Recent end-to-end approaches have shown promise in extending large language models (LLMs) to speech inputs, but face limitations in directly assessing and optimizing alignment quality and fail to achieve fine-grained alignment due to speech-text length mismatch. We introduce BLSP-KD, a novel approach for Bootstrapping Language-Speech Pretraining via Knowledge Distillation, which addresses these limitations through two key techniques. First, it optimizes speech-text alignment by minimizing the divergence between the LLM's next-token prediction distributions for speech and text inputs using knowledge distillation. Second, it employs a continuous-integrate-andfire strategy to segment speech into tokens that correspond one-to-one with text tokens, enabling fine-grained alignment. We also introduce Partial LoRA (PLoRA), a new adaptation method supporting LLM finetuning for speech inputs under knowledge distillation. Quantitative evaluation shows that BLSP-KD outperforms previous end-to-end baselines and cascaded systems with comparable scale of parameters, facilitating general instruction-following capabilities for LLMs with speech inputs. This approach provides new possibilities for extending LLMs to spoken language interactions.
Knowledge distillation is commonly employed to compress neural networks, reducing the inference costs and memory footprint. In the scenario of homogenous architecture, feature-based methods have been widely validated for their effectiveness. However, in scenarios where the teacher and student models are of heterogeneous architectures, the inherent differences in feature representation significantly degrade the performance of these methods. Recent studies have highlighted that low-frequency components constitute the majority of image features. Motivated by this, we propose a Low-Frequency Components-based Contrastive Knowledge Distillation (LFCC) framework that significantly enhances the performance of feature-based distillation between heterogeneous architectures. Specifically, we designe a set of multi-scale low-pass filters to extract the low-frequency components of intermediate features from both the teacher and student models, aligning them in a compact space to overcome architectural disparities. Moreover, leveraging the intrinsic pairing characteristic of the teacher-student framework, we design an innovative sample-level contrastive learning framework that adeptly restructures the constraints of within-sample feature similarity and between-sample feature divergence into a contrastive learning task. This strategy enables the student model to capitalize on intra-sample feature congruence while simultaneously enhancing the discrimination of features among disparate samples. Consequently, our LFCC framework accurately captures the commonalities in feature representation across heterogeneous architectures. Extensive evaluations and empirical analyses across three architectures (CNNs, Transformers, and MLPs) demonstrate that LFCC achieves superior performance on the challenging benchmarks of ImageNet-1K and CIFAR-100. All codes will be publicly available.
Unsupervised Continual Learning (UCL) is a burgeoning field in machine learning, focusing on enabling neural networks to sequentially learn tasks without explicit label information. Catastrophic Forgetting (CF), where models forget previously learned tasks upon learning new ones, poses a significant challenge in continual learning, especially in UCL, where labeled information of data is not accessible. CF mitigation strategies, such as knowledge distillation and replay buffers, often face memory inefficiency and privacy issues. Although current research in UCL has endeavored to refine data representations and address CF in streaming data contexts, there is a noticeable lack of algorithms specifically designed for unsupervised clustering. To fill this gap, in this paper, we introduce the concept of Unsupervised Continual Clustering (UCC). We propose Forward-Backward Knowledge Distillation for unsupervised Continual Clustering (FBCC) to counteract CF within the context of UCC. FBCC employs a single continual learner (the ``teacher'') with a cluster projector, along with multiple student models, to address the CF issue. The proposed method consists of two phases: Forward Knowledge Distillation, where the teacher learns new clusters while retaining knowledge from previous tasks with guidance from specialized student models, and Backward Knowledge Distillation, where a student model mimics the teacher's behavior to retain task-specific knowledge, aiding the teacher in subsequent tasks. FBCC marks a pioneering approach to UCC, demonstrating enhanced performance and memory efficiency in clustering across various tasks, outperforming the application of clustering algorithms to the latent space of state-of-the-art UCL algorithms.
Knowledge distillation (KD) has emerged as a promising technique for addressing the computational challenges associated with deploying large-scale recommender systems. KD transfers the knowledge of a massive teacher system to a compact student model, to reduce the huge computational burdens for inference while retaining high accuracy. The existing KD studies primarily focus on one-time distillation in static environments, leaving a substantial gap in their applicability to real-world scenarios dealing with continuously incoming users, items, and their interactions. In this work, we delve into a systematic approach to operating the teacher-student KD in a non-stationary data stream. Our goal is to enable efficient deployment through a compact student, which preserves the high performance of the massive teacher, while effectively adapting to continuously incoming data. We propose Continual Collaborative Distillation (CCD) framework, where both the teacher and the student continually and collaboratively evolve along the data stream. CCD facilitates the student in effectively adapting to new data, while also enabling the teacher to fully leverage accumulated knowledge. We validate the effectiveness of CCD through extensive quantitative, ablative, and exploratory experiments on two real-world datasets. We expect this research direction to contribute to narrowing the gap between existing KD studies and practical applications, thereby enhancing the applicability of KD in real-world systems.
Sequential Recommendation (SR) task involves predicting the next item a user is likely to interact with, given their past interactions. The SR models examine the sequence of a user's actions to discern more complex behavioral patterns and temporal dynamics. Recent research demonstrates the great impact of LLMs on sequential recommendation systems, either viewing sequential recommendation as language modeling or serving as the backbone for user representation. Although these methods deliver outstanding performance, there is scant evidence of the necessity of a large language model and how large the language model is needed, especially in the sequential recommendation scene. Meanwhile, due to the huge size of LLMs, it is inefficient and impractical to apply a LLM-based model in real-world platforms that often need to process billions of traffic logs daily. In this paper, we explore the influence of LLMs' depth by conducting extensive experiments on large-scale industry datasets. Surprisingly, our motivational experiments reveal that most intermediate layers of LLMs are redundant, indicating that pruning the remaining layers can still maintain strong performance. Motivated by this insight, we empower small language models for SR, namely SLMRec, which adopt a simple yet effective knowledge distillation method. Moreover, SLMRec is orthogonal to other post-training efficiency techniques, such as quantization and pruning, so that they can be leveraged in combination. Comprehensive experimental results illustrate that the proposed SLMRec model attains the best performance using only 13% of the parameters found in LLM-based recommendation models while simultaneously achieving up to 6.6x and 8.0x speedups in training and inference time costs, respectively. Besides, we provide a theoretical justification for why small language models can perform comparably to large language models in SR.
Personalized learning is a proposed approach to address the problem of data heterogeneity in collaborative machine learning. In a decentralized setting, the two main challenges of personalization are client clustering and data privacy. In this paper, we address these challenges by developing P4 (Personalized Private Peer-to-Peer) a method that ensures that each client receives a personalized model while maintaining differential privacy guarantee of each client's local dataset during and after the training. Our approach includes the design of a lightweight algorithm to identify similar clients and group them in a private, peer-to-peer (P2P) manner. Once grouped, we develop differentially-private knowledge distillation for clients to co-train with minimal impact on accuracy. We evaluate our proposed method on three benchmark datasets (FEMNIST or Federated EMNIST, CIFAR-10 and CIFAR-100) and two different neural network architectures (Linear and CNN-based networks) across a range of privacy parameters. The results demonstrate the potential of P4, as it outperforms the state-of-the-art of differential private P2P by up to 40 percent in terms of accuracy. We also show the practicality of P4 by implementing it on resource constrained devices, and validating that it has minimal overhead, e.g., about 7 seconds to run collaborative training between two clients.
High-performance Transformer trackers have shown excellent results, yet they often bear a heavy computational load. Observing that a smaller input can immediately and conveniently reduce computations without changing the model, an easy solution is to adopt the low-resolution input for efficient Transformer tracking. Albeit faster, this hurts tracking accuracy much due to information loss in low resolution tracking. In this paper, we aim to mitigate such information loss to boost the performance of the low-resolution Transformer tracking via dual knowledge distillation from a frozen high-resolution (but not a larger) Transformer tracker. The core lies in two simple yet effective distillation modules, comprising query-key-value knowledge distillation (QKV-KD) and discrimination knowledge distillation (Disc-KD), across resolutions. The former, from the global view, allows the low-resolution tracker to inherit the features and interactions from the high-resolution tracker, while the later, from the target-aware view, enhances the target-background distinguishing capacity via imitating discriminative regions from its high-resolution counterpart. With the dual knowledge distillation, our Low-Resolution Transformer Tracker (LoReTrack) enjoys not only high efficiency owing to reduced computation but also enhanced accuracy by distilling knowledge from the high-resolution tracker. In extensive experiments, LoReTrack with a 256x256 resolution consistently improves baseline with the same resolution, and shows competitive or even better results compared to 384x384 high-resolution Transformer tracker, while running 52% faster and saving 56% MACs. Moreover, LoReTrack is resolution-scalable. With a 128x128 resolution, it runs 25 fps on a CPU with 64.9%/46.4% SUC scores on LaSOT/LaSOText, surpassing all other CPU real-time trackers. Code will be released.
This work addresses the challenge of achieving zero-shot adversarial robustness while preserving zero-shot generalization in large-scale foundation models, with a focus on the popular Contrastive Language-Image Pre-training (CLIP). Although foundation models were reported to have exceptional zero-shot generalization, they are highly vulnerable to adversarial perturbations. Existing methods achieve a comparable good tradeoff between zero-shot adversarial robustness and generalization under small adversarial perturbations. However, they fail to achieve a good tradeoff under large adversarial perturbations. To this end, we propose a novel Text-Image Mutual Awareness (TIMA) method that strikes a balance between zero-shot adversarial robustness and generalization. More precisely, we propose an Image-Aware Text (IAT) tuning mechanism that increases the inter-class distance of text embeddings by incorporating the Minimum Hyperspherical Energy (MHE). Simultaneously, fixed pre-trained image embeddings are used as cross-modal auxiliary supervision to maintain the similarity between the MHE-tuned and original text embeddings by the knowledge distillation, preserving semantic information between different classes. Besides, we introduce a Text-Aware Image (TAI) tuning mechanism, which increases inter-class distance between image embeddings during the training stage by Text-distance based Adaptive Margin (TAM). Similarly, a knowledge distillation is utilized to retain the similarity between fine-tuned and pre-trained image embeddings. Extensive experimental results demonstrate the effectiveness of our approach, showing impressive zero-shot performance against a wide range of adversarial perturbations while preserving the zero-shot generalization capabilities of the original CLIP model.
The adoption of large cloud-based models for inference in privacy-sensitive domains, such as homeless care systems and medical imaging, raises concerns about end-user data privacy. A common solution is adding locally differentially private (LDP) noise to queries before transmission, but this often reduces utility. LDPKiT, which stands for Local Differentially-Private and Utility-Preserving Inference via Knowledge Transfer, addresses the concern by generating a privacy-preserving inference dataset aligned with the private data distribution. This dataset is used to train a reliable local model for inference on sensitive inputs. LDPKiT employs a two-layer noise injection framework that leverages LDP and its post-processing property to create a privacy-protected inference dataset. The first layer ensures privacy, while the second layer helps to recover utility by creating a sufficiently large dataset for subsequent local model extraction using noisy labels returned from a cloud model on privacy-protected noisy inputs. Our experiments on Fashion-MNIST, SVHN and PathMNIST medical datasets demonstrate that LDPKiT effectively improves utility while preserving privacy. Moreover, the benefits of using LDPKiT increase at higher, more privacy-protective noise levels. For instance, on SVHN, LDPKiT achieves similar inference accuracy with $\epsilon=1.25$ as it does with $\epsilon=2.0$, providing stronger privacy guarantees with less than a 2% drop in accuracy. Furthermore, we perform extensive sensitivity analyses to evaluate the impact of dataset sizes on LDPKiT's effectiveness and systematically analyze the latent space representations to offer a theoretical explanation for its accuracy improvements. Lastly, we qualitatively and quantitatively demonstrate that the type of knowledge distillation performed by LDPKiT is ethical and fundamentally distinct from adversarial model extraction attacks.
Current methods for developing foundation models in medical image segmentation rely on two primary assumptions: a fixed set of classes and the immediate availability of a substantial and diverse training dataset. However, this can be impractical due to the evolving nature of imaging technology and patient demographics, as well as labor-intensive data curation, limiting their practical applicability and scalability. To address these challenges, we introduce a novel segmentation paradigm enabling the segmentation of a variable number of classes within a single classifier-free network, featuring an architecture independent of class number. This network is trained using contrastive learning and produces discriminative feature representations that facilitate straightforward interpretation. Additionally, we integrate this strategy into a knowledge distillation-based incremental learning framework, facilitating the gradual assimilation of new information from non-stationary data streams while avoiding catastrophic forgetting. Our approach provides a unified solution for tackling both class- and domain-incremental learning scenarios. We demonstrate the flexibility of our method in handling varying class numbers within a unified network and its capacity for incremental learning. Experimental results on an incompletely annotated, multi-modal, multi-source dataset for medical image segmentation underscore its superiority over state-of-the-art alternative approaches.
Most recent multispectral object detectors employ a two-branch structure to extract features from RGB and thermal images. While the two-branch structure achieves better performance than a single-branch structure, it overlooks inference efficiency. This conflict is increasingly aggressive, as recent works solely pursue higher performance rather than both performance and efficiency. In this paper, we address this issue by improving the performance of efficient single-branch structures. We revisit the reasons causing the performance gap between these structures. For the first time, we reveal the information interference problem in the naive early-fusion strategy adopted by previous single-branch structures. Besides, we find that the domain gap between multispectral images, and weak feature representation of the single-branch structure are also key obstacles for performance. Focusing on these three problems, we propose corresponding solutions, including a novel shape-priority early-fusion strategy, a weakly supervised learning method, and a core knowledge distillation technique. Experiments demonstrate that single-branch networks equipped with these three contributions achieve significant performance enhancements while retaining high efficiency. Our code will be available at \url{https://github.com/XueZ-phd/Efficient-RGB-T-Early-Fusion-Detection}.
In the field of medical image compression, Implicit Neural Representation (INR) networks have shown remarkable versatility due to their flexible compression ratios, yet they are constrained by a one-to-one fitting approach that results in lengthy encoding times. Our novel method, ``\textbf{UniCompress}'', innovatively extends the compression capabilities of INR by being the first to compress multiple medical data blocks using a single INR network. By employing wavelet transforms and quantization, we introduce a codebook containing frequency domain information as a prior input to the INR network. This enhances the representational power of INR and provides distinctive conditioning for different image blocks. Furthermore, our research introduces a new technique for the knowledge distillation of implicit representations, simplifying complex model knowledge into more manageable formats to improve compression ratios. Extensive testing on CT and electron microscopy (EM) datasets has demonstrated that UniCompress outperforms traditional INR methods and commercial compression solutions like HEVC, especially in complex and high compression scenarios. Notably, compared to existing INR techniques, UniCompress achieves a 4$\sim$5 times increase in compression speed, marking a significant advancement in the field of medical image compression. Codes will be publicly available.
Federated learning (FL) is a machine learning approach where nodes collaboratively train a global model. As more nodes participate in a round of FL, the effectiveness of individual model updates by nodes also diminishes. In this study, we increase the effectiveness of client updates by dividing the network into smaller partitions, or cohorts. We introduce Cohort-Parallel Federated Learning (CPFL): a novel learning approach where each cohort independently trains a global model using FL, until convergence, and the produced models by each cohort are then unified using knowledge distillation. The insight behind CPFL is that smaller, isolated networks converge quicker than in a one-network setting where all nodes participate. Through exhaustive experiments involving realistic traces and non-IID data distributions on the CIFAR-10 and FEMNIST image classification tasks, we investigate the balance between the number of cohorts, model accuracy, training time, and compute resources. Compared to traditional FL, CPFL with four cohorts, non-IID data distribution, and CIFAR-10 yields a 1.9x reduction in train time and a 1.3x reduction in resource usage, with a minimal drop in test accuracy.
Partial multi-task learning where training examples are annotated for one of the target tasks is a promising idea in remote sensing as it allows combining datasets annotated for different tasks and predicting more tasks with fewer network parameters. The na\"ive approach to partial multi-task learning is sub-optimal due to the lack of all-task annotations for learning joint representations. This paper proposes using knowledge distillation to replace the need of ground truths for the alternate task and enhance the performance of such approach. Experiments conducted on the public ISPRS 2D Semantic Labeling Contest dataset show the effectiveness of the proposed idea on partial multi-task learning for semantic tasks including object detection and semantic segmentation in aerial images.
Point cloud data labeling is considered a time-consuming and expensive task in autonomous driving, whereas annotation-free learning training can avoid it by learning point cloud representations from unannotated data. In this paper, we propose AFOV, a novel 3D \textbf{A}nnotation-\textbf{F}ree framework assisted by 2D \textbf{O}pen-\textbf{V}ocabulary segmentation models. It consists of two stages: In the first stage, we innovatively integrate high-quality textual and image features of 2D open-vocabulary models and propose the Tri-Modal contrastive Pre-training (TMP). In the second stage, spatial mapping between point clouds and images is utilized to generate pseudo-labels, enabling cross-modal knowledge distillation. Besides, we introduce the Approximate Flat Interaction (AFI) to address the noise during alignment and label confusion. To validate the superiority of AFOV, extensive experiments are conducted on multiple related datasets. We achieved a record-breaking 47.73\% mIoU on the annotation-free 3D segmentation task in nuScenes, surpassing the previous best model by 3.13\% mIoU. Meanwhile, the performance of fine-tuning with 1\% data on nuScenes and SemanticKITTI reached a remarkable 51.75\% mIoU and 48.14\% mIoU, outperforming all previous pre-trained models
The deployment of large language models (LLMs) faces considerable challenges concerning resource constraints and inference efficiency. Recent research has increasingly focused on smaller, task-specific models enhanced by distilling knowledge from LLMs. However, prior studies have often overlooked the diversity and quality of knowledge, especially the untapped potential of negative knowledge. Constructing effective negative knowledge remains severely understudied. In this paper, we introduce a novel framework called quality-guided contrastive rationale distillation aimed at enhancing reasoning capabilities through contrastive knowledge learning. For positive knowledge, we enrich its diversity through temperature sampling and employ self-consistency for further denoising and refinement. For negative knowledge, we propose an innovative self-adversarial approach that generates low-quality rationales by sampling previous iterations of smaller language models, embracing the idea that one can learn from one's own weaknesses. A contrastive loss is developed to distill both positive and negative knowledge into smaller language models, where an online-updating discriminator is integrated to assess qualities of rationales and assign them appropriate weights, optimizing the training process. Through extensive experiments across multiple reasoning tasks, we demonstrate that our method consistently outperforms existing distillation techniques, yielding higher-quality rationales.
Text-to-image diffusion models have demonstrated unprecedented capabilities for flexible and realistic image synthesis. Nevertheless, these models rely on a time-consuming sampling procedure, which has motivated attempts to reduce their latency. When improving efficiency, researchers often use the original diffusion model to train an additional network designed specifically for fast image generation. In contrast, our approach seeks to reduce latency directly, without any retraining, fine-tuning, or knowledge distillation. In particular, we find the repeated calculation of attention maps to be costly yet redundant, and instead suggest reusing them during sampling. Our specific reuse strategies are based on ODE theory, which implies that the later a map is reused, the smaller the distortion in the final image. We empirically compare these reuse strategies with few-step sampling procedures of comparable latency, finding that reuse generates images that are closer to those produced by the original high-latency diffusion model.
In computational digital pathology, accurate nuclear segmentation of Hematoxylin and Eosin (H&E) stained whole slide images (WSIs) is a critical step for many analyses and tissue characterizations. One popular deep learning-based nuclear segmentation approach, HoverNet, offers remarkably accurate results but lacks the high-throughput performance needed for clinical deployment in resource-constrained settings. Our approach, HoverFast, aims to provide fast and accurate nuclear segmentation in H&E images using knowledge distillation from HoverNet. By redesigning the tool with software engineering best practices, HoverFast introduces advanced parallel processing capabilities, efficient data handling, and optimized postprocessing. These improvements facilitate scalable high-throughput performance, making HoverFast more suitable for real-time analysis and application in resource-limited environments. Using a consumer grade Nvidia A5000 GPU, HoverFast showed a 21x speed improvement as compared to HoverNet; reducing mean analysis time for 40x WSIs from ~2 hours to 6 minutes while retaining a concordant mean Dice score of 0.91 against the original HoverNet output. Peak memory usage was also reduced 71% from 44.4GB, to 12.8GB, without requiring SSD-based caching. To ease adoption in research and clinical contexts, HoverFast aligns with best-practices in terms of (a) installation, and (b) containerization, while (c) providing outputs compatible with existing popular open-source image viewing tools such as QuPath. HoverFast has been made open-source and is available at andrewjanowczyk.com/open-source-tools/hoverfast.
This research addresses the challenges of diagnosing chest X-rays (CXRs) at low resolutions, a common limitation in resource-constrained healthcare settings. High-resolution CXR imaging is crucial for identifying small but critical anomalies, such as nodules or opacities. However, when images are downsized for processing in Computer-Aided Diagnosis (CAD) systems, vital spatial details and receptive fields are lost, hampering diagnosis accuracy. To address this, this paper presents the Multilevel Collaborative Attention Knowledge (MLCAK) method. This approach leverages the self-attention mechanism of Vision Transformers (ViT) to transfer critical diagnostic knowledge from high-resolution images to enhance the diagnostic efficacy of low-resolution CXRs. MLCAK incorporates local pathological findings to boost model explainability, enabling more accurate global predictions in a multi-task framework tailored for low-resolution CXR analysis. Our research, utilizing the Vindr CXR dataset, shows a considerable enhancement in the ability to diagnose diseases from low-resolution images (e.g. 28 x 28), suggesting a critical transition from the traditional reliance on high-resolution imaging (e.g. 224 x 224).
Knowledge Distillation (KD) could transfer the ``dark knowledge" of a well-performed yet large neural network to a weaker but lightweight one. From the view of output logits and softened probabilities, this paper goes deeper into the dark knowledge provided by teachers with different capacities. Two fundamental observations are: (1) a larger teacher tends to produce probability vectors that are less distinct between non-ground-truth classes; (2) teachers with different capacities are basically consistent in their cognition of relative class affinity. Abundant experimental studies verify these observations and in-depth empirical explanations are provided. The difference in dark knowledge leads to the peculiar phenomenon named ``capacity mismatch" that a more accurate teacher does not necessarily perform as well as a smaller teacher when teaching the same student network. Enlarging the distinctness between non-ground-truth class probabilities for larger teachers could address the capacity mismatch problem. This paper explores multiple simple yet effective ways to achieve this goal and verify their success by comparing them with popular KD methods that solve the capacity mismatch.