After discovering that Language Models (LMs) can be good in-context few-shot learners, numerous strategies have been proposed to optimize in-context sequence configurations. Recently, researchers in Vision-Language (VL) domains also develop their few-shot learners, while they only use the simplest way, i.e., randomly sampling, to configure in-context image-text pairs. In order to explore the effects of varying configurations on VL in-context learning, we devised four strategies for image selection and four for caption assignment to configure in-context image-text pairs for image captioning. Here Image Captioning is used as the case study since it can be seen as the visually-conditioned LM. Our comprehensive experiments yield two counter-intuitive but valuable insights, highlighting the distinct characteristics of VL in-context learning due to multi-modal synergy, as compared to the NLP case.
Under partial-label learning (PLL) where, for each training instance, only a set of ambiguous candidate labels containing the unknown true label is accessible, contrastive learning has recently boosted the performance of PLL on vision tasks, attributed to representations learned by contrasting the same/different classes of entities. Without access to true labels, positive points are predicted using pseudo-labels that are inherently noisy, and negative points often require large batches or momentum encoders, resulting in unreliable similarity information and a high computational overhead. In this paper, we rethink a state-of-the-art contrastive PLL method PiCO[24], inspiring the design of a simple framework termed PaPi (Partial-label learning with a guided Prototypical classifier), which demonstrates significant scope for improvement in representation learning, thus contributing to label disambiguation. PaPi guides the optimization of a prototypical classifier by a linear classifier with which they share the same feature encoder, thus explicitly encouraging the representation to reflect visual similarity between categories. It is also technically appealing, as PaPi requires only a few components in PiCO with the opposite direction of guidance, and directly eliminates the contrastive learning module that would introduce noise and consume computational resources. We empirically demonstrate that PaPi significantly outperforms other PLL methods on various image classification tasks.
During the continuous evolution of one organism's ancestry, its genes accumulate extensive experiences and knowledge, enabling newborn descendants to rapidly adapt to their specific environments. Motivated by this observation, we propose a novel machine learning paradigm \textit{Learngene} to enable learning models to incorporate three key characteristics of genes. (i) Accumulating: the knowledge is accumulated during the continuous learning of an \textbf{ancestry model}. (ii) Condensing: the exhaustive accumulated knowledge is condensed into a much more compact information piece, \ie \textbf{learngene}. (iii): Inheriting: the condensed \textbf{learngene} is inherited to make it easier for \textbf{descendant models} to adapt to new environments. Since accumulating has been studied in some well-developed paradigms like large-scale pre-training and lifelong learning, we focus on condensing and inheriting, which induces three key issues and we provide the preliminary solutions to these issues in this paper: (i) \textit{Learngene} Form: the \textbf{learngene} is set to a few integral layers that can preserve the most commonality. (ii) \textit{Learngene} Condensing: we identify which layers among the ancestry model have the most similarity as one pseudo descendant model. (iii) \textit{Learngene} Inheriting: to construct distinct descendant models for specific downstream tasks, we stack some randomly initialized layers to the \textbf{learngene} layers. Extensive experiments of various settings, including using different network architectures like Vision Transformer (ViT) and Convolutional Neural Networks (CNNs) on different datasets, are carried out to confirm five advantages and two characteristics of \textit{Learngene}.
Despite the significant success of deep learning in computer vision tasks, cross-domain tasks still present a challenge in which the model's performance will degrade when the training set and the test set follow different distributions. Most existing methods employ adversarial learning or instance normalization for achieving data augmentation to solve this task. In contrast, considering that the batch normalization (BN) layer may not be robust for unseen domains and there exist the differences between local patches of an image, we propose a novel method called patch-aware batch normalization (PBN). To be specific, we first split feature maps of a batch into non-overlapping patches along the spatial dimension, and then independently normalize each patch to jointly optimize the shared BN parameter at each iteration. By exploiting the differences between local patches of an image, our proposed PBN can effectively enhance the robustness of the model's parameters. Besides, considering the statistics from each patch may be inaccurate due to their smaller size compared to the global feature maps, we incorporate the globally accumulated statistics with the statistics from each batch to obtain the final statistics for normalizing each patch. Since the proposed PBN can replace the typical BN, it can be integrated into most existing state-of-the-art methods. Extensive experiments and analysis demonstrate the effectiveness of our PBN in multiple computer vision tasks, including classification, object detection, instance retrieval, and semantic segmentation.
Denoising diffusion models have been a mainstream approach for image generation, however, training these models often suffers from slow convergence. In this paper, we discovered that the slow convergence is partly due to conflicting optimization directions between timesteps. To address this issue, we treat the diffusion training as a multi-task learning problem, and introduce a simple yet effective approach referred to as Min-SNR-$\gamma$. This method adapts loss weights of timesteps based on clamped signal-to-noise ratios, which effectively balances the conflicts among timesteps. Our results demonstrate a significant improvement in converging speed, 3.4$\times$ faster than previous weighting strategies. It is also more effective, achieving a new record FID score of 2.06 on the ImageNet $256\times256$ benchmark using smaller architectures than that employed in previous state-of-the-art. The code is available at https://github.com/TiankaiHang/Min-SNR-Diffusion-Training.
Label distribution (LD) uses the description degree to describe instances, which provides more fine-grained supervision information when learning with label ambiguity. Nevertheless, LD is unavailable in many real-world applications. To obtain LD, label enhancement (LE) has emerged to recover LD from logical label. Existing LE approach have the following problems: (\textbf{i}) They use logical label to train mappings to LD, but the supervision information is too loose, which can lead to inaccurate model prediction; (\textbf{ii}) They ignore feature redundancy and use the collected features directly. To solve (\textbf{i}), we use the topology of the feature space to generate more accurate label-confidence. To solve (\textbf{ii}), we proposed a novel supervised LE dimensionality reduction approach, which projects the original data into a lower dimensional feature space. Combining the above two, we obtain the augmented data for LE. Further, we proposed a novel nonlinear LE model based on the label-confidence and reduced features. Extensive experiments on 12 real-world datasets are conducted and the results show that our method consistently outperforms the other five comparing approaches.
Label distribution learning (LDL) trains a model to predict the relevance of a set of labels (called label distribution (LD)) to an instance. The previous LDL methods all assumed the LDs of the training instances are accurate. However, annotating highly accurate LDs for training instances is time-consuming and very expensive, and in reality the collected LD is usually inaccurate and disturbed by annotating errors. For the first time, this paper investigates the problem of inaccurate LDL, i.e., developing an LDL model with noisy LDs. Specifically, we assume the noisy LD matrix is the linear combination of an ideal LD matrix and a sparse noisy matrix. Accordingly, inaccurate LDL becomes an inverse problem, i.e., recovering the ideal LD and noise matrix from the inaccurate LDs. To this end, we assume the ideal LD matrix is low-rank due to the correlation of labels. Besides, we use the local geometric structure of instances captured by a graph to assist the ideal LD recovery as if two instances are similar to each other, they are likely to share the same LD. The proposed model is finally formulated as a graph-regularized low-rank and sparse decomposition problem and numerically solved by the alternating direction method of multipliers. Extensive experiments demonstrate that our method can recover a relatively accurate LD from the inaccurate LD and promote the performance of different LDL methods with inaccurate LD.
Partial label learning (PLL) is a typical weakly supervised learning problem in which each instance is associated with a candidate label set, and among which only one is true. However, the assumption that the ground-truth label is always among the candidate label set would be unrealistic, as the reliability of the candidate label sets in real-world applications cannot be guaranteed by annotators. Therefore, a generalized PLL named Unreliable Partial Label Learning (UPLL) is proposed, in which the true label may not be in the candidate label set. Due to the challenges posed by unreliable labeling, previous PLL methods will experience a marked decline in performance when applied to UPLL. To address the issue, we propose a two-stage framework named Unreliable Partial Label Learning with Recursive Separation (UPLLRS). In the first stage, the self-adaptive recursive separation strategy is proposed to separate the training set into a reliable subset and an unreliable subset. In the second stage, a disambiguation strategy is employed to progressively identify the ground-truth labels in the reliable subset. Simultaneously, semi-supervised learning methods are adopted to extract valuable information from the unreliable subset. Our method demonstrates state-of-the-art performance as evidenced by experimental results, particularly in situations of high unreliability.
Knowledge distillation has been widely adopted in a variety of tasks and has achieved remarkable successes. Since its inception, many researchers have been intrigued by the dark knowledge hidden in the outputs of the teacher model. Recently, a study has demonstrated that knowledge distillation and label smoothing can be unified as learning from soft labels. Consequently, how to measure the effectiveness of the soft labels becomes an important question. Most existing theories have stringent constraints on the teacher model or data distribution, and many assumptions imply that the soft labels are close to the ground-truth labels. This paper studies whether biased soft labels are still effective. We present two more comprehensive indicators to measure the effectiveness of such soft labels. Based on the two indicators, we give sufficient conditions to ensure biased soft label based learners are classifier-consistent and ERM learnable. The theory is applied to three weakly-supervised frameworks. Experimental results validate that biased soft labels can also teach good students, which corroborates the soundness of the theory.