Enhances model compression with multi-teacher knowledge distillation.
problem Uncertainty evaluation and diverse teacher expertise in model deployment.
method Bayesian inference and teacher-informed prior with entropy-based weighting.
result Improved predictive accuracy and robust uncertainty quantification.
Proposes RaT to mitigate bias in student-teacher estimation.
problem Systematic bias in teacher's predictions propagates to student model.
method Uses teacher to estimate residuals in student's predictions.
result RaT method reduces teacher bias effect and achieves optimal rate.
The teacher-student (T/S) learning has been shown to be effective for a variety of problems such as domain adaptation and model compression. One shortcoming of the T/S learning is that a teacher model, not always perfect, sporadically produces wrong guidance in form of posterior probabilities that misleads the student …
In this paper, we make an important step towards the black-box machine teaching by considering the cross-space machine teaching, where the teacher and the learner use different feature representations and the teacher can not fully observe the learner's model. In such scenario, we study how the teacher is still able to …
Distilled models often fail to match teacher models, despite improving generalization.
problem The discrepancy between teacher and student predictive distributions remains large.
method Investigated the optimization difficulties and dataset details affecting student performance.
result Optimizing for matching the teacher does not always lead to better generalization.
Aims to learn from multiple unpredictable teachers with minimal interaction.
problem Learning from multiple non-deterministic teachers with low interaction cost.
method Develops a framework and an active learning algorithm to estimate a distribution over policy space.
result Significantly reduces interaction with teachers without compromising performance.
This paper investigates teacher hacking during language model distillation and proposes methods to mitigate it.
problem Teacher hacking during language model distillation, leading to suboptimal performance.
method A controlled experimental setup involving an oracle LM, teacher LM, and student LM, using fixed offline or online data generation techniques.
result Data diversity is the key factor in preventing teacher hacking during distillation.
Improved learning to reweight using deep interactions between student and teacher models.
problem Limitation of existing learning to reweight methods in utilizing student model's internal states.
method Proposes an algorithm that uses the student model's internal states to the teacher model, which returns adaptive weights to enhance student model training.
result Significant improvement over previous methods in image classification and neural machine translation experiments.
Random feature models can outperform a weak teacher with early stopping.
problem Generalization from a weak to a strong model in random feature networks.
method Random feature models, early stopping, proving weak-to-strong generalization.
result Random feature models can outperform a weak teacher with early stopping.
A new KD layer lets student models learn and apply teacher knowledge explicitly.
problem Implicit action of traditional KD on student's feature transform limits its use in intermediate layers.
method Proposes a learnable KD layer that explicitly embeds teacher's knowledge in feature transform.
result Improves KD with two abilities: leveraging teacher's knowledge and feeding forward knowledge deeper.
Machine teaching addresses the problem of finding the best training data that can guide a learning algorithm to a target model with minimal effort. In conventional settings, a teacher provides data that are consistent with the true data distribution. However, for sequential learners which actively choose their queries,…
We present a novel framework of knowledge distillation that is capable of learning powerful and efficient student models from ensemble teacher networks. Our approach addresses the inherent model capacity issue between teacher and student and aims to maximize benefit from teacher models during distillation by reducing t…
The paper reveals three mechanisms for weak-to-strong generalization.
problem Understanding the mechanisms behind weak-to-strong generalization in imperfect labeling scenarios.
method Theoretical analysis of simple models including ridge regression and weighted ridge regression, and a nonlinear multi-index setting.
result A student model can compensate for a teacher's under-regularization and achieve lower test error.
Knowledge Distillation (KD) consists of transferring âknowledgeâ from one machine learning model (the teacher) to another (the student). Commonly, the teacher is a high-capacity model with formidable performance, while the student is more compact. By transferring knowledge, one hopes to benefit from the studentâs…
Role-wise data augmentation improves knowledge distillation effectiveness.
problem Existing knowledge distillation methods fail to utilize the full potential of teacher-student data interaction.
method Design and implement data augmentation agents with distinct roles for teacher and student.
result Specially tailored data points enhance the demonstration of teacher's knowledge to the student.
Estimates model performance from compute budget for distillation.
problem Risk mitigation in large-scale distillation.
method Distillation scaling law based on compute budget allocation.
result Maximizes student performance with compute-optimal allocation.
Two-layer ReLU networks outperform kernel methods in teacher-student settings.
problem Understanding the excess risk of two-layer ReLU neural networks in teacher-student models.
method Investigated a two-phase training process for a student network, comparing it to kernel methods.
result The student network reaches near-global optimality and outperforms kernel methods in minimax optimal rate.
The paper analyzes knowledge distillation in wide neural networks, providing theoretical insights and practical implications.
problem Lack of theoretical understanding of knowledge distillation in wide neural networks.
method Theoretical analysis of knowledge distillation in a linearized model of a wide neural network, introducing a metric of task training difficulty.
result For a perfect teacher, a high ratio of teacher's soft labels can be beneficial. For imperfect teacher, hard labels can correct wrong predictions.
After a large "teacher" neural network has been trained on labeled data, the probabilities that the teacher assigns to incorrect classes reveal a lot of information about the way in which the teacher generalizes. By training a small "student" model to match these probabilities, it is possible to transfer most of the ge…
Despite the fact that deep neural networks are powerful models and achieve appealing results on many tasks, they are too large to be deployed on edge devices like smartphones or embedded sensor nodes. There have been efforts to compress these networks, and a popular method is knowledge distillation, where a large (teac…
A new method transfers adversarial robustness from teacher to student using feature distillation.
problem Adversarial robustness transfer across different models and tasks.
method Guided Adversarial Contrastive Distillation (GACD) with contrastive learning and sample reweighted estimation.
result GACD effectively transfers adversarial robustness from teacher to student, achieving comparable or better results.
Optimal self-distillation improves generative models' velocity risk and mode recovery.
problem Improving generative models' velocity risk and mode recovery.
method Proved optimal self-distillation for rectified flow via linear probing, derived mixing coefficient, and provided validation tuning.
result Optimal self-distillation improves velocity risk and mode recovery.
Researchers study how teachers' advising relationships influence their perceptions of satisfaction and students, not policy influence.
problem Understanding the relationship between teachers' advising relationships and their perceptions of satisfaction and students.
method Proposed a novel joint model of network and item responses (JNIRM) with correlated latent variables.
result Teachers' advising relationships contribute more to satisfaction and students than to influence over educational policies.
Paper generalizes teacher-student model for realistic data.
problem Capturing learning curves for realistic datasets.
method Introduces a Gaussian covariate generalization of the teacher-student model.
result Generalized model captures learning curves for various realistic data sets.
A new training method improves MLIPs for faster, lighter simulations.
problem High computational and memory costs of complex MLIPs for large-scale MD simulations.
method Teacher-student training framework using latent atomic energy knowledge.
result Lightweight student MLIPs achieve faster MD speeds and comparable accuracy to teachers.
FairDTD improves fairness in GNNs by distilling dual teacher knowledge, balancing utility and bias.
problem Bias in GNN predictions due to sensitive attributes.
method Dual-Teacher Distillation with a causal graph model, feature and structure teachers, and graph-level distillation.
result Achieves optimal fairness while preserving high model utility.
Knowledge distillation improves model accuracy by mimicking teacher model probabilities.
problem Improving model accuracy through model compression.
method Casting knowledge distillation as a semiparametric inference problem, deriving new guarantees, and developing enhancements.
result Enhancements improve student performance by mitigating teacher overfitting and underfitting.
Paper analyzes how neural networks learn from a teacher in a specific setting.
problem Understanding how two-layer ReLU neural networks learn from a teacher in a regression model.
method Used gradient descent with specific regularization and over-parameterization, combined with measure representation and sparse estimation.
result Student network can identify teacher network parameters with high probability via gradient descent.
This study improves knowledge distillation for RNN-T models with noisy labels.
problem Challenges in distilling knowledge from RNN-T models with variable quality teachers.
method Full-sum distillation and sequence-level knowledge distillation.
result Full-sum distillation outperforms other methods for RNN-T models, especially for bad teachers.
We analyze the dynamics of training deep ReLU networks and their implications on generalization capability. Using a teacher-student setting, we discovered a novel relationship between the gradient received by hidden student nodes and the activations of teacher nodes for deep ReLU networks. With this relationship and th…
Recent advances in deep learning have facilitated the demand of neural models for real applications. In practice, these applications often need to be deployed with limited resources while keeping high accuracy. This paper touches the core of neural models in NLP, word embeddings, and presents a new embedding distillati…
Student-teacher learning improves generalization with noisy inputs.
problem Transfer knowledge from clean inputs to noisy inputs.
method Analyzes student-teacher learning using deep linear networks and experiments with nonlinear networks.
result Three factors are vital for success: zero training loss, teacher knowledge, and feature decomposition.
Improved machine learning models outperform their simpler counterparts by using imperfect labels.
problem Improving model performance using imperfect labels.
method Random feature ridge regression (RFRR) with a deterministic equivalent for excess test error.
result The student model can outperform the teacher model regardless of the teacher's scaling law, achieving the minimax optimal rate.
A new method for distilling predictions from a teacher model to a student model without original training data.
problem Training without original labeled data.
method Prediction-only distillation scheme for linear and logistic regression.
result Prediction mixing can outperform both the teacher and pure-distilled models.
Self-training with noisy student-teacher boosts keyword spotting accuracy.
problem Robust keyword spotting in challenging conditions.
method Aggressive data augmentation and self-training with noisy student-teacher approach.
result Significant accuracy improvement in difficult conditions, up to 60%.
We propose a novel way to train ranking models, such as recommender systems, that are both effective and efficient. Knowledge distillation (KD) was shown to be successful in image recognition to achieve both effectiveness and efficiency. We propose a KD technique for learning to rank problems, called \emph{ranking dist…
Self-distillation improves model performance by increasing teacher diversity and smoothing predictions.
problem Improving model generalization and performance through self-distillation.
method Interpreting self-distillation as MAP estimation and proposing instance-specific label smoothing.
result Self-distillation enhances model performance by increasing teacher diversity and smoothing predictions.
Distillation works even with hard labels from overparameterized teacher, leading to better performance.
problem Improving model performance with hard labels from overparameterized teacher.
method Training a student model on a large held-out dataset labeled by a highly overparameterized teacher.
result Student model outperforms traditional approaches due to double descent phenomenon.
Recent advances in pre-training huge models on large amounts of text through self supervision have obtained state-of-the-art results in various natural language processing tasks. However, these huge and expensive models are difficult to use in practise for downstream tasks. Some recent efforts use knowledge distillatio…
A new teacher-class network method compresses DNNs by distributing knowledge to multiple student networks.
problem Overwhelming size of Deep Neural Networks (DNNs).
method Single teacher with multiple student networks, transferring knowledge to each student.
result The combined knowledge of the class of students achieves better performance and reduces parameters.
A new distillation method transfers channel information from teacher to student.
problem Transfer knowledge from teacher to student with fewer parameters and calculations.
method Channel Distillation (CD) and Guided Knowledge Distillation (GKD) with loss decay.
result Achieved 27.68% top-1 error on ImageNet with ResNet18, outperforming state-of-the-art methods.
Fastest video anomaly detection via teacher-student distillation.
problem Anomaly detection in video at high speed.
method Adversarial knowledge distillation from object-level teacher models.
result 7-62 times faster than state-of-the-art methods.
Some machine learning applications involve training data that is sensitive, such as the medical histories of patients in a clinical trial. A model may inadvertently and implicitly store some of its training data; careful analysis of the model may therefore reveal sensitive information. To address this problem, we demon…
Knowledge distillation is a widely used technique for model compression. We posit that the teacher model used in a distillation setup, captures relationships between classes, that extend beyond the original dataset. We empirically show that a teacher model can transfer this knowledge to a student model even on an {\it …
A zoo of deep nets is available these days for almost any given task, and it is increasingly unclear which net to start with when addressing a new task, or which net to use as an initialization for fine-tuning a new model. To address this issue, in this paper, we develop knowledge flow which moves 'knowledge' from mult…
ReOPD uses pre-collected teacher trajectories to distill knowledge from multi-turn interactions.
problem The cost of fully online on-policy distillation for multi-turn interactions.
method ReOPD, an off-environment alternative that reuses pre-collected teacher trajectories as replayed prefixes, addressing the prefix trap and distribution shift.
result ReOPD preserves or improves OPD-level accuracy, uses zero tool calls, and is at least 4imes faster per training step. PLD distills knowledge using choice-theoretic Plackett-Luce model.
problem Model compression and knowledge transfer between large and small networks.
method PLD uses a weighted list-wise ranking loss based on the Plackett-Luce model.
result PLD achieves consistent gains across diverse architectures and distillation methods.
New model shows weak teachers can help strong students learn even with imperfect labels.
problem Improving strong student's performance with weak teacher's imperfect pseudolabels.
method Stylized overparameterized spiked covariance model with Gaussian covariates, proving two phases of generalization.
result Provable successful and random guessing phases of strong student's generalization.