Knowledge distillation (KD) is widely used for compressing a teacher model to a smaller student model, reducing its inference cost and memory footprint while preserving model capabilities. However, current KD methods for auto-regressive sequence models (e.g., large language models) suffer from missing a standardized objective function. Moreover, the recent use of student-generated outputs to address training-inference mismatches has significantly escalated computational costs. To tackle these issues, we introduce DistiLLM, a more effective and efficient KD framework for auto-regressive language models. DistiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs. Extensive experiments, including instruction-following tasks, demonstrate the effectiveness of DistiLLM in building high-performing student models while achieving up to 4.3$\times$ speedup compared to recent KD methods.
Albeit great performance of Transformer-based speech selfsupervised learning (SSL) models, their large parameter size and computational cost make them unfavorable to utilize. In this study, we propose to compress the speech SSL models by distilling speech temporal relation (STaR). Unlike previous works that directly match the representation for each speech frame, STaR distillation transfers temporal relation between speech frames, which is more suitable for lightweight student with limited capacity. We explore three STaR distillation objectives and select the best combination as the final STaR loss. Our model distilled from HuBERT BASE achieves an overall score of 79.8 on SUPERB benchmark, the best performance among models with up to 27 million parameters. We show that our method is applicable across different speech SSL models and maintains robust performance with further reduced parameters.
The recent progress in diffusion-based text-to-image generation models has significantly expanded generative capabilities via conditioning the text descriptions. However, since relying solely on text prompts is still restrictive for fine-grained customization, we aim to extend the boundaries of conditional generation to incorporate diverse types of modalities, e.g., sketch, box, and style embedding, simultaneously. We thus design a multimodal text-to-image diffusion model, coined as DiffBlender, that achieves the aforementioned goal in a single model by training only a few small hypernetworks. DiffBlender facilitates a convenient scaling of input modalities, without altering the parameters of an existing large-scale generative model to retain its well-established knowledge. Furthermore, our study sets new standards for multimodal generation by conducting quantitative and qualitative comparisons with existing approaches. By diversifying the channels of conditioning modalities, DiffBlender faithfully reflects the provided information or, in its absence, creates imaginative generation.
Respiratory sound contains crucial information for the early diagnosis of fatal lung diseases. Since the COVID-19 pandemic, there has been a growing interest in contact-free medical care based on electronic stethoscopes. To this end, cutting-edge deep learning models have been developed to diagnose lung diseases; however, it is still challenging due to the scarcity of medical data. In this study, we demonstrate that the pretrained model on large-scale visual and audio datasets can be generalized to the respiratory sound classification task. In addition, we introduce a straightforward Patch-Mix augmentation, which randomly mixes patches between different samples, with Audio Spectrogram Transformer (AST). We further propose a novel and effective Patch-Mix Contrastive Learning to distinguish the mixed representations in the latent space. Our method achieves state-of-the-art performance on the ICBHI dataset, outperforming the prior leading score by an improvement of 4.08%.
Transformer-based speech self-supervised learning (SSL) models, such as HuBERT, show surprising performance in various speech processing tasks. However, huge number of parameters in speech SSL models necessitate the compression to a more compact model for wider usage in academia or small companies. In this study, we suggest to reuse attention maps across the Transformer layers, so as to remove key and query parameters while retaining the number of layers. Furthermore, we propose a novel masking distillation strategy to improve the student model's speech representation quality. We extend the distillation loss to utilize both masked and unmasked speech frames to fully leverage the teacher model's high-quality representation. Our universal compression strategy yields the student model that achieves phoneme error rate (PER) of 7.72% and word error rate (WER) of 9.96% on the SUPERB benchmark.
Deep learning in general domains has constantly been extended to domain-specific tasks requiring the recognition of fine-grained characteristics. However, real-world applications for fine-grained tasks suffer from two challenges: a high reliance on expert knowledge for annotation and necessity of a versatile model for various downstream tasks in a specific domain (e.g., prediction of categories, bounding boxes, or pixel-wise annotations). Fortunately, the recent self-supervised learning (SSL) is a promising approach to pretrain a model without annotations, serving as an effective initialization for any downstream tasks. Since SSL does not rely on the presence of annotation, in general, it utilizes the large-scale unlabeled dataset, referred to as an open-set. In this sense, we introduce a novel Open-Set Self-Supervised Learning problem under the assumption that a large-scale unlabeled open-set is available, as well as the fine-grained target dataset, during a pretraining phase. In our problem setup, it is crucial to consider the distribution mismatch between the open-set and target dataset. Hence, we propose SimCore algorithm to sample a coreset, the subset of an open-set that has a minimum distance to the target dataset in the latent space. We demonstrate that SimCore significantly improves representation learning performance through extensive experimental settings, including eleven fine-grained datasets and seven open-sets in various downstream tasks.
Most of the recent few-shot learning algorithms are based on transfer learning, where a model is pre-trained using a large amount of source data, and the pre-trained model is updated using a small amount of target data afterward. In transfer-based few-shot learning, sophisticated pre-training methods have been widely studied for universal and improved representation. However, there is little study on updating pre-trained models for few-shot learning. In this paper, we compare the two popular updating methods, fine-tuning (i.e., updating the entire network) and linear probing (i.e., updating only the linear classifier), considering the distribution shift between the source and target data. We find that fine-tuning is better than linear probing as the number of samples increases, regardless of distribution shift. Next, we investigate the effectiveness and ineffectiveness of data augmentation when pre-trained models are fine-tuned. Our fundamental analyses demonstrate that careful considerations of the details about updating pre-trained models are required for better few-shot performance.
Cross-domain few-shot learning (CD-FSL), where there are few target samples under extreme differences between source and target domains, has recently attracted huge attention. For CD-FSL, recent studies generally have developed transfer learning based approaches that pre-train a neural network on popular labeled source domain datasets and then transfer it to target domain data. Although the labeled datasets may provide suitable initial parameters for the target data, the domain difference between the source and target might hinder the fine-tuning on the target domain. This paper proposes a simple yet powerful method that re-randomizes the parameters fitted on the source domain before adapting to the target data. The re-randomization resets source-specific parameters of the source pre-trained model and thus facilitates fine-tuning on the target domain, improving few-shot performance.
Cross-domain few-shot learning has drawn increasing attention for handling large differences between the source and target domains--an important concern in real-world scenarios. To overcome these large differences, recent works have considered exploiting small-scale unlabeled data from the target domain during the pre-training stage. This data enables self-supervised pre-training on the target domain, in addition to supervised pre-training on the source domain. In this paper, we empirically investigate scenarios under which it is advantageous to use each pre-training scheme, based on domain similarity and few-shot difficulty: performance gain of self-supervised pre-training over supervised pre-training increases when domain similarity is smaller or few-shot difficulty is lower. We further design two pre-training schemes, mixed-supervised and two-stage learning, that improve performance. In this light, we present seven findings for CD-FSL which are supported by extensive experiments and analyses on three source and eight target benchmark datasets with varying levels of domain similarity and few-shot difficulty. Our code is available at https://anonymous.4open.science/r/understandingCDFSL.