Cur-rent methods handle incremental retrospectively by making updated similar to one. By contrast, we suggest prospectively prepare for future updates, propose ForwArd Compatible...

Deep models, e.g., CNNs and Vision Transformers, have achieved impressive achievements in many vision tasks the closed world. However, novel classes emerge from time to our ever-changing world, requiring a learning system acquire new knowledge continually. Class-Incremental Learning (CIL) enables learner incorporate of incrementally build universal classifier among all seen classes. Correspondingly, when directly training model with class instances, fatal problem occurs -- tends...

Multivariate time series data comprises various channels of variables. The multivariate forecasting models need to capture the relationship between accurately predict future values. However, recently, there has been an emergence methods that employ Channel Independent (CI) strategy. These view as separate univariate and disregard correlation channels. Surprisingly, our empirical results have shown trained with CI strategy outperform those Dependent (CD) strategy, usually by a significant...

Deep models, e.g., CNNs and Vision Transformers, have achieved impressive achievements in many vision tasks the closed world.However, novel classes emerge from time to our ever-changing world, requiring a learning system acquire new knowledge continually.Class-Incremental Learning (CIL) enables learner incorporate of incrementally build universal classifier among all seen classes.Correspondingly, when directly training model with class instances, fatal problem occurs -the tends...

New classes arise frequently in our ever-changing world, e.g., emerging topics social media and new types of products e-commerce. A model should recognize meanwhile maintain discriminability over old classes. Under severe circumstances, only limited novel instances are available to incrementally update the model. The task recognizing few-shot without forgetting is called class-incremental learning (FSCIL). In this work, we propose a paradigm for FSCIL based on meta-learning by LearnIng...

Nowadays, real-world applications often face streaming data, which requires the learning system to absorb new knowledge as data evolves. Continual Learning (CL) aims achieve this goal and meanwhile overcome catastrophic forgetting of former when ones. Typical CL methods build model from scratch grow with incoming data. However, advent pre-trained (PTM) era has sparked immense research interest, particularly in leveraging PTMs' robust representational capabilities. This paper presents a...

Class-Incremental Learning (CIL) or continual learning is a desired capability in the real world, which requires system to adapt new tasks without forgetting former ones. While traditional CIL methods focus on visual information grasp core features, recent advances Vision-Language Models (VLM) have shown promising capabilities generalizable representations with aid of textual information. However, when continually trained classes, VLMs often suffer from catastrophic knowledge. Applying poses...

Classifiers trained with class-imbalanced data are known to perform poorly on test of the "minor" classes, which we have insufficient training data. In this paper, investigate learning a ConvNet classifier under such scenario. We found that significantly over-fits minor is quite opposite traditional machine algorithms often under-fit classes. conducted series analysis and discovered feature deviation phenomenon -- learned generates deviated features between classes explains how over-fitting...

Traditional learning systems are trained in closed-world for a fixed number of classes, and need pre-collected datasets advance. However, new classes often emerge real-world applications should be learned incrementally. For example, electronic commerce, types products appear daily, social media community, topics frequently. Under such circumstances, incremental models learn several at time without forgetting. We find strong correlation between old learning, which can applied to relate...

While deep learning demonstrates its strong ability to handle independent and identically distributed (IID) data, it often suffers from out-of-distribution (OoD) generalization, where the test data come another distribution (w.r.t. training one). Designing a general OoD generalization framework for wide range of applications is challenging, mainly due different kinds shifts in real world, such as shift across domains or extrapolation correlation. Most previous approaches can only solve one...

Few-shot learning (FSL) aims to generate a classifier using limited labeled examples. Many existing works take the meta-learning approach, constructing few-shot learner (a meta-model) that can learn from examples classifier. Typically, is constructed or meta-trained by sampling multiple tasks in turn and optimizing learner's performance generating classifiers for those tasks. The measured how well resulting classify test (i.e., query) of In this paper, we point out two potential weaknesses...

Real-world applications require the classification model to adapt new classes without forgetting old ones. Correspondingly, Class-Incremental Learning (CIL) aims train a with limited memory size meet this requirement. Typical CIL methods tend save representative exemplars from former resist forgetting, while recent works find that storing models history can substantially boost performance. However, stored are not counted into budget, which implicitly results in unfair comparisons. We when...

Multivariate time series forecasting plays a crucial role in various fields such as finance, traffic management, energy, and healthcare. Recent studies have highlighted the advantages of channel independence to resist distribution drift but neglect correlations, limiting further enhancements. Several methods utilize mechanisms like attention or mixer address this by capturing they either introduce excessive complexity rely too heavily on correlation achieve satisfactory results under drifts,...

The inherent complexity of image semantics engenders a fascinating variability in relationships between images. For instance, under certain condition, two images may demonstrate similarity, while different circumstances, the same pair could exhibit absolute dissimilarity. A singular feature space is therefore insufficient for capturing nuanced semantic that exist samples. Conditional Similarity Learning (CSL) aims to address this gap by learning multiple, distinct spaces. Existing approaches...