Mikhail Petrov

Ethan Harvey, Mikhail Petrov, Michael C. Hughes

We pursue transfer learning to improve classifier accuracy on a target task with few labeled examples available for training. Recent work suggests that using a source task to learn a prior distribution over neural net weights, not just an initialization, can boost target task performance. In this study, we carefully compare transfer learning with and without source task informed priors across 5 datasets. We find that standard transfer learning informed by an initialization only performs far better than reported in previous comparisons. The relative gains of methods using informative priors over standard transfer learning vary in magnitude across datasets. For the scenario of 5-300 examples per class, we find negative or negligible gains on 2 datasets, modest gains (between 1.5-3 points of accuracy) on 2 other datasets, and substantial gains (>8 points) on one dataset. Among methods using informative priors, we find that an isotropic covariance appears competitive with learned low-rank covariance matrix while being substantially simpler to understand and tune. Further analysis suggests that the mechanistic justification for informed priors -- hypothesized improved alignment between train and test loss landscapes -- is not consistently supported due to high variability in empirical landscapes. We release code to allow independent reproduction of all experiments.

Ethan Harvey, Mikhail Petrov, Michael C. Hughes

A number of popular transfer learning methods rely on grid search to select regularization hyperparameters that control over-fitting. This grid search requirement has several key disadvantages: the search is computationally expensive, requires carving out a validation set that reduces the size of available data for model training, and requires practitioners to specify candidate values. In this paper, we propose an alternative to grid search: directly learning regularization hyperparameters on the full training set via model selection techniques based on the evidence lower bound ("ELBo") objective from variational methods. For deep neural networks with millions of parameters, we specifically recommend a modified ELBo that upweights the influence of the data likelihood relative to the prior while remaining a valid bound on the evidence for Bayesian model selection. Our proposed technique overcomes all three disadvantages of grid search. We demonstrate effectiveness on image classification tasks on several datasets, yielding heldout accuracy comparable to existing approaches with far less compute time.

Ethan Harvey, Mikhail Petrov, Michael C. Hughes

When training large flexible models on limited data, avoiding overfitting is a practical concern. Common grid search or smarter search methods rely on expensive separate runs at each candidate hyperparameter while carving out a validation set that reduces available training data. In this paper, we consider direct gradient-based learning of regularization hyperparameters on the full training set via the evidence lower bound ("ELBo") objective from Bayesian variational methods. We focus on scenarios where the model is over-parameterized for flexibility while the approximate posterior is chosen to be Gaussian with isotropic covariance for tractability, even though it cannot match the true posterior exactly. In such scenarios, we find the ELBo prioritizes posteriors that match the prior variance, which leads to severely underfitting the data. Instead, we recommend a data-emphasized ELBo that upweights the influence of the data likelihood relative to the prior. In Bayesian transfer learning of classifiers for text and images, our method reduces 88+ hour grid searches of past work to under 3 hours while delivering comparable accuracy. We further demonstrate how our approach enables efficient yet accurate approximations of Gaussian processes with learnable length-scale kernels.