Streamlined Knowledge Distillation is proposed, a simple yet effective logit-based method that transfers only two essential forms of knowledge without requiring additional alignment or relational modeling and introduces a Mahalanobis distance-based direction-wise loss stabilized through Tikhonov regularization and Cholesky decomposition.
Knowledge distillation (KD) has become a pivotal technique for transferring knowledge from large-scale teacher models to lightweight student models. However, traditional feature-based distillation methods necessitate the direct exposure of the teacher’s intermediate representations, raising concerns regarding data privacy and the leakage of proprietary model details. These concerns often hinder the deployment of distillation in collaborative or cloud-based scenarios. To address these challenges, we propose a privacy-aware teacher-oriented projected feature distillation framework that reduces direct feature exposure while preserving distillation effectiveness. Specifically, our method employs a low-rank projection strategy to obfuscate the teacher’s features into a compact subspace. Authorized student models leverage this projection matrix to align their own features, enabling effective knowledge transfer while reducing direct exposure of the teacher’s original spatial patterns. Extensive experiments on benchmark datasets demonstrate empirical feature obfuscation and reconstruction resistance while maintaining competitive performance in object detection and semantic segmentation, with gains in several settings even when learning from projected representations.
Jun-Fei Yi, Sihao Lin, Hui Zhang et al.· IEEE Transactions on Image P...· 0 citations
Logit-based knowledge distillation (KD) is pivotal for efficient model compression and cross-architecture learning. However, conventional methods typically rely on static, single-scale logit alignment, thereby overlooking the semantic evolution trajectory embedded in cross-scale prediction transitions. To bridge this gap, we propose Scale-Difference Evolution distillation (SDE), formulated in a structure-aware manner. Unlike the traditional prediction-imitation paradigm, SDE explicitly models the difference vectors between multi-scale logits to capture dynamic logical transitions from local features to global semantics. To optimize the distillation signal-to-noise ratio, SDE decouples scale differences into two complementary modules: Category-Dominant Difference (CDD), which isolates evidence fluctuations for core categories via a Top-K attention mask, and Relational Structure Difference (RSD), which preserves the consistency of inter-class topological evolution. Extensive experiments demonstrate that SDE consistently outperforms state-of-the-art methods across CIFAR-100, Tiny-ImageNet-200, CUB-200 and Stanford Cars. Notably, SDE achieves a significant 8.10% accuracy boost on the CUB-200 fine-grained benchmark, highlighting its superior capability in resolving inter-class ambiguities through cross-scale semantic modeling.
Hejie Lu· Proceedings of the 32nd ACM...· 0 citations
Token-level knowledge distillation (KD) matches two conditional distributions per position, yet the standard objectives compare them pointwise: a Kullback-Leibler gradient is blind to which wrong token receives probability mass. We develop a distributional view in which the teacher is represented not by a single softened output but by a family of multi-temperature views - marginals of the annealing path of its logits - and the student is trained against a geometry-aware aggregate of these views under an embedding-based ground cost. We formalize the resulting design space (mixtures, log-linear pooling, entropic Wasserstein barycenters, and a debiased Sinkhorn-divergence flagship in hub and path forms), prove an exact collapse result showing log-linear pooling of tempered views is equivalent to a single temperature, and give a multi-marginal Schrodinger-bridge reading that yields falsifiable predictions. On instruction-tuned Pythia pairs, experiments yield three empirical laws: (i) dispersion law - the benefit of multi-temperature aggregation grows monotonically with the effective temperature dispersion of the views, not with their number; (ii) dispersed views unlock the aggregation operator - the barycenter separates from the arithmetic mixture exactly when transport-based aggregation starts to beat averaging; and (iii) two-regime picture governed by the ceiling gap $\Gamma=\mathrm{PPL}_{\mathrm{SFT}}-\mathrm{PPL}_{T}$: when the fine-tuned teacher barely beats a supervised student the gentle transport objective is the best KD loss but no KD beats supervised fine-tuning, whereas at a real ceiling the ranking inverts - and the sign of the fidelity-generalization correlation flips. We argue that"which distillation loss is the best"is not a fixed property of the loss but a function of $\Gamma$.
This work revisits on-policy Reverse Kullback-Leibler distillation and decomposes its objective into a teacher-fitting term and a student-entropy term, without introducing an explicit FKL branch, and proposes Adaptive Entropy Distillation (AED), which uses the teacher's entropy to dynamically calibrate token-level imitation strength.
Shizheng Li, Zhiyu Shen, Yuyin Lu et al.· 0 citations
A practitioner's study of how to make distillation training efficient is presented, organised around two systems contributions, and a fused, chunked KL loss is introduced, making peak memory linear in the sequence length.
Bakbergen Ryskulov, Iker García-Ferrero, David Montero et al.· 0 citations
Knowledge distillation (KD) is a standard approach for compressing sequence-to-sequence models, but its per-sample effects are rarely examined. On the BanSum Bangla summarization benchmark, we find that standard KD improves ROUGE-L by only +0.0003 over a cross-entropy baseline, and that approximately 51.3% of training samples are estimated to actively harm student validation loss under standard KD. We propose two complementary reliability-aware distillation methods. CHAD (Counterfactual Harm-Aware Distillation) measures per-sample KD usefulness via gradient alignment with the validation loss direction and trains a lightweight gate that generalizes this counterfactual judgment to the full training set. EWAD+CPDP combines token-level entropy-weighted adaptive distillation with a capacity-proportional geometric constraint from a second, vocabulary-incompatible teacher. On BanSum, both methods substantially outperform standard KD: CHAD by +0.0173 ROUGE-L and EWAD+CPDP by +0.0219 ROUGE-L, where standard KD itself improves ROUGE-L by only +0.0003; despite using only 60M parameters, both outperform a fine-tuned Qwen 2.5-3B model (50x larger). We further evaluate the stronger method, EWAD+CPDP, across 15 typologically diverse XL-Sum languages organised into three sets, beating the CE-only baseline on 10/15 languages; gains are most reliable where the two teachers contribute complementary signal, and weakest where they have saturated or jointly weak target-language coverage. We release code and trained models to support reproducibility and further research on selective distillation.
Dipto Sumit, Ankan Kumar Roy Srizon, Sadia Khair Rodela et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
