Sparse Logit Sampling: Accelerating Knowledge Distillation in LLMs

• URL: http://arxiv.org/abs/2503.16870v1
• Date: Fri, 21 Mar 2025 05:58:18 GMT
• Title: Sparse Logit Sampling: Accelerating Knowledge Distillation in LLMs
• Authors: Anshumann, Mohd Abbas Zaidi, Akhil Kedia, Jinwoo Ahn, Taehwak Kwon, Kangwook Lee, Haejun Lee, Joohyung Lee,
• Abstract summary: We prove that naive approaches for sparse knowledge distillation such as caching Top-K probabilities, while intuitive, provide biased estimates of teacher probability distribution to the student.<n>We propose an importance-sampling-based method Random Sampling Knowledge Distillation', which provides unbiased estimates, preserves the gradient in expectation, and requires storing significantly sparser logits.
• Score: 12.73155638335145
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Knowledge distillation can be a cost-effective technique to distill knowledge in Large Language Models, if the teacher output logits can be pre-computed and cached. However, successfully applying this to pre-training remains largely unexplored. In this work, we prove that naive approaches for sparse knowledge distillation such as caching Top-K probabilities, while intuitive, provide biased estimates of teacher probability distribution to the student, resulting in suboptimal performance and calibration. We propose an importance-sampling-based method `Random Sampling Knowledge Distillation', which provides unbiased estimates, preserves the gradient in expectation, and requires storing significantly sparser logits. Our method enables faster training of student models with marginal overhead (<10%) compared to cross-entropy based training, while maintaining competitive performance compared to full distillation, across a range of model sizes from 300M to 3B.

Related papers

• Learning from Stochastic Teacher Representations Using Student-Guided Knowledge Distillation [64.15918654558816]
  Self-distillation (SSD) training strategy is introduced for filtering and weighting teacher representation to distill from task-relevant representations only. Experimental results on real-world affective computing, wearable/biosignal datasets from the UCR Archive, the HAR dataset, and image classification datasets show that the proposed SSD method can outperform state-of-the-art methods.
  arXiv  Detail & Related papers  (2025-04-19T14:08:56Z)
• Learning Effective Representations for Retrieval Using Self-Distillation with Adaptive Relevance Margins [29.88235846291593]
  biencoders estimate relevance of a document to a query by calculating the similarity of their respective embeddings. Current state-of-the-art biencoders are trained using an expensive training regime involving knowledge distillation from a teacher model and batch-sampling. We propose a novel parameter-free loss function for self-supervision that exploits the pre-trained language modeling capabilities of the encoder model as a training signal.
  arXiv  Detail & Related papers  (2024-07-31T10:33:32Z)
• Self-Knowledge Distillation for Learning Ambiguity [11.755814660833549]
  Recent language models often over-confidently predict a single label without consideration for its correctness. We propose a novel self-knowledge distillation method that enables models to learn label distributions more accurately. We validate our method on diverse NLU benchmark datasets and the experimental results demonstrate its effectiveness in producing better label distributions.
  arXiv  Detail & Related papers  (2024-06-14T05:11:32Z)
• BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping [64.54271680071373]
  Diffusion models have demonstrated excellent potential for generating diverse images. Knowledge distillation has been recently proposed as a remedy that can reduce the number of inference steps to one or a few. We present a novel technique called BOOT, that overcomes limitations with an efficient data-free distillation algorithm.
  arXiv  Detail & Related papers  (2023-06-08T20:30:55Z)
• HomoDistil: Homotopic Task-Agnostic Distillation of Pre-trained Transformers [49.79405257763856]
  This paper focuses on task-agnostic distillation. It produces a compact pre-trained model that can be easily fine-tuned on various tasks with small computational costs and memory footprints. We propose Homotopic Distillation (HomoDistil), a novel task-agnostic distillation approach equipped with iterative pruning.
  arXiv  Detail & Related papers  (2023-02-19T17:37:24Z)
• AI-KD: Adversarial learning and Implicit regularization for self-Knowledge Distillation [20.80341206396388]
  We present a novel adversarial penalized self-knowledge distillation method, named adversarial learning and implicit regularization for self-knowledge distillation (AI-KD) Our model not only distills the deterministic and progressive knowledge which are from the pre-trained and previous epoch predictive probabilities but also transfers the knowledge of the deterministic predictive distributions using adversarial learning. We demonstrate the effectiveness of the proposed method with network architectures on multiple datasets and show the proposed method achieves better performance than state-of-the-art methods.
  arXiv  Detail & Related papers  (2022-11-20T10:30:58Z)
• Parameter-Efficient and Student-Friendly Knowledge Distillation [83.56365548607863]
  We present a parameter-efficient and student-friendly knowledge distillation method, namely PESF-KD, to achieve efficient and sufficient knowledge transfer. Experiments on a variety of benchmarks show that PESF-KD can significantly reduce the training cost while obtaining competitive results compared to advanced online distillation methods.
  arXiv  Detail & Related papers  (2022-05-28T16:11:49Z)
• Robust and Resource-Efficient Data-Free Knowledge Distillation by Generative Pseudo Replay [5.3330804968579795]
  Data-Free Knowledge Distillation (KD) allows knowledge transfer from a trained neural network (teacher) to a more compact one (student) in the absence of original training data. Existing works use a validation set to monitor the accuracy of the student over real data and report the highest performance throughout the entire process. However, validation data may not be available at distillation time either, making it infeasible to record the student snapshot that achieved the peak accuracy. This is challenging because the student experiences knowledge degradation due to the distribution shift of the synthetic data. We propose to model the distribution of the previously observed synthetic samples
  arXiv  Detail & Related papers  (2022-01-09T14:14:28Z)
• Conditional Generative Data-Free Knowledge Distillation based on Attention Transfer [0.8594140167290099]
  We propose a conditional generative data-free knowledge distillation (CGDD) framework to train efficient portable network without any real data. In this framework, except using the knowledge extracted from teacher model, we introduce preset labels as additional auxiliary information. We show that trained portable network learned with proposed data-free distillation method obtains 99.63%, 99.07% and 99.84% relative accuracy on CIFAR10, CIFAR100 and Caltech101.
  arXiv  Detail & Related papers  (2021-12-31T09:23:40Z)
• Efficient training of lightweight neural networks using Online Self-Acquired Knowledge Distillation [51.66271681532262]
  Online Self-Acquired Knowledge Distillation (OSAKD) is proposed, aiming to improve the performance of any deep neural model in an online manner. We utilize k-nn non-parametric density estimation technique for estimating the unknown probability distributions of the data samples in the output feature space.
  arXiv  Detail & Related papers  (2021-08-26T14:01:04Z)
• Scalable Marginal Likelihood Estimation for Model Selection in Deep Learning [78.83598532168256]
  Marginal-likelihood based model-selection is rarely used in deep learning due to estimation difficulties. Our work shows that marginal likelihoods can improve generalization and be useful when validation data is unavailable.
  arXiv  Detail & Related papers  (2021-04-11T09:50:24Z)
• Deep Semi-supervised Knowledge Distillation for Overlapping Cervical Cell Instance Segmentation [54.49894381464853]
  We propose to leverage both labeled and unlabeled data for instance segmentation with improved accuracy by knowledge distillation. We propose a novel Mask-guided Mean Teacher framework with Perturbation-sensitive Sample Mining. Experiments show that the proposed method improves the performance significantly compared with the supervised method learned from labeled data only.
  arXiv  Detail & Related papers  (2020-07-21T13:27:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.