Knowledge Distillation (KD), a learning manner with a larger teacher network guiding a smaller student network, transfers dark knowledge from the teacher to the student via logits or intermediate features, with the aim of producing a well-performed lightweight model. Notably, many subsequent feature-based KD methods outperformed the earliest logit-based KD method and iteratively generated numerous state-of-the-art distillation methods. Nevertheless, recent work has uncovered the potential of the logit-based method, bringing the simple KD form based on logits back into the limelight. Features or logits? They partially implement the KD with entirely distinct perspectives; therefore, choosing between logits and features is not straightforward. This paper provides a unified perspective of feature alignment in order to obtain a better comprehension of their fundamental distinction. Inheriting the design philosophy and insights of feature-based and logit-based methods, we introduce a block-wise logit distillation framework to apply implicit logit-based feature alignment by gradually replacing teacher's blocks as intermediate stepping-stone models to bridge the gap between the student and the teacher. Our method obtains comparable or superior results to state-of-the-art distillation methods. This paper demonstrates the great potential of combining logit and features, and we hope it will inspire future research to revisit KD from a higher vantage point.
To help MLOps engineers decide which operator to use in which deployment scenario, this study aims to empirically assess the accuracy vs latency trade-off of white-box (training-based) and black-box operators (non-training-based) and their combinations in an Edge AI setup. We perform inference experiments including 3 white-box (i.e., QAT, Pruning, Knowledge Distillation), 2 black-box (i.e., Partition, SPTQ), and their combined operators (i.e., Distilled SPTQ, SPTQ Partition) across 3 tiers (i.e., Mobile, Edge, Cloud) on 4 commonly-used Computer Vision and Natural Language Processing models to identify the effective strategies, considering the perspective of MLOps Engineers. Our Results indicate that the combination of Distillation and SPTQ operators (i.e., DSPTQ) should be preferred over non-hybrid operators when lower latency is required in the edge at small to medium accuracy drop. Among the non-hybrid operators, the Distilled operator is a better alternative in both mobile and edge tiers for lower latency performance at the cost of small to medium accuracy loss. Moreover, the operators involving distillation show lower latency in resource-constrained tiers (Mobile, Edge) compared to the operators involving Partitioning across Mobile and Edge tiers. For textual subject models, which have low input data size requirements, the Cloud tier is a better alternative for the deployment of operators than the Mobile, Edge, or Mobile-Edge tier (the latter being used for operators involving partitioning). In contrast, for image-based subject models, which have high input data size requirements, the Edge tier is a better alternative for operators than Mobile, Edge, or their combination.
Large Language Models (LLMs) demonstrate promising capabilities in solving simple scientific problems but often produce hallucinations for complex ones. While integrating LLMs with tools can increase reliability, this approach typically results in over-reliance on tools, diminishing the model's ability to solve simple problems through basic reasoning. In contrast, human experts first assess problem complexity using domain knowledge before choosing an appropriate solution approach. Inspired by this human problem-solving process, we propose a novel two-component fine-tuning method. In the first component World Knowledge Distillation (WKD), LLMs learn directly from solutions generated using tool's information to internalize domain knowledge. In the second component Tool Usage Adaptation (TUA), we partition problems into easy and hard categories based on the model's direct answering accuracy. While maintaining the same alignment target for easy problems as in WKD, we train the model to intelligently switch to tool usage for more challenging problems. We validate our method on six scientific benchmark datasets, spanning mathematics, climate science and epidemiology. On average, our models demonstrate a 28.18% improvement in answer accuracy and a 13.89% increase in tool usage precision across all datasets, surpassing state-of-the-art models including GPT-4o and Claude-3.5.
Multi-Object Tracking (MOT) aims to associate multiple objects across video frames and is a challenging vision task due to inherent complexities in the tracking environment. Most existing approaches train and track within a single domain, resulting in a lack of cross-domain generalizability to data from other domains. While several works have introduced natural language representation to bridge the domain gap in visual tracking, these textual descriptions often provide too high-level a view and fail to distinguish various instances within the same class. In this paper, we address this limitation by developing IP-MOT, an end-to-end transformer model for MOT that operates without concrete textual descriptions. Our approach is underpinned by two key innovations: Firstly, leveraging a pre-trained vision-language model, we obtain instance-level pseudo textual descriptions via prompt-tuning, which are invariant across different tracking scenes; Secondly, we introduce a query-balanced strategy, augmented by knowledge distillation, to further boost the generalization capabilities of our model. Extensive experiments conducted on three widely used MOT benchmarks, including MOT17, MOT20, and DanceTrack, demonstrate that our approach not only achieves competitive performance on same-domain data compared to state-of-the-art models but also significantly improves the performance of query-based trackers by large margins for cross-domain inputs.
The use of large language models (LLMs) as evaluators has garnered significant attention due to their potential to rival human-level evaluations in long-form response assessments. However, current LLM evaluators rely heavily on static, human-defined criteria, limiting their ability to generalize across diverse generative tasks and incorporate context-specific knowledge. In this paper, we propose a novel Self-Assessing LLM framework that integrates Context-Aware Criteria (SALC) with dynamic knowledge tailored to each evaluation instance. This instance-level knowledge enhances the LLM evaluator's performance by providing relevant and context-aware insights that pinpoint the important criteria specific to the current instance. Additionally, the proposed framework adapts seamlessly to various tasks without relying on predefined human criteria, offering a more flexible evaluation approach. Empirical evaluations demonstrate that our approach significantly outperforms existing baseline evaluation frameworks, yielding improvements on average 4.8% across a wide variety of datasets. Furthermore, by leveraging knowledge distillation techniques, we fine-tuned smaller language models for criteria generation and evaluation, achieving comparable or superior performance to larger models with much lower cost. Our method also exhibits a improvement in LC Win-Rate in AlpacaEval2 leaderboard up to a 12% when employed for preference data generation in Direct Preference Optimization (DPO), underscoring its efficacy as a robust and scalable evaluation framework.
This paper investigates the industrial setting of real-time classification of early media exchanged during the initialization phase of voice calls. We explore the application of state-of-the-art audio tagging models and highlight some limitations when applied to the classification of early media. While most existing approaches leverage convolutional neural networks, we propose a novel approach for low-resource requirements based on gradient-boosted trees. Our approach not only demonstrates a substantial improvement in runtime performance, but also exhibits a comparable accuracy. We show that leveraging knowledge distillation and class aggregation techniques to train a simpler and smaller model accelerates the classification of early media in voice calls. We provide a detailed analysis of the results on a proprietary and publicly available dataset, regarding accuracy and runtime performance. We additionally report a case study of the achieved performance improvements at a regional data center in India.
Radio-Frequency (RF)-based Human Activity Recognition (HAR) rises as a promising solution for applications unamenable to techniques requiring computer visions. However, the scarcity of labeled RF data due to their non-interpretable nature poses a significant obstacle. Thanks to the recent breakthrough of foundation models (FMs), extracting deep semantic insights from unlabeled visual data become viable, yet these vision-based FMs fall short when applied to small RF datasets. To bridge this gap, we introduce FM-Fi, an innovative cross-modal framework engineered to translate the knowledge of vision-based FMs for enhancing RF-based HAR systems. FM-Fi involves a novel cross-modal contrastive knowledge distillation mechanism, enabling an RF encoder to inherit the interpretative power of FMs for achieving zero-shot learning. It also employs the intrinsic capabilities of FM and RF to remove extraneous features for better alignment between the two modalities. The framework is further refined through metric-based few-shot learning techniques, aiming to boost the performance for predefined HAR tasks. Comprehensive evaluations evidently indicate that FM-Fi rivals the effectiveness of vision-based methodologies, and the evaluation results provide empirical validation of FM-Fi's generalizability across various environments.
This paper proposes a medical literature summary generation method based on the BERT model to address the challenges brought by the current explosion of medical information. By fine-tuning and optimizing the BERT model, we develop an efficient summary generation system that can quickly extract key information from medical literature and generate coherent, accurate summaries. In the experiment, we compared various models, including Seq-Seq, Attention, Transformer, and BERT, and demonstrated that the improved BERT model offers significant advantages in the Rouge and Recall metrics. Furthermore, the results of this study highlight the potential of knowledge distillation techniques to further enhance model performance. The system has demonstrated strong versatility and efficiency in practical applications, offering a reliable tool for the rapid screening and analysis of medical literature.
Large language models (LLMs) are expensive to deploy. Parameter sharing offers a possible path towards reducing their size and cost, but its effectiveness in modern LLMs remains fairly limited. In this work, we revisit "layer tying" as form of parameter sharing in Transformers, and introduce novel methods for converting existing LLMs into smaller "Recursive Transformers" that share parameters across layers, with minimal loss of performance. Here, our Recursive Transformers are efficiently initialized from standard pretrained Transformers, but only use a single block of unique layers that is then repeated multiple times in a loop. We further improve performance by introducing Relaxed Recursive Transformers that add flexibility to the layer tying constraint via depth-wise low-rank adaptation (LoRA) modules, yet still preserve the compactness of the overall model. We show that our recursive models (e.g., recursive Gemma 1B) outperform both similar-sized vanilla pretrained models (such as TinyLlama 1.1B and Pythia 1B) and knowledge distillation baselines -- and can even recover most of the performance of the original "full-size" model (e.g., Gemma 2B with no shared parameters). Finally, we propose Continuous Depth-wise Batching, a promising new inference paradigm enabled by the Recursive Transformer when paired with early exiting. In a theoretical analysis, we show that this has the potential to lead to significant (2-3x) gains in inference throughput.
Large language models (LLMs) have demonstrated remarkable performance across various downstream tasks. However, the high computational and memory requirements of LLMs are a major bottleneck. To address this, parameter-efficient fine-tuning (PEFT) methods such as low-rank adaptation (LoRA) have been proposed to reduce computational costs while ensuring minimal loss in performance. Additionally, knowledge distillation (KD) has been a popular choice for obtaining compact student models from teacher models. In this work, we present KD-LoRA, a novel fine-tuning method that combines LoRA with KD. Our results demonstrate that KD-LoRA achieves performance comparable to full fine-tuning (FFT) and LoRA while significantly reducing resource requirements. Specifically, KD-LoRA retains 98% of LoRA's performance on the GLUE benchmark, while being 40% more compact. Additionally, KD-LoRA reduces GPU memory usage by 30% compared to LoRA, while decreasing inference time by 30% compared to both FFT and LoRA. We evaluate KD-LoRA across three encoder-only models: BERT, RoBERTa, and DeBERTaV3. Code is available at https://github.com/rambodazimi/KD-LoRA.
Despite the success of Large Language Models (LLMs), they still face challenges related to high inference costs and memory requirements. To address these issues, Knowledge Distillation (KD) has emerged as a popular method for model compression, with student-generated outputs (SGOs) being particularly notable for reducing the mismatch between training and inference. However, SGOs often produce noisy and biased sequences, which can lead to misguidance from the teacher model, especially in long sequences. To mitigate these challenges, we propose SWITCH (Studying WIth TeaCHer for Knowledge Distillation), a novel approach that strategically incorporates the teacher model during the student's sequence generation. SWITCH identifies discrepancies between the token probabilities of the teacher and student models, allowing the teacher to intervene selectively, particularly in long sequences that are more prone to teacher misguidance. Extensive experimental results across three model families and five instruction-following datasets show that SWITCH surpasses traditional KD methods, particularly excelling in the generation of long sequential data.
Speech Emotion Captioning (SEC) has gradually become an active research task. The emotional content conveyed through human speech are often complex, and classifying them into fixed categories may not be enough to fully capture speech emotions. Describing speech emotions through natural language may be a more effective approach. However, existing SEC methods often produce hallucinations and lose generalization on unseen speech. To overcome these problems, we propose AlignCap, which Aligning Speech Emotion Captioning to Human Preferences based on large language model (LLM) with two properties: 1) Speech-Text Alignment, which minimizing the divergence between the LLM's response prediction distributions for speech and text inputs using knowledge distillation (KD) Regularization. 2) Human Preference Alignment, where we design Preference Optimization (PO) Regularization to eliminate factuality and faithfulness hallucinations. We also extract emotional clues as a prompt for enriching fine-grained information under KD-Regularization. Experiments demonstrate that AlignCap presents stronger performance to other state-of-the-art methods on Zero-shot SEC task.
In the rapidly evolving field of financial sentiment analysis, the efficiency and accuracy of predictive models are critical due to their significant impact on financial markets. Transformer based models like BERT and large language models (LLMs) like GPT-4, have advanced NLP tasks considerably. Despite their advantages, BERT-based models face challenges with computational intensity in edge computing environments, and the substantial size and compute requirements of LLMs limit their practical deployment. This study proposes leveraging the generative capabilities of LLMs, such as GPT-4 Omni, to create synthetic, domain-specific training data. This approach addresses the challenge of data scarcity and enhances the performance of smaller models by making them competitive with their larger counterparts. The research specifically aims to enhance FinBERT, a BERT model fine-tuned for financial sentiment analysis, and develop TinyFinBERT, a compact transformer model, through a structured, two-tiered knowledge distillation strategy. Using data augmented by GPT-4 Omni, which involves generating new training examples and transforming existing data, we significantly improved the accuracy of FinBERT, preparing it to serve as a teacher model. This enhanced FinBERT then distilled knowledge to TinyFinBERT, employing both GPT-4 Omni and GPT-3.5 Turbo augmented data. The distillation strategy incorporated both logit and intermediate layer distillation. The training and evaluation of TinyFinBERT utilized the PhraseBank dataset and the FiQA 2018 Task1 dataset, achieving performance comparable to FinBERT while being substantially smaller and more efficient. This research demonstrates how LLMs can effectively contribute to the advancement of financial sentiment analysis by enhancing the capabilities of smaller, more efficient models through innovative data augmentation and distillation techniques.
A growing number of machine learning scenarios rely on knowledge distillation where one uses the output of a surrogate model as labels to supervise the training of a target model. In this work, we provide a sharp characterization of this process for ridgeless, high-dimensional regression, under two settings: (i) model shift, where the surrogate model is arbitrary, and (ii) distribution shift, where the surrogate model is the solution of empirical risk minimization with out-of-distribution data. In both cases, we characterize the precise risk of the target model through non-asymptotic bounds in terms of sample size and data distribution under mild conditions. As a consequence, we identify the form of the optimal surrogate model, which reveals the benefits and limitations of discarding weak features in a data-dependent fashion. In the context of weak-to-strong (W2S) generalization, this has the interpretation that (i) W2S training, with the surrogate as the weak model, can provably outperform training with strong labels under the same data budget, but (ii) it is unable to improve the data scaling law. We validate our results on numerical experiments both on ridgeless regression and on neural network architectures.
Leading open-source large language models (LLMs) such as Llama-3.1-Instruct-405B are extremely capable at generating text, answering questions, and solving a variety of natural language understanding tasks. However, they incur higher inference cost and latency compared to smaller LLMs. Knowledge distillation provides a way to use outputs from these large, capable teacher models to train smaller student models which can be used for inference at lower cost and latency, while retaining comparable accuracy. We investigate the efficacy of distillation using the Llama-3.1-405B-Instruct teacher and the smaller Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct student models. Contributions of this work include (a) We evaluate the generalizability of distillation with the above Llama-3.1 teacher-student pairs across different tasks and datasets (b) We show that using synthetic data during distillation significantly improves the accuracy of 8B and 70B models, and when used with reasoning chains, even matches or surpasses the zero-shot accuracy of 405B model on some datasets (c) We empirically show that distillation enables 8B and 70B models to internalize 405B's reasoning ability by using only standard fine-tuning (without customizing any loss function). This allows cost and latency-efficient student model inference. (d) We show pitfalls in evaluation of distillation, and present task-specific evaluation, including both human and LLM-grading, and ground-truth based traditional accuracy benchmarks. This methodical study brings out the fundamental importance of synthetic data quality in knowledge distillation, and of combining multiple, task-specific ways of accuracy and quality evaluation in assessing the effectiveness of distillation.
Large Language Models (LLMs) can transfer their reasoning skills to smaller models by teaching them to generate the intermediate reasoning process required to solve multistep reasoning tasks. While LLMs can accurately solve reasoning tasks through a variety of strategies, even without fine-tuning, smaller models are not expressive enough to fit the LLMs distribution on all strategies when distilled and tend to prioritize one strategy over the others. This reliance on one strategy poses a challenge for smaller models when attempting to solve reasoning tasks that may be difficult with their preferred strategy. To address this, we propose a distillation method SIKeD (Self-guided Iterative Knowledge Distillation for mathematical reasoning), where the LLM teaches the smaller model to approach a task using different strategies and the smaller model uses its self-generated on-policy outputs to choose the most suitable strategy for the given task. The training continues in a self-guided iterative manner, where for each training iteration, a decision is made on how to combine the LLM data with the self-generated outputs. Unlike traditional distillation methods, SIKeD allows the smaller model to learn which strategy is suitable for a given task while continuously learning to solve a task using different strategies. Our experiments on various mathematical reasoning datasets show that SIKeD significantly outperforms traditional distillation techniques across smaller models of different sizes. Our code is available at: https://github.com/kumar-shridhar/SIKeD
Recent advancements in Text-to-Speech (TTS) technology have led to natural-sounding speech for English, primarily due to the availability of large-scale, high-quality web data. However, many other languages lack access to such resources, relying instead on limited studio-quality data. This scarcity results in synthesized speech that often suffers from intelligibility issues, particularly with low-frequency character bigrams. In this paper, we propose three solutions to address this challenge. First, we leverage high-quality data from linguistically or geographically related languages to improve TTS for the target language. Second, we utilize low-quality Automatic Speech Recognition (ASR) data recorded in non-studio environments, which is refined using denoising and speech enhancement models. Third, we apply knowledge distillation from large-scale models using synthetic data to generate more robust outputs. Our experiments with Hindi demonstrate significant reductions in intelligibility issues, as validated by human evaluators. We propose this methodology as a viable alternative for languages with limited access to high-quality data, enabling them to collectively benefit from shared resources.
Vertical Federated Learning (VFL) enables collaborative model training across different participants with distinct features and common samples, while preserving data privacy. Existing VFL methodologies often struggle with realistic data partitions, typically incurring high communication costs and significant operational complexity. In this work, we introduce a novel simplified approach to VFL, Active Participant-Centric VFL (APC-VFL), that, to the best of our knowledge, is the first to require only a single communication round between participants, and allows the active participant to do inference in a non collaborative fashion. This method integrates unsupervised representation learning with knowledge distillation to achieve comparable accuracy to traditional VFL methods based on vertical split learning in classical settings, reducing required communication rounds by up to $4200\times$, while being more flexible. Our approach also shows improvements compared to non-federated local models, as well as a comparable VFL proposal, VFedTrans, offering an efficient and flexible solution for collaborative learning.
Data-free knowledge distillation (DFKD) has emerged as a pivotal technique in the domain of model compression, substantially reducing the dependency on the original training data. Nonetheless, conventional DFKD methods that employ synthesized training data are prone to the limitations of inadequate diversity and discrepancies in distribution between the synthesized and original datasets. To address these challenges, this paper introduces an innovative approach to DFKD through diverse diffusion augmentation (DDA). Specifically, we revise the paradigm of common data synthesis in DFKD to a composite process through leveraging diffusion models subsequent to data synthesis for self-supervised augmentation, which generates a spectrum of data samples with similar distributions while retaining controlled variations. Furthermore, to mitigate excessive deviation in the embedding space, we introduce an image filtering technique grounded in cosine similarity to maintain fidelity during the knowledge distillation process. Comprehensive experiments conducted on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets showcase the superior performance of our method across various teacher-student network configurations, outperforming the contemporary state-of-the-art DFKD methods. Code will be available at:https://github.com/SLGSP/DDA.
Knowledge distillation (KD) is widely used to train small, high-performing student language models (LMs) using large teacher LMs. While effective in fine-tuning, KD during pre-training faces challenges in efficiency, flexibility, and effectiveness. Existing methods either incur high computational costs due to online teacher inference, require tokenization matching between teacher and student LMs, or risk losing the difficulty and diversity of the teacher-generated training data. To address these issues, we propose MiniPLM, a KD framework for pre-training LMs by refining the training data distribution with the teacher's knowledge. For efficiency, MiniPLM performs offline teacher LM inference, allowing KD for multiple student LMs without adding training-time costs. For flexibility, MiniPLM operates solely on the training corpus, enabling KD across model families. For effectiveness, MiniPLM leverages the differences between large and small LMs to enhance the difficulty and diversity of the training data, helping student LMs acquire versatile and sophisticated knowledge. Extensive experiments demonstrate that MiniPLM boosts the student LMs' performance on 9 widely used downstream tasks, improves the language modeling capabilities, and reduces pre-training computation. The benefit of MiniPLM extends to large pre-training scales, evidenced by the extrapolation of the scaling curves. Further analysis reveals that MiniPLM supports KD across model families and enhances the utilization of pre-training data. Our model, code, and data are available at https://github.com/thu-coai/MiniPLM.