Traditional knowledge distillation (KD) relies on a proficient teacher trained on the target task, which is not always available. In this setting, cross-task distillation can be used, enabling the use of any teacher model trained on a different task. However, many KD methods prove ineffective when applied to this cross-task setting. To address this limitation, we propose a simple modification: the use of an inverted projection. We show that this drop-in replacement for a standard projector is effective by learning to disregard any task-specific features which might degrade the student's performance. We find that this simple modification is sufficient for extending many KD methods to the cross-task setting, where the teacher and student tasks can be very different. In doing so, we obtain up to a 1.9% improvement in the cross-task setting compared to the traditional projection, at no additional cost. Our method can obtain significant performance improvements (up to 7%) when using even a randomly-initialised teacher on various tasks such as depth estimation, image translation, and semantic segmentation, despite the lack of any learned knowledge to transfer. To provide conceptual and analytical insights into this result, we show that using an inverted projection allows the distillation loss to be decomposed into a knowledge transfer and a spectral regularisation component. Through this analysis we are additionally able to propose a novel regularisation loss that allows teacher-free distillation, enabling performance improvements of up to 8.57% on ImageNet with no additional training costs.
Pornographic content occurring in human-machine interaction dialogues can cause severe side effects for users in open-domain dialogue systems. However, research on detecting pornographic language within human-machine interaction dialogues is an important subject that is rarely studied. To advance in this direction, we introduce CensorChat, a dialogue monitoring dataset aimed at detecting whether the dialogue session contains pornographic content. To this end, we collect real-life human-machine interaction dialogues in the wild and break them down into single utterances and single-turn dialogues, with the last utterance spoken by the chatbot. We propose utilizing knowledge distillation of large language models to annotate the dataset. Specifically, first, the raw dataset is annotated by four open-source large language models, with the majority vote determining the label. Second, we use ChatGPT to update the empty label from the first step. Third, to ensure the quality of the validation and test sets, we utilize GPT-4 for label calibration. If the current label does not match the one generated by GPT-4, we employ a self-criticism strategy to verify its correctness. Finally, to facilitate the detection of pornographic text, we develop a series of text classifiers using a pseudo-labeled dataset. Detailed data analysis demonstrates that leveraging knowledge distillation techniques with large language models provides a practical and cost-efficient method for developing pornographic text detectors.
Logit knowledge distillation attracts increasing attention due to its practicality in recent studies. However, it often suffers inferior performance compared to the feature knowledge distillation. In this paper, we argue that existing logit-based methods may be sub-optimal since they only leverage the global logit output that couples multiple semantic knowledge. This may transfer ambiguous knowledge to the student and mislead its learning. To this end, we propose a simple but effective method, i.e., Scale Decoupled Distillation (SDD), for logit knowledge distillation. SDD decouples the global logit output into multiple local logit outputs and establishes distillation pipelines for them. This helps the student to mine and inherit fine-grained and unambiguous logit knowledge. Moreover, the decoupled knowledge can be further divided into consistent and complementary logit knowledge that transfers the semantic information and sample ambiguity, respectively. By increasing the weight of complementary parts, SDD can guide the student to focus more on ambiguous samples, improving its discrimination ability. Extensive experiments on several benchmark datasets demonstrate the effectiveness of SDD for wide teacher-student pairs, especially in the fine-grained classification task. Code is available at: https://github.com/shicaiwei123/SDD-CVPR2024
While various models and computational tools have been proposed for structure and property analysis of molecules, generating molecules that conform to all desired structures and properties remains a challenge. Here, we introduce a multi-constraint molecular generation large language model, TSMMG, which, akin to a student, incorporates knowledge from various small models and tools, namely, the 'teachers'. To train TSMMG, we construct a large set of text-molecule pairs by extracting molecular knowledge from these 'teachers', enabling it to generate novel molecules that conform to the descriptions through various text prompts. We experimentally show that TSMMG remarkably performs in generating molecules meeting complex, natural language-described property requirements across two-, three-, and four-constraint tasks, with an average molecular validity of over 99% and success ratio of 82.58%, 68.03%, and 67.48%, respectively. The model also exhibits adaptability through zero-shot testing, creating molecules that satisfy combinations of properties that have not been encountered. It can comprehend text inputs with various language styles, extending beyond the confines of outlined prompts, as confirmed through empirical validation. Additionally, the knowledge distillation feature of TSMMG contributes to the continuous enhancement of small models, while the innovative approach to dataset construction effectively addresses the issues of data scarcity and quality, which positions TSMMG as a promising tool in the domains of drug discovery and materials science.
Data augmentation improves the generalization power of deep learning models by synthesizing more training samples. Sample-mixing is a popular data augmentation approach that creates additional data by combining existing samples. Recent sample-mixing methods, like Mixup and Cutmix, adopt simple mixing operations to blend multiple inputs. Although such a heuristic approach shows certain performance gains in some computer vision tasks, it mixes the images blindly and does not adapt to different datasets automatically. A mixing strategy that is effective for a particular dataset does not often generalize well to other datasets. If not properly configured, the methods may create misleading mixed images, which jeopardize the effectiveness of sample-mixing augmentations. In this work, we propose an automated approach, TransformMix, to learn better transformation and mixing augmentation strategies from data. In particular, TransformMix applies learned transformations and mixing masks to create compelling mixed images that contain correct and important information for the target tasks. We demonstrate the effectiveness of TransformMix on multiple datasets in transfer learning, classification, object detection, and knowledge distillation settings. Experimental results show that our method achieves better performance as well as efficiency when compared with strong sample-mixing baselines.
Federated Learning (FL) has emerged as a prominent alternative to the traditional centralized learning approach. Generally speaking, FL is a decentralized approach that allows for collaborative training of Machine Learning (ML) models across multiple local nodes, ensuring data privacy and security while leveraging diverse datasets. Conventional FL, however, is susceptible to gradient inversion attacks, restrictively enforces a uniform architecture on local models, and suffers from model heterogeneity (model drift) due to non-IID local datasets. To mitigate some of these challenges, the new paradigm of Federated Knowledge Distillation (FKD) has emerged. FDK is developed based on the concept of Knowledge Distillation (KD), which involves extraction and transfer of a large and well-trained teacher model's knowledge to lightweight student models. FKD, however, still faces the model drift issue. Intuitively speaking, not all knowledge is universally beneficial due to the inherent diversity of data among local nodes. This calls for innovative mechanisms to evaluate the relevance and effectiveness of each client's knowledge for others, to prevent propagation of adverse knowledge. In this context, the paper proposes Effective Knowledge Fusion (KnFu) algorithm that evaluates knowledge of local models to only fuse semantic neighbors' effective knowledge for each client. The KnFu is a personalized effective knowledge fusion scheme for each client, that analyzes effectiveness of different local models' knowledge prior to the aggregation phase. Comprehensive experiments were performed on MNIST and CIFAR10 datasets illustrating effectiveness of the proposed KnFu in comparison to its state-of-the-art counterparts. A key conclusion of the work is that in scenarios with large and highly heterogeneous local datasets, local training could be preferable to knowledge fusion-based solutions.
Background: Image classification can be considered one of the key pillars of medical image analysis. Deep learning (DL) faces challenges that prevent its practical applications despite the remarkable improvement in medical image classification. The data distribution differences can lead to a drop in the efficiency of DL, known as the domain shift problem. Besides, requiring bulk annotated data for model training, the large size of models, and the privacy-preserving of patients are other challenges of using DL in medical image classification. This study presents a strategy that can address the mentioned issues simultaneously. Method: The proposed domain adaptive model based on knowledge distillation can classify images by receiving limited annotated data of different distributions. The designed multiple teachers-meticulous student model trains a student network that tries to solve the challenges by receiving the parameters of several teacher networks. The proposed model was evaluated using six available datasets of different distributions by defining the respiratory motion artefact detection task. Results: The results of extensive experiments using several datasets show the superiority of the proposed model in addressing the domain shift problem and lack of access to bulk annotated data. Besides, the privacy preservation of patients by receiving only the teacher network parameters instead of the original data and consolidating the knowledge of several DL models into a model with almost similar performance are other advantages of the proposed model. Conclusions: The proposed model can pave the way for practical clinical applications of deep classification methods by achieving the mentioned objectives simultaneously.
A typical assumption in state-of-the-art self-localization models is that an annotated training dataset is available in the target workspace. However, this does not always hold when a robot travels in a general open-world. This study introduces a novel training scheme for open-world distributed robot systems. In our scheme, a robot ("student") can ask the other robots it meets at unfamiliar places ("teachers") for guidance. Specifically, a pseudo-training dataset is reconstructed from the teacher model and thereafter used for continual learning of the student model. Unlike typical knowledge transfer schemes, our scheme introduces only minimal assumptions on the teacher model, such that it can handle various types of open-set teachers, including uncooperative, untrainable (e.g., image retrieval engines), and blackbox teachers (i.e., data privacy). Rather than relying on the availability of private data of teachers as in existing methods, we propose to exploit an assumption that holds universally in self-localization tasks: "The teacher model is a self-localization system" and to reuse the self-localization system of a teacher as a sole accessible communication channel. We particularly focus on designing an excellent student/questioner whose interactions with teachers can yield effective question-and-answer sequences that can be used as pseudo-training datasets for the student self-localization model. When applied to a generic recursive knowledge distillation scenario, our approach exhibited stable and consistent performance improvement.
Knowledge Distillation (KD) aims to transfer a more capable teacher model's knowledge to a lighter student model in order to improve the efficiency of the model, making it faster and more deployable. However, the student model's optimization process over the noisy pseudo labels (generated by the teacher model) is tricky and the amount of pseudo labels one can generate is limited due to Out of Memory (OOM) error. In this paper, we propose FlyKD (Knowledge Distillation on the Fly) which enables the generation of virtually unlimited number of pseudo labels, coupled with Curriculum Learning that greatly alleviates the optimization process over the noisy pseudo labels. Empirically, we observe that FlyKD outperforms vanilla KD and the renown Local Structure Preserving Graph Convolutional Network (LSPGCN). Lastly, with the success of Curriculum Learning, we shed light on a new research direction of improving optimization over noisy pseudo labels.
Test-Time Training (TTT) proposes to adapt a pre-trained network to changing data distributions on-the-fly. In this work, we propose the first TTT method for 3D semantic segmentation, TTT-KD, which models Knowledge Distillation (KD) from foundation models (e.g. DINOv2) as a self-supervised objective for adaptation to distribution shifts at test-time. Given access to paired image-pointcloud (2D-3D) data, we first optimize a 3D segmentation backbone for the main task of semantic segmentation using the pointclouds and the task of 2D $\to$ 3D KD by using an off-the-shelf 2D pre-trained foundation model. At test-time, our TTT-KD updates the 3D segmentation backbone for each test sample, by using the self-supervised task of knowledge distillation, before performing the final prediction. Extensive evaluations on multiple indoor and outdoor 3D segmentation benchmarks show the utility of TTT-KD, as it improves performance for both in-distribution (ID) and out-of-distribution (ODO) test datasets. We achieve a gain of up to 13% mIoU (7% on average) when the train and test distributions are similar and up to 45% (20% on average) when adapting to OOD test samples.
Artificial General Intelligence falls short when communicating role specific nuances to other systems. This is more pronounced when building autonomous LLM agents capable and designed to communicate with each other for real world problem solving. Humans can communicate context and domain specific nuances along with knowledge, and that has led to refinement of skills. In this work we propose and evaluate a novel method that leads to knowledge distillation among LLM agents leading to realtime human role play preserving unique contexts without relying on any stored data or pretraining. We also evaluate how our system performs better in simulated real world tasks compared to state of the art.
We present a hybrid-view-based knowledge distillation framework, termed HVDistill, to guide the feature learning of a point cloud neural network with a pre-trained image network in an unsupervised man- ner. By exploiting the geometric relationship between RGB cameras and LiDAR sensors, the correspondence between the two modalities based on both image- plane view and bird-eye view can be established, which facilitates representation learning. Specifically, the image-plane correspondences can be simply ob- tained by projecting the point clouds, while the bird- eye-view correspondences can be achieved by lifting pixels to the 3D space with the predicted depths un- der the supervision of projected point clouds. The image teacher networks provide rich semantics from the image-plane view and meanwhile acquire geometric information from the bird-eye view. Indeed, image features from the two views naturally comple- ment each other and together can ameliorate the learned feature representation of the point cloud stu- dent networks. Moreover, with a self-supervised pre- trained 2D network, HVDistill requires neither 2D nor 3D annotations. We pre-train our model on nuScenes dataset and transfer it to several downstream tasks on nuScenes, SemanticKITTI, and KITTI datasets for evaluation. Extensive experimental results show that our method achieves consistent improvements over the baseline trained from scratch and significantly out- performs the existing schemes. Codes are available at [email omitted]:zhangsha1024/HVDistill.git.
Quantization-aware training (QAT) and Knowledge Distillation (KD) are combined to achieve competitive performance in creating low-bit deep learning models. However, existing works applying KD to QAT require tedious hyper-parameter tuning to balance the weights of different loss terms, assume the availability of labeled training data, and require complex, computationally intensive training procedures for good performance. To address these limitations, this paper proposes a novel Self-Supervised Quantization-Aware Knowledge Distillation (SQAKD) framework. SQAKD first unifies the forward and backward dynamics of various quantization functions, making it flexible for incorporating various QAT works. Then it formulates QAT as a co-optimization problem that simultaneously minimizes the KL-Loss between the full-precision and low-bit models for KD and the discretization error for quantization, without supervision from labels. A comprehensive evaluation shows that SQAKD substantially outperforms the state-of-the-art QAT and KD works for a variety of model architectures. Our code is at: https://github.com/kaiqi123/SQAKD.git.
Since the advent of the Segment Anything Model(SAM) approximately one year ago, it has engendered significant academic interest and has spawned a large number of investigations and publications from various perspectives. However, the deployment of SAM in practical assembly line scenarios has yet to materialize due to its large image encoder, which weighs in at an imposing 632M. In this study, we have replaced the heavyweight image encoder with a lightweight one, thereby enabling the deployment of SAM in practical assembly line scenarios. Specifically, we have employed decoupled distillation to train the encoder of MobileSAM in a resource-limited setting. The entire knowledge distillation experiment can be completed in a single day on a single RTX 4090. The resulting lightweight SAM, referred to as Group-Mix SAM, had 37.63% (2.16M) fewer parameters and 42.5% (15614.7M) fewer floating-point operations compared to MobileSAM. However, on our constructed industrial dataset, MALSD, its mIoU was only marginally lower than that of MobileSAM, at 0.615. Finally, we conducted a comprehensive comparative experiment to demonstrate the superiority of Group-Mix SAM in the industrial domain. With its exceptional performance, our Group-Mix SAM is more suitable for practical assembly line applications.
Histo-genomic multi-modal methods have recently emerged as a powerful paradigm, demonstrating significant potential for improving cancer prognosis. However, genome sequencing, unlike histopathology imaging, is still not widely accessible in underdeveloped regions, limiting the application of these multi-modal approaches in clinical settings. To address this, we propose a novel Genome-informed Hyper-Attention Network, termed G-HANet, which is capable of effectively distilling the histo-genomic knowledge during training to elevate uni-modal whole slide image (WSI)-based inference for the first time. Compared with traditional knowledge distillation methods (i.e., teacher-student architecture) in other tasks, our end-to-end model is superior in terms of training efficiency and learning cross-modal interactions. Specifically, the network comprises the cross-modal associating branch (CAB) and hyper-attention survival branch (HSB). Through the genomic data reconstruction from WSIs, CAB effectively distills the associations between functional genotypes and morphological phenotypes and offers insights into the gene expression profiles in the feature space. Subsequently, HSB leverages the distilled histo-genomic associations as well as the generated morphology-based weights to achieve the hyper-attention modeling of the patients from both histopathology and genomic perspectives to improve cancer prognosis. Extensive experiments are conducted on five TCGA benchmarking datasets and the results demonstrate that G-HANet significantly outperforms the state-of-the-art WSI-based methods and achieves competitive performance with genome-based and multi-modal methods. G-HANet is expected to be explored as a useful tool by the research community to address the current bottleneck of insufficient histo-genomic data pairing in the context of cancer prognosis and precision oncology.
We present Recurrent Drafter (ReDrafter), an advanced speculative decoding approach that achieves state-of-the-art speedup for large language models (LLMs) inference. The performance gains are driven by three key aspects: (1) leveraging a recurrent neural network (RNN) as the draft model conditioning on LLM's hidden states, (2) applying a dynamic tree attention algorithm over beam search results to eliminate duplicated prefixes in candidate sequences, and (3) training through knowledge distillation from the LLM. ReDrafter accelerates Vicuna inference in MT-Bench by up to 2.8x with a PyTorch implementation on Nvidia H100 GPUs. To demonstrate its practicality in real environments, we also validated its effectiveness for on-device applications by implementing the approach in MLX and benchmarking performance on Metal GPUs in Apple Silicon chips, achieving up to 2.3x speedup.
Computational high-throughput studies, especially in research on high-entropy materials and catalysts, are hampered by high-dimensional composition spaces and myriad structural microstates. They present bottlenecks to the conventional use of density functional theory calculations, and consequently, the use of machine-learned potentials is becoming increasingly prevalent in atomic structure simulations. In this communication, we show the results of adjusting and fine-tuning the pretrained EquiformerV2 model from the Open Catalyst Project to infer adsorption energies of *OH and *O on the out-of-domain high-entropy alloy Ag-Ir-Pd-Pt-Ru. By applying an energy filter based on the local environment of the binding site the zero-shot inference is markedly improved and through few-shot fine-tuning the model yields state-of-the-art accuracy. It is also found that EquiformerV2, assuming the role of general machine learning potential, is able to inform a smaller, more focused direct inference model. This knowledge distillation setup boosts performance on complex binding sites. Collectively, this shows that foundational knowledge learned from ordered intermetallic structures, can be extrapolated to the highly disordered structures of solid-solutions. With the vastly accelerated computational throughput of these models, hitherto infeasible research in the high-entropy material space is now readily accessible.
Large Language Models (LLM) have demonstrated their strong ability in the field of machine translation (MT), yet they suffer from high computational cost and latency. Therefore, transferring translation knowledge from giant LLMs to medium-sized machine translation models is a promising research direction. However, traditional knowledge distillation methods do not take the capability of student and teacher models into consideration, therefore repeatedly teaching student models on the knowledge they have learned, and failing to extend to novel contexts and knowledge. In this paper, we propose a framework called MT-Patcher, which transfers knowledge from LLMs to existing MT models in a selective, comprehensive and proactive manner. Considering the current translation ability of student MT models, we only identify and correct their translation errors, instead of distilling the whole translation from the teacher. Leveraging the strong language abilities of LLMs, we instruct LLM teachers to synthesize diverse contexts and anticipate more potential errors for the student. Experiment results on translating both specific language phenomena and general MT benchmarks demonstrate that finetuning the student MT model on about 10% examples can achieve comparable results to the traditional knowledge distillation method, and synthesized potential errors and diverse contexts further improve translation performances on unseen contexts and words.
Large-scale vision-language models (VLMs) have shown a strong zero-shot generalization capability on unseen-domain data. However, adapting pre-trained VLMs to a sequence of downstream tasks often leads to the forgetting of previously learned knowledge and a reduction in zero-shot classification performance. To tackle this problem, we propose a unique Selective Dual-Teacher Knowledge Transfer framework that leverages the most recent fine-tuned and the original pre-trained VLMs as dual teachers to preserve the previously learned knowledge and zero-shot capabilities, respectively. With only access to an unlabeled reference dataset, our proposed framework performs a selective knowledge distillation mechanism by measuring the feature discrepancy from the dual-teacher VLMs. Consequently, our selective dual-teacher knowledge distillation mitigates catastrophic forgetting of previously learned knowledge while preserving the zero-shot capabilities of pre-trained VLMs. Extensive experiments on benchmark datasets demonstrate that our framework is favorable against state-of-the-art continual learning approaches for preventing catastrophic forgetting and zero-shot degradation. Project page: https://chuyu.org/research/snd
Reconstructing a sequence of sharp images from the blurry input is crucial for enhancing our insights into the captured scene and poses a significant challenge due to the limited temporal features embedded in the image. Spike cameras, sampling at rates up to 40,000 Hz, have proven effective in capturing motion features and beneficial for solving this ill-posed problem. Nonetheless, existing methods fall into the supervised learning paradigm, which suffers from notable performance degradation when applied to real-world scenarios that diverge from the synthetic training data domain. Moreover, the quality of reconstructed images is capped by the generated images based on motion analysis interpolation, which inherently differs from the actual scene, affecting the generalization ability of these methods in real high-speed scenarios. To address these challenges, we propose the first self-supervised framework for the task of spike-guided motion deblurring. Our approach begins with the formulation of a spike-guided deblurring model that explores the theoretical relationships among spike streams, blurry images, and their corresponding sharp sequences. We subsequently develop a self-supervised cascaded framework to alleviate the issues of spike noise and spatial-resolution mismatching encountered in the deblurring model. With knowledge distillation and re-blurring loss, we further design a lightweight deblur network to generate high-quality sequences with brightness and texture consistency with the original input. Quantitative and qualitative experiments conducted on our real-world and synthetic datasets with spikes validate the superior generalization of the proposed framework. Our code, data and trained models will be available at \url{https://github.com/chenkang455/S-SDM}.