An official website of the United States government
Recent approaches have shown promises distilling diffusion models into efficient one-step generators. Among them, Distribution Matching Distillation (DMD) produces one-step generators that match their teacher in distribution, without enforcing a one-to-one correspondence with the sampling trajectories of their teachers. However, to ensure stable training, DMD requires an additional regression loss computed using a large set of noise-image pairs generated by the teacher with many steps of a deterministic sampler. This is costly for large-scale text-to-image synthesis and limits the student's quality, tying it too closely to the teacher's original sampling paths. We introduce DMD2, a set of techniques that lift this limitation and improve DMD training. First, we eliminate the regression loss and the need for expensive dataset construction. We show that the resulting instability is due to the fake critic not estimating the distribution of generated samples accurately and propose a two time-scale update rule as a remedy. Second, we integrate a GAN loss into the distillation procedure, discriminating between generated samples and real images. This lets us train the student model on real data, mitigating the imperfect real score estimation from the teacher model, and enhancing quality. Lastly, we modify the training procedure to enable multi-step sampling. We identify and address the training-inference input mismatch problem in this setting, by simulating inference-time generator samples during training time. Taken together, our improvements set new benchmarks in one-step image generation, with FID scores of 1.28 on ImageNet-64x64 and 8.35 on zero-shot COCO 2014, surpassing the original teacher despite a 500X reduction in inference cost. Further, we show our approach can generate megapixel images by distilling SDXL, demonstrating exceptional visual quality among few-step methods.
more »
« less
Understanding the gains from repeated self-distillation
Self-Distillation is a special type of knowledge distillation where the student model has the same architecture as the teacher model. Despite using the same architecture and the same training data, self-distillation has been empirically observed to improve performance, especially when applied repeatedly. For such a process, there is a fundamental question of interest: How much gain is possible by applying multiple steps of self-distillation? To investigate this relative gain, we propose using the simple but canonical task of linear regression. Our analysis shows that the excess risk achieved by multi-step self-distillation can significantly improve upon a single step of self-distillation, reducing the excess risk by a factor of , where is the input dimension. Empirical results on regression tasks from the UCI repository show a reduction in the learnt model's risk (MSE) by up to %.
more »
« less
- Award ID(s):
- 2229876
- PAR ID:
- 10662752
- Publisher / Repository:
- 38th Annual Conference on Neural Information Processing Systems (NeurIPS 2024)
- Date Published:
- Page Range / eLocation ID:
- 7759-7796
- Format(s):
- Medium: X
- Location:
- Vancouver, Canada
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Fan, Yanan; Nott, David; Smith, Michael S; Dortet-Bernadet, Jean-Luc. (Ed.)Quantile regression is widely seen as an ideal tool to understand complex predictor-response relations. Its biggest promise rests in its ability to quantify whether and how predictor effects vary across response quantile levels. But this promise has not been fully met due to a lack of statistical estimation methods that perform a rigorous, joint analysis of all quantile levels. This gap has been recently bridged by Yang and Tokdar [18]. Here we demonstrate how their joint quantile regression method, as encoded in the R package qrjoint, offers a comprehensive and model-based regression analysis framework. This chapter is an R vignette where we illustrate how to fit models, interpret coefficients, improve and compare models and obtain predictions under this framework. Our case study is an application to ecology where we analyse how the abundance of red maple trees depends on topographical and geographical features of the location. A complete absence of the species contributes excess zeros in the response data. We treat such excess zeros as left censoring in the spirit of a Tobit regression analysis. By utilising the generative nature of the joint quantile regression model, we not only adjust for censoring but also treat it as an object of independent scientific interest.more » « less
-
Abstract We propose a sparse deep ReLU network (SDRN) estimator of the regression function obtained from regularized empirical risk minimization with a Lipschitz loss function. Our framework can be applied to a variety of regression and classification problems. We establish novel nonasymptotic excess risk bounds for our SDRN estimator when the regression function belongs to a Sobolev space with mixed derivatives. We obtain a new, nearly optimal, risk rate in the sense that the SDRN estimator can achieve nearly the same optimal minimax convergence rate as one-dimensional nonparametric regression with the dimension involved in a logarithm term only when the feature dimension is fixed. The estimator has a slightly slower rate when the dimension grows with the sample size. We show that the depth of the SDRN estimator grows with the sample size in logarithmic order, and the total number of nodes and weights grows in polynomial order of the sample size to have the nearly optimal risk rate. The proposed SDRN can go deeper with fewer parameters to well estimate the regression and overcome the overfitting problem encountered by conventional feedforward neural networks.more » « less
-
Brain-Computer interfaces (BCIs) are typically designed to be lightweight and responsive in real-time to provide users timely feedback. Classical feature engineering is computationally efficient but has low accuracy, whereas the recent neural networks (DNNs) improve accuracy but are computationally expensive and incur high latency. As a promising alternative, the low-dimensional computing (LDC) classifier based on vector symbolic architecture (VSA), achieves small model size yet higher accuracy than classical feature engineering methods. However, its accuracy still lags behind that of modern DNNs, making it challenging to process complex brain signals. To improve the accuracy of a small model, knowledge distillation is a popular method. However, maintaining a constant level of distillation between the teacher and student models may not be the best way for a growing student during its progressive learning stages. In this work, we propose a simple scheduled knowledge distillation method based on curriculum data order to enable the student to gradually build knowledge from the teacher model, controlled by an alpha scheduler. Meanwhile, we employ the LDC/VSA as the student model to enhance the on-device inference efficiency for tiny BCI devices that demand low latency. The empirical results have demonstrated that our approach achieves better tradeoff between accuracy and hardware efficiency compared to other methods.more » « less
-
Brain-Computer interfaces (BCIs) are typically designed to be lightweight and responsive in real-time to provide users timely feedback. Classical feature engineering is computationally efficient but has low accuracy, whereas the recent neural networks (DNNs) improve accuracy but are computationally expensive and incur high latency. As a promising alternative, the low-dimensional computing (LDC) classifier based on vector symbolic architecture (VSA), achieves small model size yet higher accuracy than classical feature engineering methods. However, its accuracy still lags behind that of modern DNNs, making it challenging to process complex brain signals. To improve the accuracy of a small model, knowledge distillation is a popular method. However, maintaining a constant level of distillation between the teacher and student models may not be the best way for a growing student during its progressive learning stages. In this work, we propose a simple scheduled knowledge distillation method based on curriculum data order to enable the student to gradually build knowledge from the teacher model, controlled by an scheduler. Meanwhile, we employ the LDC/VSA as the student model to enhance the on-device inference efficiency for tiny BCI devices that demand low latency. The empirical results have demonstrated that our approach achieves better tradeoff between accuracy and hardware efficiency compared to other methods.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript
- Conference Paper:
- The DOI is not currently available.
