Hydra: Preserving Ensemble Diversity for Model Distillation

By Linh Hoang Tran, Bastiaan S. Veeling, Kevin Roth, Jakub Bartlomiej Swiatkowski, Joshua V. Dillon, Jasper Snoek, Stephan Mandt, Tim Salimans, Sebastian Nowozin, Rodolphe Jenatton. 2020

65 citations recorded by the source corpus.

Ensembles of models have been empirically shown to improve predictive performance and to yield robust measures of uncertainty. However, they are expensive in computation and memory. Therefore, recent research has focused on distilling ensembles into a single compact model, reducing the computational and memory burden of the ensemble while trying to preserve its predictive behavior. Most existing distillation formulations summarize the ensemble by capturing its average predictions. As a result, the diversity of the ensemble predictions, stemming from each individual member, is lost. Thus, the distilled model cannot provide a measure of uncertainty comparable to that of the original ensemble. To retain more faithfully the diversity of the ensemble, we propose a distillation method based on a single multi-headed neural network, which we refer to as Hydra. The shared body network learns a joint feature representation that enables each head to capture the predictive behavior of each ensemble member. We demonstrate that with a slight increase in parameter count, Hydra improves distillation performance on classification and regression settings while capturing the uncertainty behaviour of the original ensemble over both in-domain and out-of-distribution tasks.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Group Knowledge Transfer: Collaborative Training of Large CNNs on the Edge (2020, background): “Moreover, all teachers and students in distillation share the same dataset [38, 40, 41, 42], while in our setting each member (client) can only access its own independent dataset.”
Robust fine-tuning of zero-shot models (2021, background): “through distillation into a single model [65, 94]).”
Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning (2020, background): “The distillation mechanism allows to approximate the prediction of an ensemble by a single neural network (Hinton et al., 2015; Balan et al., 2015; Tran et al., 2020), whereas fast dropout (Wang & Manning, 2013) and deterministic variational inference (Wu et al., 2018) allow to approximate the…”
Group Knowledge Transfer: Federated Learning of Large CNNs at the Edge (2020, background): “Moreover, all teachers and students in distillation share the same dataset [42, 44, 45, 46], while in our setting each member (client) can only access its own independent dataset.”
A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges (2020, background): “• The results of our research show that although a variety of ensemble methods [474], [475], [476], [477], [478], [479], [480] also have a high ability to deal with uncertainty along with good performance and optimizing the performance of other methods, these high capabilities of these methods have ”
Knowledge Distillation and Student-Teacher Learning for Visual Intelligence: A Review and New Outlooks (2020, methodology): “Anil [9] X 7 7 7 7 X X LCE+LEns Chen [28] X X 7 7 7 X X LCE+LEns+LKD Dvornik [51] X 7 7 X 7 X X LCE+LEns+LKD Fukuda [56] X 7 7 7 7 7 7 LCE+LEns Furlanello [57] X 7 7 7 7 7 7 LCE+LEns He [80] 7 X 7 7 X 7 7 L1+LKD Jung [101] X 7 7 7 7 7 7 LCE+LKD Lan [114] X 7 7 7 7 X 7 LCE+LEns Lee [118] X X 7 7 X X ”