skip to main content


The NSF Public Access Repository (NSF-PAR) system and access will be unavailable from 11:00 PM ET on Friday, May 17 until 8:00 AM ET on Saturday, May 18 due to maintenance. We apologize for the inconvenience.

Title: Does Knowledge Distillation Really Work?
Knowledge distillation is a popular technique for training a small student network to emulate a larger teacher model, such as an ensemble of networks. We show that while knowledge distillation can improve student generalization, it does not typically work as it is commonly understood: there often remains a surprisingly large discrepancy between the predictive distributions of the teacher and the student, even in cases when the student has the capacity to perfectly match the teacher. We identify difficulties in optimization as a key reason for why the student is unable to match the teacher. We also show how the details of the dataset used for distillation play a role in how closely the student matches the teacher --- and that more closely matching the teacher paradoxically does not always lead to better student generalization.  more » « less
Award ID(s):
Author(s) / Creator(s):
; ; ; ;
Date Published:
Journal Name:
Advances in neural information processing systems
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Knowledge Distillation (KD) (Hinton et al., 2015) is one of the most effective approaches for deploying large-scale pre-trained language models in low-latency environments by transferring the knowledge contained in the largescale models to smaller student models. Previous KD approaches use the soft labels and intermediate activations generated by the teacher to transfer knowledge to the student model parameters alone. In this paper, we show that having access to non-parametric memory in the form of a knowledge base with the teacher’s soft labels and predictions can further enhance student capacity and improve generalization. To enable the student to retrieve from the knowledge base effectively, we propose a new Retrieval-augmented KD framework with a loss function that aligns the relational knowledge in teacher and student embedding spaces. We show through extensive experiments that our retrieval mechanism can achieve state-of-the-art performance for taskspecific knowledge distillation on the GLUE benchmark (Wang et al., 2018a). 
    more » « less
  2. We propose a novel knowledge distillation (KD) method to selectively instill teacher knowledge into a student model motivated by situations where the student’s capacity is significantly smaller than that of the teachers. In vanilla KD, the teacher primarily sets a predictive target for the student to follow, and we posit that this target is overly optimistic due to the student’s lack of capacity. We develop a novel scaffolding scheme where the teacher, in addition to setting a predictive target, also scaffolds the student’s prediction by censoring hard-to-learn examples. The student model utilizes the same information as the teacher’s soft-max predictions as inputs, and in this sense, our proposal can be viewed as a natural variant of vanilla KD. We show on synthetic examples that censoring hard-examples leads to smoothening the student’s loss landscape so that the student encounters fewer local minima. As a result, it has good generalization properties. Against vanilla KD, we achieve improved performance and are comparable to more intrusive techniques that leverage feature matching on benchmark datasets. 
    more » « less
  3. Abstract Objective . Deep-learning (DL)-based dose engines have been developed to alleviate the intrinsic compromise between the calculation accuracy and efficiency of the traditional dose calculation algorithms. However, current DL-based engines typically possess high computational complexity and require powerful computing devices. Therefore, to mitigate their computational burdens and broaden their applicability to a clinical setting where resource-limited devices are available, we proposed a compact dose engine via knowledge distillation (KD) framework that offers an ultra-fast calculation speed with high accuracy for prostate Volumetric Modulated Arc Therapy (VMAT). Approach . The KD framework contains two sub-models: a large pre-trained teacher and a small to-be-trained student. The student receives knowledge transferred from the teacher for better generalization. The trained student serves as the final engine for dose calculation. The model input is patient computed tomography and VMAT dose in water, and the output is DL-calculated patient dose. The ground-truth \dose was computed by the Monte Carlo module of the Monaco treatment planning system. Twenty and ten prostate cases were included for model training and assessment, respectively. The model’s performance (teacher/student/student-only) was evaluated by Gamma analysis and inference efficiency. Main results . The dosimetric comparisons (input/DL-calculated/ground-truth doses) suggest that the proposed engine can effectively convert low-accuracy doses in water to high-accuracy patient doses. The Gamma passing rate (2%/2 mm, 10% threshold) between the DL-calculated and ground-truth doses was 98.64 ± 0.62% (teacher), 98.13 ± 0.76% (student), and 96.95 ± 1.02% (student-only). The inference time was 16 milliseconds (teacher) and 11 milliseconds (student/student-only) using a graphics processing unit device, while it was 936 milliseconds (teacher) and 374 milliseconds (student/student-only) using a central processing unit device. Significance . With the KD framework, a compact dose engine can achieve comparable accuracy to that of a larger one. Its compact size reduces the computational burdens and computing device requirements, and thus such an engine can be more clinically applicable. 
    more » « less
  4. Policy distillation, which transfers a teacher policy to a student policy has achieved great success in challenging tasks of deep reinforcement learning. This teacher-student framework requires a well-trained teacher model which is computationally expensive. Moreover, the performance of the student model could be limited by the teacher model if the teacher model is not optimal. In the light of collaborative learning, we study the feasibility of involving joint intellectual efforts from diverse perspectives of student models. In this work, we introduce dual policy distillation (DPD), a student-student framework in which two learners operate on the same environment to explore different perspectives of the environment and extract knowledge from each other to enhance their learning. The key challenge in developing this dual learning framework is to identify the beneficial knowledge from the peer learner for contemporary learning-based reinforcement learning algorithms, since it is unclear whether the knowledge distilled from an imperfect and noisy peer learner would be helpful. To address the challenge, we theoretically justify that distilling knowledge from a peer learner will lead to policy improvement and propose a disadvantageous distillation strategy based on the theoretical results. The conducted experiments on several continuous control tasks show that the proposed framework achieves superior performance with a learning-based agent and function approximation without the use of expensive teacher models.

    more » « less
  5. The commonsense natural language inference (CNLI) tasks aim to select the most likely follow-up statement to a contextual description of ordinary, everyday events and facts. Current approaches to transfer learning of CNLI models across tasks require many labeled data from the new task. This paper presents a way to reduce this need for additional annotated training data from the new task by leveraging symbolic knowledge bases, such as ConceptNet. We formulate a teacher-student framework for mixed symbolic-neural reasoning, with the large-scale symbolic knowledge base serving as the teacher and a trained CNLI model as the student. This hybrid distillation process involves two steps. The first step is a symbolic reasoning process. Given a collection of unlabeled data, we use an abductive reasoning framework based on Grenander's pattern theory to create weakly labeled data. Pattern theory is an energy-based graphical probabilistic framework for reasoning among random variables with varying dependency structures. In the second step, the weakly labeled data, along with a fraction of the labeled data, is used to transfer-learn the CNLI model into the new task. The goal is to reduce the fraction of labeled data required. We demonstrate the efficacy of our approach by using three publicly available datasets (OpenBookQA, SWAG, and HellaSWAG) and evaluating three CNLI models (BERT, LSTM, and ESIM) that represent different tasks. We show that, on average, we achieve 63% of the top performance of a fully supervised BERT model with no labeled data. With only 1000 labeled samples, we can improve this performance to 72%. Interestingly, without training, the teacher mechanism itself has significant inference power. The pattern theory framework achieves 32.7% accuracy on OpenBookQA, outperforming transformer-based models such as GPT (26.6%), GPT-2 (30.2%), and BERT (27.1%) by a significant margin. We demonstrate that the framework can be generalized to successfully train neural CNLI models using knowledge distillation under unsupervised and semi-supervised learning settings. Our results show that it outperforms all unsupervised and weakly supervised baselines and some early supervised approaches, while offering competitive performance with fully supervised baselines. Additionally, we show that the abductive learning framework can be adapted for other downstream tasks, such as unsupervised semantic textual similarity, unsupervised sentiment classification, and zero-shot text classification, without significant modification to the framework. Finally, user studies show that the generated interpretations enhance its explainability by providing key insights into its reasoning mechanism. 
    more » « less