Knowledge distillation (KD) has shown great promise in transferring knowledge from larger teacher models to smaller student models. However, existing KD strategies for large language models often minimize output distributions between student and teacher models indiscriminately for each token. This overlooks the imbalanced nature of tokens and their varying transfer difficulties. In response, we propose a distillation strategy called Self-Evolution KD. The core of this approach involves dynamically integrating teacher distribution and one-hot distribution of ground truth into the student distribution as prior knowledge, which promotes the distillation process. It adjusts the ratio of prior knowledge based on token learning difficulty, fully leveraging the teacher model's potential. Experimental results show our method brings an average improvement of approximately 1.4 SacreBLEU points across four translation directions in the WMT22 test sets. Further analysis indicates that the improvement comes from better knowledge transfer from teachers, confirming our hypothesis.
This paper explores the potential of recurrent neural networks (RNNs) and other subquadratic architectures as competitive alternatives to transformer-based models in low-resource language modeling scenarios. We utilize HGRN2 (Qin et al., 2024), a recently proposed RNN-based architecture, and comparatively evaluate its effectiveness against transformer-based baselines and other subquadratic architectures (LSTM, xLSTM, Mamba). Our experimental results show that BABYHGRN, our HGRN2 language model, outperforms transformer-based models in both the 10M and 100M word tracks of the challenge, as measured by their performance on the BLiMP, EWoK, GLUE and BEAR benchmarks. Further, we show the positive impact of knowledge distillation. Our findings challenge the prevailing focus on transformer architectures and indicate the viability of RNN-based models, particularly in resource-constrained environments.
This work proposes a novel framework, Uncertainty-Guided Cross Attention Ensemble Mean Teacher (UG-CEMT), for achieving state-of-the-art performance in semi-supervised medical image segmentation. UG-CEMT leverages the strengths of co-training and knowledge distillation by combining a Cross-attention Ensemble Mean Teacher framework (CEMT) inspired by Vision Transformers (ViT) with uncertainty-guided consistency regularization and Sharpness-Aware Minimization emphasizing uncertainty. UG-CEMT improves semi-supervised performance while maintaining a consistent network architecture and task setting by fostering high disparity between sub-networks. Experiments demonstrate significant advantages over existing methods like Mean Teacher and Cross-pseudo Supervision in terms of disparity, domain generalization, and medical image segmentation performance. UG-CEMT achieves state-of-the-art results on multi-center prostate MRI and cardiac MRI datasets, where object segmentation is particularly challenging. Our results show that using only 10\% labeled data, UG-CEMT approaches the performance of fully supervised methods, demonstrating its effectiveness in exploiting unlabeled data for robust medical image segmentation. The code is publicly available at \url{https://github.com/Meghnak13/UG-CEMT}
Recent advances in diffusion generative models have yielded remarkable progress. While the quality of generated content continues to improve, these models have grown considerably in size and complexity. This increasing computational burden poses significant challenges, particularly in resource-constrained deployment scenarios such as mobile devices. The combination of model pruning and knowledge distillation has emerged as a promising solution to reduce computational demands while preserving generation quality. However, this technique inadvertently propagates undesirable behaviors, including the generation of copyrighted content and unsafe concepts, even when such instances are absent from the fine-tuning dataset. In this paper, we propose a novel bilevel optimization framework for pruned diffusion models that consolidates the fine-tuning and unlearning processes into a unified phase. Our approach maintains the principal advantages of distillation-namely, efficient convergence and style transfer capabilities-while selectively suppressing the generation of unwanted content. This plug-in framework is compatible with various pruning and concept unlearning methods, facilitating efficient, safe deployment of diffusion models in controlled environments.
Compositional understanding allows visual language models to interpret complex relationships between objects, attributes, and relations in images and text. However, most existing methods often rely on hard negative examples and fine-tuning, which can overestimate improvements and are limited by the difficulty of obtaining hard negatives. In this work, we introduce Zero-Shot Compositional Understanding (ZS-CU), a novel task that enhances compositional understanding without requiring hard negative training data. We propose YUKINO (Yielded Compositional Understanding Knowledge via Textual Inversion with NO), which uses textual inversion to map unlabeled images to pseudo-tokens in a pre-trained CLIP model. We propose introducing "no" logical regularization to address the issue of token interaction in inversion. Additionally, we suggest using knowledge distillation to reduce the time complexity of textual inversion. Experimental results show that YUKINO outperforms the existing multi-modal SOTA models by over 8% on the SugarCREPE benchmark, and also achieves significant improvements in image retrieval tasks.
Knowledge distillation (KD) has become a prevalent technique for compressing large language models (LLMs). Existing KD methods are constrained by the need for identical tokenizers (i.e., vocabularies) between teacher and student models, limiting their versatility in handling LLMs of different architecture families. In this paper, we introduce the Multi-Level Optimal Transport (MultiLevelOT), a novel approach that advances the optimal transport for universal cross-tokenizer knowledge distillation. Our method aligns the logit distributions of the teacher and the student at both token and sequence levels using diverse cost matrices, eliminating the need for dimensional or token-by-token correspondence. At the token level, MultiLevelOT integrates both global and local information by jointly optimizing all tokens within a sequence to enhance robustness. At the sequence level, we efficiently capture complex distribution structures of logits via the Sinkhorn distance, which approximates the Wasserstein distance for divergence measures. Extensive experiments on tasks such as extractive QA, generative QA, and summarization demonstrate that the MultiLevelOT outperforms state-of-the-art cross-tokenizer KD methods under various settings. Our approach is robust to different student and teacher models across model families, architectures, and parameter sizes. Codes and models are available at https://github.com/2018cx/Multi-Level-OT.
Educational data mining (EDM) is a part of applied computing that focuses on automatically analyzing data from learning contexts. Early prediction for identifying at-risk students is a crucial and widely researched topic in EDM research. It enables instructors to support at-risk students to stay on track, preventing student dropout or failure. Previous studies have predicted students' learning performance to identify at-risk students by using machine learning on data collected from e-learning platforms. However, most studies aimed to identify at-risk students utilizing the entire course data after the course finished. This does not correspond to the real-world scenario that at-risk students may drop out before the course ends. To address this problem, we introduce an RNN-Attention-KD (knowledge distillation) framework to predict at-risk students early throughout a course. It leverages the strengths of Recurrent Neural Networks (RNNs) in handling time-sequence data to predict students' performance at each time step and employs an attention mechanism to focus on relevant time steps for improved predictive accuracy. At the same time, KD is applied to compress the time steps to facilitate early prediction. In an empirical evaluation, RNN-Attention-KD outperforms traditional neural network models in terms of recall and F1-measure. For example, it obtained recall and F1-measure of 0.49 and 0.51 for Weeks 1--3 and 0.51 and 0.61 for Weeks 1--6 across all datasets from four years of a university course. Then, an ablation study investigated the contributions of different knowledge transfer methods (distillation objectives). We found that hint loss from the hidden layer of RNN and context vector loss from the attention module on RNN could enhance the model's prediction performance for identifying at-risk students. These results are relevant for EDM researchers employing deep learning models.
3D object detection is one of the fundamental perception tasks for autonomous vehicles. Fulfilling such a task with a 4D millimeter-wave radar is very attractive since the sensor is able to acquire 3D point clouds similar to Lidar while maintaining robust measurements under adverse weather. However, due to the high sparsity and noise associated with the radar point clouds, the performance of the existing methods is still much lower than expected. In this paper, we propose a novel Semi-supervised Cross-modality Knowledge Distillation (SCKD) method for 4D radar-based 3D object detection. It characterizes the capability of learning the feature from a Lidar-radar-fused teacher network with semi-supervised distillation. We first propose an adaptive fusion module in the teacher network to boost its performance. Then, two feature distillation modules are designed to facilitate the cross-modality knowledge transfer. Finally, a semi-supervised output distillation is proposed to increase the effectiveness and flexibility of the distillation framework. With the same network structure, our radar-only student trained by SCKD boosts the mAP by 10.38% over the baseline and outperforms the state-of-the-art works on the VoD dataset. The experiment on ZJUODset also shows 5.12% mAP improvements on the moderate difficulty level over the baseline when extra unlabeled data are available. Code is available at https://github.com/Ruoyu-Xu/SCKD.
The emergence of large-scale Mixture of Experts (MoE) models represents a significant advancement in artificial intelligence, offering enhanced model capacity and computational efficiency through conditional computation. However, deploying and running inference on these models presents significant challenges in computational resources, latency, and energy efficiency. This comprehensive survey analyzes optimization techniques for MoE models across the entire system stack. We first establish a taxonomical framework that categorizes optimization approaches into model-level, system-level, and hardware-level optimizations. At the model level, we examine architectural innovations including efficient expert design, attention mechanisms, various compression techniques such as pruning, quantization, and knowledge distillation, as well as algorithm improvement including dynamic routing strategies and expert merging methods. At the system level, we investigate distributed computing approaches, load balancing mechanisms, and efficient scheduling algorithms that enable scalable deployment. Furthermore, we delve into hardware-specific optimizations and co-design strategies that maximize throughput and energy efficiency. This survey provides both a structured overview of existing solutions and identifies key challenges and promising research directions in MoE inference optimization. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE inference optimization research, we have established a repository accessible at https://github.com/MoE-Inf/awesome-moe-inference/.
We applied few-shot in-context learning on the OPT-1.3B model for the natural language inference task and employed knowledge distillation to internalize the context information, reducing model parameter from 1.3B to 125M and achieving a size reduction from 2.5GB to 0.25GB. Compared to using in-context learning alone on similarly sized models, this context distillation approach achieved a nearly 50% improvement in out-of-domain accuracy, demonstrating superior knowledge transfer capabilities over prompt-based methods. Furthermore, this approach reduced memory consumption by up to 60% while delivering a 20% improvement in out-of-domain accuracy compared to conventional pattern-based fine-tuning.
The limited availability of labeled data has driven advancements in semi-supervised learning for medical image segmentation. Modern large-scale models tailored for general segmentation, such as the Segment Anything Model (SAM), have revealed robust generalization capabilities. However, applying these models directly to medical image segmentation still exposes performance degradation. In this paper, we propose a learnable prompting SAM-induced Knowledge distillation framework (KnowSAM) for semi-supervised medical image segmentation. Firstly, we propose a Multi-view Co-training (MC) strategy that employs two distinct sub-networks to employ a co-teaching paradigm, resulting in more robust outcomes. Secondly, we present a Learnable Prompt Strategy (LPS) to dynamically produce dense prompts and integrate an adapter to fine-tune SAM specifically for medical image segmentation tasks. Moreover, we propose SAM-induced Knowledge Distillation (SKD) to transfer useful knowledge from SAM to two sub-networks, enabling them to learn from SAM's predictions and alleviate the effects of incorrect pseudo-labels during training. Notably, the predictions generated by our subnets are used to produce mask prompts for SAM, facilitating effective inter-module information exchange. Extensive experimental results on various medical segmentation tasks demonstrate that our model outperforms the state-of-the-art semi-supervised segmentation approaches. Crucially, our SAM distillation framework can be seamlessly integrated into other semi-supervised segmentation methods to enhance performance. The code will be released upon acceptance of this manuscript at: https://github.com/taozh2017/KnowSAM
Data-free knowledge distillation aims to learn a compact student network from a pre-trained large teacher network without using the original training data of the teacher network. Existing collection-based and generation-based methods train student networks by collecting massive real examples and generating synthetic examples, respectively. However, they inevitably become weak in practical scenarios due to the difficulties in gathering or emulating sufficient real-world data. To solve this problem, we propose a novel method called \textbf{H}ybr\textbf{i}d \textbf{D}ata-\textbf{F}ree \textbf{D}istillation (HiDFD), which leverages only a small amount of collected data as well as generates sufficient examples for training student networks. Our HiDFD comprises two primary modules, \textit{i.e.}, the teacher-guided generation and student distillation. The teacher-guided generation module guides a Generative Adversarial Network (GAN) by the teacher network to produce high-quality synthetic examples from very few real-world collected examples. Specifically, we design a feature integration mechanism to prevent the GAN from overfitting and facilitate the reliable representation learning from the teacher network. Meanwhile, we drive a category frequency smoothing technique via the teacher network to balance the generative training of each category. In the student distillation module, we explore a data inflation strategy to properly utilize a blend of real and synthetic data to train the student network via a classifier-sharing-based feature alignment technique. Intensive experiments across multiple benchmarks demonstrate that our HiDFD can achieve state-of-the-art performance using 120 times less collected data than existing methods. Code is available at https://github.com/tangjialiang97/HiDFD.
Knowledge distillation (KD) remains challenging due to the opaque nature of the knowledge transfer process from a Teacher to a Student, making it difficult to address certain issues related to KD. To address this, we proposed UniCAM, a novel gradient-based visual explanation method, which effectively interprets the knowledge learned during KD. Our experimental results demonstrate that with the guidance of the Teacher's knowledge, the Student model becomes more efficient, learning more relevant features while discarding those that are not relevant. We refer to the features learned with the Teacher's guidance as distilled features and the features irrelevant to the task and ignored by the Student as residual features. Distilled features focus on key aspects of the input, such as textures and parts of objects. In contrast, residual features demonstrate more diffused attention, often targeting irrelevant areas, including the backgrounds of the target objects. In addition, we proposed two novel metrics: the feature similarity score (FSS) and the relevance score (RS), which quantify the relevance of the distilled knowledge. Experiments on the CIFAR10, ASIRRA, and Plant Disease datasets demonstrate that UniCAM and the two metrics offer valuable insights to explain the KD process.
Language models have proven successful across a wide range of software engineering tasks, but their significant computational costs often hinder their practical adoption. To address this challenge, researchers have begun applying various compression strategies to improve the efficiency of language models for code. These strategies aim to optimize inference latency and memory usage, though often at the cost of reduced model effectiveness. However, there is still a significant gap in understanding how these strategies influence the efficiency and effectiveness of language models for code. Here, we empirically investigate the impact of three well-known compression strategies -- knowledge distillation, quantization, and pruning -- across three different classes of software engineering tasks: vulnerability detection, code summarization, and code search. Our findings reveal that the impact of these strategies varies greatly depending on the task and the specific compression method employed. Practitioners and researchers can use these insights to make informed decisions when selecting the most appropriate compression strategy, balancing both efficiency and effectiveness based on their specific needs.
OpenAI o1 represents a significant milestone in Artificial Inteiligence, which achieves expert-level performances on many challanging tasks that require strong reasoning ability.OpenAI has claimed that the main techinique behinds o1 is the reinforcement learining. Recent works use alternative approaches like knowledge distillation to imitate o1's reasoning style, but their effectiveness is limited by the capability ceiling of the teacher model. Therefore, this paper analyzes the roadmap to achieving o1 from the perspective of reinforcement learning, focusing on four key components: policy initialization, reward design, search, and learning. Policy initialization enables models to develop human-like reasoning behaviors, equipping them with the ability to effectively explore solution spaces for complex problems. Reward design provides dense and effective signals via reward shaping or reward modeling, which is the guidance for both search and learning. Search plays a crucial role in generating high-quality solutions during both training and testing phases, which can produce better solutions with more computation. Learning utilizes the data generated by search for improving policy, which can achieve the better performance with more parameters and more searched data. Existing open-source projects that attempt to reproduce o1 can be seem as a part or a variant of our roadmap. Collectively, these components underscore how learning and search drive o1's advancement, making meaningful contributions to the development of LLM.
Multi-camera 3D object detection aims to detect and localize objects in 3D space using multiple cameras, which has attracted more attention due to its cost-effectiveness trade-off. However, these methods often struggle with the lack of accurate depth estimation caused by the natural weakness of the camera in ranging. Recently, multi-modal fusion and knowledge distillation methods for 3D object detection have been proposed to solve this problem, which are time-consuming during the training phase and not friendly to memory cost. In light of this, we propose PromptDet, a lightweight yet effective 3D object detection framework motivated by the success of prompt learning in 2D foundation model. Our proposed framework, PromptDet, comprises two integral components: a general camera-based detection module, exemplified by models like BEVDet and BEVDepth, and a LiDAR-assisted prompter. The LiDAR-assisted prompter leverages the LiDAR points as a complementary signal, enriched with a minimal set of additional trainable parameters. Notably, our framework is flexible due to our prompt-like design, which can not only be used as a lightweight multi-modal fusion method but also as a camera-only method for 3D object detection during the inference phase. Extensive experiments on nuScenes validate the effectiveness of the proposed PromptDet. As a multi-modal detector, PromptDet improves the mAP and NDS by at most 22.8\% and 21.1\% with fewer than 2\% extra parameters compared with the camera-only baseline. Without LiDAR points, PromptDet still achieves an improvement of at most 2.4\% mAP and 4.0\% NDS with almost no impact on camera detection inference time.
In this paper, we introduce Modality-Inconsistent Continual Learning (MICL), a new continual learning scenario for Multimodal Large Language Models (MLLMs) that involves tasks with inconsistent modalities (image, audio, or video) and varying task types (captioning or question-answering). Unlike existing vision-only or modality-incremental settings, MICL combines modality and task type shifts, both of which drive catastrophic forgetting. To address these challenges, we propose MoInCL, which employs a Pseudo Targets Generation Module to mitigate forgetting caused by task type shifts in previously seen modalities. It also incorporates Instruction-based Knowledge Distillation to preserve the model's ability to handle previously learned modalities when new ones are introduced. We benchmark MICL using a total of six tasks and conduct experiments to validate the effectiveness of our proposed MoInCL. The experimental results highlight the superiority of MoInCL, showing significant improvements over representative and state-of-the-art continual learning baselines.
The intrinsic dynamics and event-driven nature of spiking neural networks (SNNs) make them excel in processing temporal information by naturally utilizing embedded time sequences as time steps. Recent studies adopting this approach have demonstrated SNNs' effectiveness in speech command recognition, achieving high performance by employing large time steps for long time sequences. However, the large time steps lead to increased deployment burdens for edge computing applications. Thus, it is important to balance high performance and low energy consumption when detecting temporal patterns in edge devices. Our solution comprises two key components. 1). We propose a high-performance fully spike-driven framework termed SpikeSCR, characterized by a global-local hybrid structure for efficient representation learning, which exhibits long-term learning capabilities with extended time steps. 2). To further fully embrace low energy consumption, we propose an effective knowledge distillation method based on curriculum learning (KDCL), where valuable representations learned from the easy curriculum are progressively transferred to the hard curriculum with minor loss, striking a trade-off between power efficiency and high performance. We evaluate our method on three benchmark datasets: the Spiking Heidelberg Dataset (SHD), the Spiking Speech Commands (SSC), and the Google Speech Commands (GSC) V2. Our experimental results demonstrate that SpikeSCR outperforms current state-of-the-art (SOTA) methods across these three datasets with the same time steps. Furthermore, by executing KDCL, we reduce the number of time steps by 60% and decrease energy consumption by 54.8% while maintaining comparable performance to recent SOTA results. Therefore, this work offers valuable insights for tackling temporal processing challenges with long time sequences in edge neuromorphic computing systems.
We highlight two significant issues leading to the passivity of current merchant non-player characters (NPCs): pricing and communication. While immersive interactions have been a focus, negotiations between merchant NPCs and players on item prices have not received sufficient attention. First, we define passive pricing as the limited ability of merchants to modify predefined item prices. Second, passive communication means that merchants can only interact with players in a scripted manner. To tackle these issues and create an active merchant NPC, we propose a merchant framework based on large language models (LLMs), called MART, which consists of an appraiser module and a negotiator module. We conducted two experiments to guide game developers in selecting appropriate implementations by comparing different training methods and LLM sizes. Our findings indicate that finetuning methods, such as supervised finetuning (SFT) and knowledge distillation (KD), are effective in using smaller LLMs to implement active merchant NPCs. Additionally, we found three irregular cases arising from the responses of LLMs. We expect our findings to guide developers in using LLMs for developing active merchant NPCs.
In this paper, we propose $\textbf{Ne}$ural-$\textbf{Sy}$mbolic $\textbf{C}$ollaborative $\textbf{D}$istillation ($\textbf{NesyCD}$), a novel knowledge distillation method for learning the complex reasoning abilities of Large Language Models (LLMs, e.g., \textgreater 13B). We argue that complex reasoning tasks are difficult for Small Language Models (SLMs, e.g., $\leq$ 7B), as these tasks demand not only general cognitive abilities but also specialized knowledge, which is often sparse and difficult for these neural-based SLMs to effectively capture. Therefore, NesyCD distills the general capabilities and specialized knowledge in LLMs using different manners. On the one hand, we distill only general abilities from teacher LLMs into the student SLMs of parameterized neural networks. On the other hand, for the specialized abilities and uncommon knowledge of a complex reasoning task, we employ a symbolic knowledge distillation approach to obtain and store the specialized knowledge within a symbolic knowledge base (KB). By decoupling general and specialized capabilities, the proposed NesyCD can achieve superior performance cost-effectively, utilizing smaller models and blending parameterized neural networks with symbolic KB. Moreover, the specialized KB generalizes well and is comprehended and manipulated by humans. Our experiments show that NesyCD significantly boosts SLMs' complex reasoning performance on in-domain (BBH, GSM8K) and out-of-domain (AGIEval, ARC) datasets. Notably, our approach enabled the LLaMA3-8B and Qwen2-7B to surpass GPT-3.5-turbo in performance and come close to matching LLaMA3-70B, despite the latter having nine times more parameters. Our code will be available at https://github.com/Xnhyacinth/NesyCD.