Rafael Rios Müller

Andrew Qiu, Shubham Barhate, Hin Wai Lui et al.

Deep learning for conversion prediction has found widespread applications in online advertising. These models have become more complex as they are trained to jointly predict multiple objectives such as click, add-to-cart, checkout and other conversion types. Additionally, the capacity and performance of these models can often be increased with the use of embedding tables that encode high cardinality categorical features such as advertiser, user, campaign, and product identifiers (IDs). These embedding tables can be pre-trained, but also learned end-to-end jointly with the model to directly optimize the model objectives. Training these large tables is challenging due to: gradient sparsity, the high cardinality of the categorical features, the non-uniform distribution of IDs and the very high label sparsity. These issues make training prone to both slow convergence and overfitting after the first epoch. Previous works addressed the multi-epoch overfitting issue by using: stronger feature hashing to reduce cardinality, filtering of low frequency IDs, regularization of the embedding tables, re-initialization of the embedding tables after each epoch, etc. Some of these techniques reduce overfitting at the expense of reduced model performance if used too aggressively. In this paper, we share key learnings from the development of embedding table optimization and multi-epoch training in Pinterest Ads Conversion models. We showcase how our Sparse Optimizer speeds up convergence, and how multi-epoch overfitting varies in severity between different objectives in a multi-task model depending on label sparsity. We propose a new approach to deal with multi-epoch overfitting: the use of a frequency-adaptive learning rate on the embedding tables and compare it to embedding re-initialization. We evaluate both methods offline using an industrial large-scale production dataset.

Yeming Wen, Ghassen Jerfel, Rafael Muller et al.

Ensemble methods which average over multiple neural network predictions are a simple approach to improve a model's calibration and robustness. Similarly, data augmentation techniques, which encode prior information in the form of invariant feature transformations, are effective for improving calibration and robustness. In this paper, we show a surprising pathology: combining ensembles and data augmentation can harm model calibration. This leads to a trade-off in practice, whereby improved accuracy by combining the two techniques comes at the expense of calibration. On the other hand, selecting only one of the techniques ensures good uncertainty estimates at the expense of accuracy. We investigate this pathology and identify a compounding under-confidence among methods which marginalize over sets of weights and data augmentation techniques which soften labels. Finally, we propose a simple correction, achieving the best of both worlds with significant accuracy and calibration gains over using only ensembles or data augmentation individually. Applying the correction produces new state-of-the art in uncertainty calibration across CIFAR-10, CIFAR-100, and ImageNet.

Rafael Müller, Simon Kornblith, Geoffrey Hinton

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 generalization ability of the teacher to the student, often producing a much better small model than directly training the student on the training data. The transfer works best when there are many possible classes because more is then revealed about the function learned by the teacher, but in cases where there are only a few possible classes we show that we can improve the transfer by forcing the teacher to divide each class into many subclasses that it invents during the supervised training. The student is then trained to match the subclass probabilities. For datasets where there are known, natural subclasses we demonstrate that the teacher learns similar subclasses and these improve distillation. For clickthrough datasets where the subclasses are unknown we demonstrate that subclass distillation allows the student to learn faster and better.

Rafael Müller, Simon Kornblith, Geoffrey Hinton

The generalization and learning speed of a multi-class neural network can often be significantly improved by using soft targets that are a weighted average of the hard targets and the uniform distribution over labels. Smoothing the labels in this way prevents the network from becoming over-confident and label smoothing has been used in many state-of-the-art models, including image classification, language translation and speech recognition. Despite its widespread use, label smoothing is still poorly understood. Here we show empirically that in addition to improving generalization, label smoothing improves model calibration which can significantly improve beam-search. However, we also observe that if a teacher network is trained with label smoothing, knowledge distillation into a student network is much less effective. To explain these observations, we visualize how label smoothing changes the representations learned by the penultimate layer of the network. We show that label smoothing encourages the representations of training examples from the same class to group in tight clusters. This results in loss of information in the logits about resemblances between instances of different classes, which is necessary for distillation, but does not hurt generalization or calibration of the model's predictions.