With the advances in deep learning, the performance of end-to-end (E2E) single-task models for speech and audio processing has been constantly improving. However, it is still challenging to build a general-purpose model with high performance on multiple tasks, since different speech and audio processing tasks usually require different training data, input features, or model architectures to achieve optimal performance. In this work, MT2KD, a novel two-stage multi-task learning framework is proposed to build a general-purpose speech and audio encoder that jointly performs three fundamental tasks: automatic speech recognition (ASR), audio tagging (AT) and speaker verification (SV). In the first stage, multi-teacher knowledge distillation (KD) is applied to align the feature spaces of three single-task high-performance teacher encoders into a single student encoder using the same unlabelled data. In the second stage, multi-task supervised fine-tuning is carried out by initialising the model from the first stage and training on the separate labelled data of each single task. Experiments demonstrate that the proposed multi-task training pipeline significantly outperforms a baseline model trained with multi-task learning from scratch. The final system achieves good performance on ASR, AT and SV: with less than 4% relative word-error-rate increase on ASR, only 1.9 lower mean averaged precision on AT and 0.23% absolute higher equal error rate on SV compared to the best-performing single-task encoders, using only a 66M total model parameters.
Knowledge distillation typically employs the Kullback-Leibler (KL) divergence to constrain the student model's output to match the soft labels provided by the teacher model exactly. However, sometimes the optimization direction of the KL divergence loss is not always aligned with the task loss, where a smaller KL divergence could lead to erroneous predictions that diverge from the soft labels. This limitation often results in suboptimal optimization for the student. Moreover, even under temperature scaling, the KL divergence loss function tends to overly focus on the larger-valued channels in the logits, disregarding the rich inter-class information provided by the multitude of smaller-valued channels. This hard constraint proves too challenging for lightweight students, hindering further knowledge distillation. To address this issue, we propose a plug-and-play ranking loss based on Kendall's $\tau$ coefficient, called Rank-Kendall Knowledge Distillation (RKKD). RKKD balances the attention to smaller-valued channels by constraining the order of channel values in student logits, providing more inter-class relational information. The rank constraint on the top-valued channels helps avoid suboptimal traps during optimization. We also discuss different differentiable forms of Kendall's $\tau$ coefficient and demonstrate that the proposed ranking loss function shares a consistent optimization objective with the KL divergence. Extensive experiments on the CIFAR-100 and ImageNet datasets show that our RKKD can enhance the performance of various knowledge distillation baselines and offer broad improvements across multiple teacher-student architecture combinations.
Many medical image segmentation methods have achieved impressive results. Yet, most existing methods do not take into account the shape-intensity prior information. This may lead to implausible segmentation results, in particular for images of unseen datasets. In this paper, we propose a novel approach to incorporate joint shape-intensity prior information into the segmentation network. Specifically, we first train a segmentation network (regarded as the teacher network) on class-wise averaged training images to extract valuable shape-intensity information, which is then transferred to a student segmentation network with the same network architecture as the teacher via knowledge distillation. In this way, the student network regarded as the final segmentation model can effectively integrate the shape-intensity prior information, yielding more accurate segmentation results. Despite its simplicity, experiments on five medical image segmentation tasks of different modalities demonstrate that the proposed Shape-Intensity Knowledge Distillation (SIKD) consistently improves several baseline models (including recent MaxStyle and SAMed) under intra-dataset evaluation, and significantly improves the cross-dataset generalization ability. The code is available at https://github.com/whdong-whu/SIKD.
Despite being widely applied due to their exceptional capabilities, Large Language Models (LLMs) have been proven to be vulnerable to backdoor attacks. These attacks introduce targeted vulnerabilities into LLMs by poisoning training samples and full-parameter fine-tuning. However, this kind of backdoor attack is limited since they require significant computational resources, especially as the size of LLMs increases. Besides, parameter-efficient fine-tuning (PEFT) offers an alternative but the restricted parameter updating may impede the alignment of triggers with target labels. In this study, we first verify that backdoor attacks with PEFT may encounter challenges in achieving feasible performance. To address these issues and improve the effectiveness of backdoor attacks with PEFT, we propose a novel backdoor attack algorithm from weak to strong based on feature alignment-enhanced knowledge distillation (W2SAttack). Specifically, we poison small-scale language models through full-parameter fine-tuning to serve as the teacher model. The teacher model then covertly transfers the backdoor to the large-scale student model through feature alignment-enhanced knowledge distillation, which employs PEFT. Theoretical analysis reveals that W2SAttack has the potential to augment the effectiveness of backdoor attacks. We demonstrate the superior performance of W2SAttack on classification tasks across four language models, four backdoor attack algorithms, and two different architectures of teacher models. Experimental results indicate success rates close to 100% for backdoor attacks targeting PEFT.
The training of automatic speech recognition (ASR) with little to no supervised data remains an open question. In this work, we demonstrate that streaming Transformer-Transducer (TT) models can be trained from scratch in consumer and accessible GPUs in their entirety with pseudo-labeled (PL) speech from foundational speech models (FSM). This allows training a robust ASR model just in one stage and does not require large data and computational budget compared to the two-step scenario with pre-training and fine-tuning. We perform a comprehensive ablation on different aspects of PL-based streaming TT models such as the impact of (1) shallow fusion of n-gram LMs, (2) contextual biasing with named entities, (3) chunk-wise decoding for low-latency streaming applications, and (4) TT overall performance as the function of the FSM size. Our results demonstrate that TT can be trained from scratch without supervised data, even with very noisy PLs. We validate the proposed framework on 6 languages from CommonVoice and propose multiple heuristics to filter out hallucinated PLs.
Optical flow has made great progress in clean scenes, while suffers degradation under adverse weather due to the violation of the brightness constancy and gradient continuity assumptions of optical flow. Typically, existing methods mainly adopt domain adaptation to transfer motion knowledge from clean to degraded domain through one-stage adaptation. However, this direct adaptation is ineffective, since there exists a large gap due to adverse weather and scene style between clean and real degraded domains. Moreover, even within the degraded domain itself, static weather (e.g., fog) and dynamic weather (e.g., rain) have different impacts on optical flow. To address above issues, we explore synthetic degraded domain as an intermediate bridge between clean and real degraded domains, and propose a cumulative homogeneous-heterogeneous adaptation framework for real adverse weather optical flow. Specifically, for clean-degraded transfer, our key insight is that static weather possesses the depth-association homogeneous feature which does not change the intrinsic motion of the scene, while dynamic weather additionally introduces the heterogeneous feature which results in a significant boundary discrepancy in warp errors between clean and degraded domains. For synthetic-real transfer, we figure out that cost volume correlation shares a similar statistical histogram between synthetic and real degraded domains, benefiting to holistically aligning the homogeneous correlation distribution for synthetic-real knowledge distillation. Under this unified framework, the proposed method can progressively and explicitly transfer knowledge from clean scenes to real adverse weather. In addition, we further collect a real adverse weather dataset with manually annotated optical flow labels and perform extensive experiments to verify the superiority of the proposed method.
When developing Computer Aided Detection (CAD) systems for Digital Breast Tomosynthesis (DBT), the complexity arising from the volumetric nature of the modality poses significant technical challenges for obtaining large-scale accurate annotations. Without access to large-scale annotations, the resulting model may not generalize to different domains. Given the costly nature of obtaining DBT annotations, how to effectively increase the amount of data used for training DBT CAD systems remains an open challenge. In this paper, we present SelectiveKD, a semi-supervised learning framework for building cancer detection models for DBT, which only requires a limited number of annotated slices to reach high performance. We achieve this by utilizing unlabeled slices available in a DBT stack through a knowledge distillation framework in which the teacher model provides a supervisory signal to the student model for all slices in the DBT volume. Our framework mitigates the potential noise in the supervisory signal from a sub-optimal teacher by implementing a selective dataset expansion strategy using pseudo labels. We evaluate our approach with a large-scale real-world dataset of over 10,000 DBT exams collected from multiple device manufacturers and locations. The resulting SelectiveKD process effectively utilizes unannotated slices from a DBT stack, leading to significantly improved cancer classification performance (AUC) and generalization performance.
The proliferation of Artificial Neural Networks (ANNs) has led to increased energy consumption, raising concerns about their sustainability. Spiking Neural Networks (SNNs), which are inspired by biological neural systems and operate using sparse, event-driven spikes to communicate information between neurons, offer a potential solution due to their lower energy requirements. An alternative technique for reducing a neural network's footprint is quantization, which compresses weight representations to decrease memory usage and energy consumption. In this study, we present Twin Network Augmentation (TNA), a novel training framework aimed at improving the performance of SNNs while also facilitating an enhanced compression through low-precision quantization of weights. TNA involves co-training an SNN with a twin network, optimizing both networks to minimize their cross-entropy losses and the mean squared error between their output logits. We demonstrate that TNA significantly enhances classification performance across various vision datasets and in addition is particularly effective when applied when reducing SNNs to ternary weight precision. Notably, during inference , only the ternary SNN is retained, significantly reducing the network in number of neurons, connectivity and weight size representation. Our results show that TNA outperforms traditional knowledge distillation methods and achieves state-of-the-art performance for the evaluated network architecture on benchmark datasets, including CIFAR-10, CIFAR-100, and CIFAR-10-DVS. This paper underscores the effectiveness of TNA in bridging the performance gap between SNNs and ANNs and suggests further exploration into the application of TNA in different network architectures and datasets.
For reasons such as privacy, there are use cases for language models at the edge. This has given rise to small language models (SLMs) targeted for deployment in resource-constrained devices where energy efficiency is a significant concern. Spiking neural networks (SNNs) offer a promising solution due to their energy efficiency, and there are already works on realizing transformer-based models on SNNs. However, key operations like softmax and layer normalization (LN) are difficult to implement on neuromorphic hardware, and many of these early works sidestepped them. To address these challenges, we introduce Sorbet, a transformer-based spiking language model that is more neuromorphic hardware-compatible. Sorbet incorporates a novel shifting-based softmax called PTsoftmax and a power normalization method using bit-shifting (BSPN), both designed to replace the respective energy-intensive operations. By leveraging knowledge distillation and model quantization, Sorbet achieved a highly compressed binary weight model that maintains competitive performance while significantly reducing energy consumption. We validate Sorbet's effectiveness through extensive testing on the GLUE benchmark and a series of ablation studies, demonstrating its potential as an energy-efficient solution for language model inference.
By inducing privacy attacks on NLP models, attackers can obtain sensitive information such as training data and model parameters, etc. Although researchers have studied, in-depth, several kinds of attacks in NLP models, they are non-systematic analyses. It lacks a comprehensive understanding of the impact caused by the attacks. For example, we must consider which scenarios can apply to which attacks, what the common factors are that affect the performance of different attacks, the nature of the relationships between different attacks, and the influence of various datasets and models on the effectiveness of the attacks, etc. Therefore, we need a benchmark to holistically assess the privacy risks faced by NLP models. In this paper, we present a privacy attack and defense evaluation benchmark in the field of NLP, which includes the conventional/small models and large language models (LLMs). This benchmark supports a variety of models, datasets, and protocols, along with standardized modules for comprehensive evaluation of attacks and defense strategies. Based on the above framework, we present a study on the association between auxiliary data from different domains and the strength of privacy attacks. And we provide an improved attack method in this scenario with the help of Knowledge Distillation (KD). Furthermore, we propose a chained framework for privacy attacks. Allowing a practitioner to chain multiple attacks to achieve a higher-level attack objective. Based on this, we provide some defense and enhanced attack strategies. The code for reproducing the results can be found at https://github.com/user2311717757/nlp_doctor.
Animal sounds can be recognised automatically by machine learning, and this has an important role to play in biodiversity monitoring. Yet despite increasingly impressive capabilities, bioacoustic species classifiers still exhibit imbalanced performance across species and habitats, especially in complex soundscapes. In this study, we explore the effectiveness of transfer learning in large-scale bird sound classification across various conditions, including single- and multi-label scenarios, and across different model architectures such as CNNs and Transformers. Our experiments demonstrate that both fine-tuning and knowledge distillation yield strong performance, with cross-distillation proving particularly effective in improving in-domain performance on Xeno-canto data. However, when generalizing to soundscapes, shallow fine-tuning exhibits superior performance compared to knowledge distillation, highlighting its robustness and constrained nature. Our study further investigates how to use multi-species labels, in cases where these are present but incomplete. We advocate for more comprehensive labeling practices within the animal sound community, including annotating background species and providing temporal details, to enhance the training of robust bird sound classifiers. These findings provide insights into the optimal reuse of pretrained models for advancing automatic bioacoustic recognition.
We introduce the AIM 2024 UHD-IQA Challenge, a competition to advance the No-Reference Image Quality Assessment (NR-IQA) task for modern, high-resolution photos. The challenge is based on the recently released UHD-IQA Benchmark Database, which comprises 6,073 UHD-1 (4K) images annotated with perceptual quality ratings from expert raters. Unlike previous NR-IQA datasets, UHD-IQA focuses on highly aesthetic photos of superior technical quality, reflecting the ever-increasing standards of digital photography. This challenge aims to develop efficient and effective NR-IQA models. Participants are tasked with creating novel architectures and training strategies to achieve high predictive performance on UHD-1 images within a computational budget of 50G MACs. This enables model deployment on edge devices and scalable processing of extensive image collections. Winners are determined based on a combination of performance metrics, including correlation measures (SRCC, PLCC, KRCC), absolute error metrics (MAE, RMSE), and computational efficiency (G MACs). To excel in this challenge, participants leverage techniques like knowledge distillation, low-precision inference, and multi-scale training. By pushing the boundaries of NR-IQA for high-resolution photos, the UHD-IQA Challenge aims to stimulate the development of practical models that can keep pace with the rapidly evolving landscape of digital photography. The innovative solutions emerging from this competition will have implications for various applications, from photo curation and enhancement to image compression.
The purpose of face super-resolution (FSR) is to reconstruct high-resolution (HR) face images from low-resolution (LR) inputs. With the continuous advancement of deep learning technologies, contemporary prior-guided FSR methods initially estimate facial priors and then use this information to assist in the super-resolution reconstruction process. However, ensuring the accuracy of prior estimation remains challenging, and straightforward cascading and convolutional operations often fail to fully leverage prior knowledge. Inaccurate or insufficiently utilized prior information inevitably degrades FSR performance. To address this issue, we propose a prior knowledge distillation network (PKDN) for FSR, which involves transferring prior information from the teacher network to the student network. This approach enables the network to learn priors during the training stage while relying solely on low-resolution facial images during the testing stage, thus mitigating the adverse effects of prior estimation inaccuracies. Additionally, we incorporate robust attention mechanisms to design a parsing map fusion block that effectively utilizes prior information. To prevent feature loss, we retain multi-scale features during the feature extraction stage and employ them in the subsequent super-resolution reconstruction process. Experimental results on benchmark datasets demonstrate that our PKDN approach surpasses existing FSR methods in generating high-quality face images.
Given the significant potential of large language models (LLMs) in sequence modeling, emerging studies have begun applying them to time-series forecasting. Despite notable progress, existing methods still face two critical challenges: 1) their reliance on large amounts of paired text data, limiting the model applicability, and 2) a substantial modality gap between text and time series, leading to insufficient alignment and suboptimal performance. In this paper, we introduce \textbf{H}ierarchical \textbf{T}ext-\textbf{F}ree \textbf{A}lignment (\textbf{TS-HTFA}), a novel method that leverages hierarchical alignment to fully exploit the representation capacity of LLMs while eliminating the dependence on text data. Specifically, we replace paired text data with adaptive virtual text based on QR decomposition word embeddings and learnable prompt. Furthermore, we establish comprehensive cross-modal alignment at three levels: input, feature, and output. Extensive experiments on multiple time-series benchmarks demonstrate that HTFA achieves state-of-the-art performance, significantly improving prediction accuracy and generalization.
Diffusion models have shown excellent performance on various image generation tasks, but the substantial computational costs and huge memory footprint hinder their low-latency applications in real-world scenarios. Quantization is a promising way to compress and accelerate models. Nevertheless, due to the wide range and time-varying activations in diffusion models, existing methods cannot maintain both accuracy and efficiency simultaneously for low-bit quantization. To tackle this issue, we propose DilateQuant, a novel quantization framework for diffusion models that offers comparable accuracy and high efficiency. Specifically, we keenly aware of numerous unsaturated in-channel weights, which can be cleverly exploited to reduce the range of activations without additional computation cost. Based on this insight, we propose Weight Dilation (WD) that maximally dilates the unsaturated in-channel weights to a constrained range through a mathematically equivalent scaling. WD costlessly absorbs the activation quantization errors into weight quantization. The range of activations decreases, which makes activations quantization easy. The range of weights remains constant, which makes model easy to converge in training stage. Considering the temporal network leads to time-varying activations, we design a Temporal Parallel Quantizer (TPQ), which sets time-step quantization parameters and supports parallel quantization for different time steps, significantly improving the performance and reducing time cost. To further enhance performance while preserving efficiency, we introduce a Block-wise Knowledge Distillation (BKD) to align the quantized models with the full-precision models at a block level. The simultaneous training of time-step quantization parameters and weights minimizes the time required, and the shorter backpropagation paths decreases the memory footprint of the quantization process.
Sequential recommendation models user interests based on historical behaviors to provide personalized recommendation. Previous sequential recommendation algorithms primarily employ neural networks to extract features of user interests, achieving good performance. However, due to the recommendation system datasets sparsity, these algorithms often employ small-scale network frameworks, resulting in weaker generalization capability. Recently, a series of sequential recommendation algorithms based on large pre-trained language models have been proposed. Nonetheless, given the real-time demands of recommendation systems, the challenge remains in applying pre-trained language models for rapid recommendations in real scenarios. To address this, we propose a sequential recommendation algorithm based on a pre-trained language model and knowledge distillation. The key of proposed algorithm is to transfer pre-trained knowledge across domains and achieve lightweight inference by knowledge distillation. The algorithm operates in two stages: in the first stage, we fine-tune the pre-trained language model on the recommendation dataset to transfer the pre-trained knowledge to the recommendation task; in the second stage, we distill the trained language model to transfer the learned knowledge to a lightweight model. Extensive experiments on multiple public recommendation datasets show that the proposed algorithm enhances recommendation accuracy and provide timely recommendation services.
As per recent studies, Self-supervised learning (SSL) does not readily extend to smaller architectures. One direction to mitigate this shortcoming while simultaneously training a smaller network without labels is to adopt unsupervised knowledge distillation (UKD). Existing UKD approaches handcraft preservation worthy inter/intra sample relationships between the teacher and its student. However, this may overlook/ignore other key relationships present in the mapping of a teacher. In this paper, instead of heuristically constructing preservation worthy relationships between samples, we directly motivate the student to model the teacher's embedding manifold. If the mapped manifold is similar, all inter/intra sample relationships are indirectly conserved. We first demonstrate that prior methods cannot preserve teacher's latent manifold due to their sole reliance on $L_2$ normalised embedding features. Subsequently, we propose a simple objective to capture the lost information due to normalisation. Our proposed loss component, termed \textbf{space similarity}, motivates each dimension of a student's feature space to be similar to the corresponding dimension of its teacher. We perform extensive experiments demonstrating strong performance of our proposed approach on various benchmarks.
The use of pre-trained language models fine-tuned to address specific downstream tasks is a common approach in natural language processing (NLP). However, acquiring domain-specific knowledge via fine-tuning is challenging. Traditional methods involve pretraining language models using vast amounts of domain-specific data before fine-tuning for particular tasks. This study investigates emergency/non-emergency classification tasks based on electronic medical record (EMR) data obtained from pediatric emergency departments (PEDs) in Korea. Our findings reveal that existing domain-specific pre-trained language models underperform compared to general language models in handling N-lingual free-text data characteristics of non-English-speaking regions. To address these limitations, we propose a domain knowledge transfer methodology that leverages knowledge distillation to infuse general language models with domain-specific knowledge via fine-tuning. This study demonstrates the effective transfer of specialized knowledge between models by defining a general language model as the student model and a domain-specific pre-trained model as the teacher model. In particular, we address the complexities of EMR data obtained from PEDs in non-English-speaking regions, such as Korea, and demonstrate that the proposed method enhances classification performance in such contexts. The proposed methodology not only outperforms baseline models on Korean PED EMR data, but also promises broader applicability in various professional and technical domains. In future works, we intend to extend this methodology to include diverse non-English-speaking regions and address additional downstream tasks, with the aim of developing advanced model architectures using state-of-the-art KD techniques. The code is available in https://github.com/JoSangYeon/DSG-KD.
Large Language Models (LLMs), with their increasing depth and number of parameters, have demonstrated outstanding performance across a variety of natural language processing tasks. However, this growth in scale leads to increased computational demands, particularly during inference and fine-tuning. To address these challenges, we introduce EchoAtt, a novel framework aimed at optimizing transformer-based models by analyzing and leveraging the similarity of attention patterns across layers. Our analysis reveals that many inner layers in LLMs, especially larger ones, exhibit highly similar attention matrices. By exploiting this similarity, EchoAtt enables the sharing of attention matrices in less critical layers, significantly reducing computational requirements without compromising performance. We incorporate this approach within a knowledge distillation setup, where a pre-trained teacher model guides the training of a smaller student model. The student model selectively shares attention matrices in layers with high similarity while inheriting key parameters from the teacher. Our best results with TinyLLaMA-1.1B demonstrate that EchoAtt improves inference speed by 15\%, training speed by 25\%, and reduces the number of parameters by approximately 4\%, all while improving zero-shot performance. These findings highlight the potential of attention matrix sharing to enhance the efficiency of LLMs, making them more practical for real-time and resource-limited applications.
The rise of large transformer models has revolutionized Natural Language Processing, leading to significant advances in tasks like text classification. However, this progress demands substantial computational resources, escalating training duration, and expenses with larger model sizes. Efforts have been made to downsize and accelerate English models (e.g., Distilbert, MobileBert). Yet, research in this area is scarce for low-resource languages. In this study, we explore the case of the low-resource Indic language Marathi. Leveraging the marathi-topic-all-doc-v2 model as our baseline, we implement optimization techniques to reduce computation time and memory usage. Our focus is on enhancing the efficiency of Marathi transformer models while maintaining top-tier accuracy and reducing computational demands. Using the MahaNews document classification dataset and the marathi-topic-all-doc-v2 model from L3Cube, we apply Block Movement Pruning, Knowledge Distillation, and Mixed Precision methods individually and in combination to boost efficiency. We demonstrate the importance of strategic pruning levels in achieving desired efficiency gains. Furthermore, we analyze the balance between efficiency improvements and environmental impact, highlighting how optimized model architectures can contribute to a more sustainable computational ecosystem. Implementing these techniques on a single GPU system, we determine that the optimal configuration is 25\% pruning + knowledge distillation. This approach yielded a 2.56x speedup in computation time while maintaining baseline accuracy levels.