Department of Electrical Engineering Indian Institute of T echnology Bombay 21D070002 aakash.gore@iitb.ac.in Abstract--Knowledge distillation has emerged as a powerful technique for model compression, enabling the transfer of knowledge from large teacher networks to compact student models. However, traditional knowledge distillation methods treat all teacher predictions equally, regardless of the teacher's confidence in those predictions. This paper proposes an uncertainty-aware dual-student knowledge distillation framework that leverages teacher prediction uncertainty to selectively guide student learning. We introduce a peer-learning mechanism where two heterogeneous student architectures, specifically ResNet-18 and MobileNetV2, learn collaboratively from both the teacher network and each other . Experimental results on ImageNet-100 demonstrate that our approach achieves superior performance compared to baseline knowledge distillation methods, with ResNet-18 achieving 83.84% top-1 accuracy and MobileNetV2 achieving 81.46% top-1 accuracy, representing improvements of 2.04% and 0.92% respectively over traditional single-student distillation approaches. Deep neural networks have achieved remarkable success across various computer vision tasks, but their deployment on resource-constrained devices remains challenging due to high computational and memory requirements. This technique has become increasingly important as the demand for deploying sophisticated machine learning models on edge devices, mobile platforms, and embedded systems continues to grow. Traditional knowledge distillation approaches use a weighted combination of hard labels derived from ground truth annotations and soft labels generated by teacher predictions to train student networks.

Neural Information Processing Systems http://nips.cc/

We first give a derivation on the equivalence of label smoothing regularization and Eq. 7. Evidently, the objective does not regularize confidence diversity. "Scale both" corresponds to the originally proposed distillation objective in which both teacher and Plots of test accuracy and ECE against amount of temperature scaling applied are shown in Figure 1. Firstly, we observe that models trained with student scaling have ECE almost identical to that of the teacher models. As a direct contrast, we see that the student models trained without student scaling perform much better in terms of calibration error in general over its teacher. This coupled effect could be the reason for the observed conflict between ECE and accuracy.

The recently proposed Temporal Ensembling has achieved state-of-the-art results in several semi-supervised learning benchmarks. It maintains an exponential moving average of label predictions on each training example, and penalizes predictions that are inconsistent with this target. However, because the targets change only once per epoch, Temporal Ensembling becomes unwieldy when learning large datasets. To overcome this problem, we propose Mean Teacher, a method that averages model weights instead of label predictions. As an additional benefit, Mean Teacher improves test accuracy and enables training with fewer labels than Temporal Ensembling. Without changing the network architecture, Mean Teacher achieves an error rate of 4.35% on SVHN with 250 labels, outperforming Temporal Ensembling trained with 1000 labels. We also show that a good network architecture is crucial to performance. Combining Mean Teacher and Residual Networks, we improve the state of the art on CIFAR-10 with 4000 labels from 10.55% to 6.28%, and on ImageNet 2012 with 10% of the labels from 35.24% to 9.11%.

Despite the popularity and efficacy of knowledge distillation, there is limited understanding of why it helps. In order to study the generalization behavior of a distilled student, we propose a new theoretical framework that leverages supervision complexity: a measure of alignment between teacher-provided supervision and the student's neural tangent kernel. The framework highlights a delicate interplay among the teacher's accuracy, the student's margin with respect to the teacher predictions, and the complexity of the teacher predictions. Specifically, it provides a rigorous justification for the utility of various techniques that are prevalent in the context of distillation, such as early stopping and temperature scaling. Our analysis further suggests the use of online distillation, where a student receives increasingly more complex supervision from teachers in different stages of their training. We demonstrate efficacy of online distillation and validate the theoretical findings on a range of image classification benchmarks and model architectures.

Teacher-student knowledge distillation is a popular technique for compressing today's prevailing large language models into manageable sizes that fit low-latency downstream applications. Both the teacher and the choice of transfer set used for distillation are crucial ingredients in creating a high quality student. Yet, the generic corpora used to pretrain the teacher and the corpora associated with the downstream target domain are often significantly different, which raises a natural question: should the student be distilled over the generic corpora, so as to learn from high-quality teacher predictions, or over the downstream task corpora to align with finetuning? Our study investigates this trade-off using Domain Classification (DC) and Intent Classification/Named Entity Recognition (ICNER) as downstream tasks. We distill several multilingual students from a larger multilingual LM with varying proportions of generic and task-specific datasets, and report their performance after finetuning on DC and ICNER. We observe significant improvements across tasks and test sets when only task-specific corpora is used. We also report on how the impact of adding task-specific data to the transfer set correlates with the similarity between generic and task-specific data. Our results clearly indicate that, while distillation from a generic LM benefits downstream tasks, students learn better using target domain data even if it comes at the price of noisier teacher predictions. In other words, target domain data still trumps teacher knowledge.

Traditional knowledge distillation in classification problems transfers the knowledge via class correlations in the soft label produced by teacher models, which are not available in regression problems like stock trading volume prediction. To remedy this, we present a novel distillation framework for training a light-weight student model to perform trading volume prediction given historical transaction data. Specifically, we turn the regression model into a probabilistic forecasting model, by training models to predict a Gaussian distribution to which the trading volume belongs. The student model can thus learn from the teacher at a more informative distributional level, by matching its predicted distributions to that of the teacher. Two correlational distillation objectives are further introduced to encourage the student to produce consistent pair-wise relationships with the teacher model. We evaluate the framework on a real-world stock volume dataset with two different time window settings. Experiments demonstrate that our framework is superior to strong baseline models, compressing the model size by $5\times$ while maintaining $99.6\%$ prediction accuracy. The extensive analysis further reveals that our framework is more effective than vanilla distillation methods under low-resource scenarios.