Source-linked AI summary

Temporal Ensembling for Semi-Supervised Learning

Samuli Laine, Timo Aila

arXiv:1610.02242v3cs.NEcs.LG

TL;DR

Semi-supervised learning needs reliable targets when only a small portion of training data is labeled. The paper forms self-ensembles from predictions across epochs and stochastic conditions, achieving strong benchmark results and tolerance to incorrect labels. Its comparisons are qualified by differences in augmentation and pooling setups across methods.

  • Problem

    Semi-supervised learning requires useful targets when only a small portion of training data is labeled.

  • Method

    Self-ensembling uses predictions from one network across epochs, regularization, and input augmentations as targets for unlabeled examples.

  • Results

    The methods set new semi-supervised benchmark records across SVHN, CIFAR-10, and CIFAR-100, with gains under non-augmented and augmented settings.

  • Takeaways & Limitations

    Self-ensembling also improves fully supervised accuracy and provides tolerance against incorrect labels.

  • Takeaways & Limitations

    Comparisons with prior work are limited by differing augmentation and fractional max-pooling configurations.

Abstract

from arXiv · show

In this paper, we present a simple and efficient method for training deep neural networks in a semi-supervised setting where only a small portion of training data is labeled. We introduce self-ensembling, where we form a consensus prediction of the unknown labels using the outputs of the network-in-training on different epochs, and most importantly, under different regularization and input augmentation conditions. This ensemble prediction can be expected to be a better predictor for the unknown labels than the output of the network at the most recent training epoch, and can thus be used as a target for training. Using our method, we set new records for two standard semi-supervised learning benchmarks, reducing the (non-augmented) classification error rate from 18.44% to 7.05% in SVHN with 500 labels and from 18.63% to 16.55% in CIFAR-10 with 4000 labels, and further to 5.12% and 12.16% by enabling the standard augmentations. We additionally obtain a clear improvement in CIFAR-100 classification accuracy by using random images from the Tiny Images dataset as unlabeled extra inputs during training. Finally, we demonstrate good tolerance to incorrect labels.

1 INTRODUCTION

The paper extends ensemble-based prediction into semi-supervised training by using a network’s outputs across epochs, regularization conditions, and input augmentations as targets for unlabeled data. It introduces Π-model and temporal ensembling, reporting improved semi-supervised and fully supervised performance with tolerance to incorrect labels.

  • Motivation: Self-ensembling averages predictions from one network across training epochs and different regularization or augmentation conditions.The ensemble prediction is intended to better approximate unknown labels than the latest network output.
  • Semi-supervised objective: Unlabeled inputs receive inferred ensemble targets, while labeled inputs retain the supervised cross-entropy objective.The cross-entropy component is evaluated only for labeled inputs.
  • Methods: The paper presents two self-ensembling implementations: the Π-model and temporal ensembling.These methods differ in how they obtain the multiple predictions used for consistency training.
  • Reported contribution: Both approaches surpass prior state-of-the-art semi-supervised results by a considerable margin.The paper also reports improved accuracy in fully labeled settings and tolerance to incorrect labels.
  • Relation to prior work: The methods relate to transform/stability loss, the ladder network’s Γ-model, and bootstrapping for noisy labels.The Π-model is described as a special case or simplification of related consistency-based approaches.

2 SELF-ENSEMBLING DURING TRAINING

Self-ensembling trains on consistency between stochastic network evaluations or predictions accumulated from earlier epochs, alongside supervised learning on labeled examples. Temporal ensembling reduces computation and target noise relative to the Π-model, but requires stored cross-epoch predictions and an additional momentum parameter.

  • Π-model: The Π-model compares two stochastic evaluations of each input and penalizes differences between their prediction vectors.Its loss combines labeled cross-entropy with an unsupervised mean-square consistency term weighted by w(t).
  • Π-model: The Π-model compares full output vectors rather than only final classifications, imposing a stronger consistency requirement.Dropout, Gaussian noise, and input augmentations create the variation between evaluations.
  • Training schedule: The unsupervised loss weight ramps up slowly from zero during early training to avoid a degenerate solution.The implementation ramps the weight along a Gaussian curve during the first 80 epochs.
  • Temporal ensembling: Temporal ensembling accumulates epoch-level predictions with momentum α and bias-corrects them into targets for the next training updates.Recent epochs receive larger weights, while the first epoch has zero targets because no prior predictions exist.
  • Temporal ensembling: Temporal ensembling evaluates each input once per epoch, making training faster and producing less noisy targets than the Π-model.Its trade-offs are auxiliary storage across epochs and the additional hyperparameter α.
  • Open direction: The paper leaves uncertainty-aware temporal ensembling based on higher-order prediction statistics for future work.Tracking prediction variance could support weighting more certain outputs more heavily.

3 RESULTS

Across CIFAR-10, SVHN, and CIFAR-100, self-ensembling methods substantially improve semi-supervised classification, with further gains from augmentation and unlabeled extra data. The methods also show tolerance to incorrect labels and improve fully supervised results.

  • CIFAR-10: 2.1 percentage points lower CIFAR-10 classification error was achieved with 4000 labels using the non-augmented Π-model.The comparison is against earlier methods.
  • CIFAR-10: 12.16% CIFAR-10 error was achieved with temporal ensembling and standard augmentations, versus 12.36% for the Π-model.Augmentation reduced the Π-model error by 4.2 percentage points to 12.36%, and temporal ensembling trained twice as fast.
  • SVHN: 2.7 percentage points lower SVHN error was obtained with 1000 labels, from 8.11% to 5.43% without augmentation and to 4.42% with standard augmentations.With 500 labels, temporal ensembling further reduced the augmented error rate to 5.12%.
  • CIFAR-100 and Tiny Images: 43.43% and 38.65% CIFAR-100 error rates were obtained without and with augmentation, improving supervised learning by 7.8 and 5.9 percentage points.The experiment used 10000 labels.
  • CIFAR-100 and Tiny Images: 2.7 percentage points lower CIFAR-100 error resulted from adding 500k randomly selected unlabeled Tiny Images, from 26.30% to 23.63%.The result indicates that randomly selected natural images can provide useful unlabeled inputs.
  • Tolerance to incorrect labels: Temporal ensembling retained over ninety percent SVHN classification accuracy when 80% of labels were randomized, while standard supervised training degraded rapidly.With half the labels randomized, temporal ensembling showed almost perfect resistance to disinformation.

4 RELATED WORK

The paper situates self-ensembling among semi-supervised, noisy-label, and ensemble-based approaches, while identifying related architectures and possible extensions.

  • Prior semi-supervised work includes ladder-network variants, transform/stability loss, label propagation, and generative adversarial networks.
  • The Π-model shares the transform/stability principle and simplifies the Γ-model by comparing network outputs directly.
  • Self-ensembling derives variability from dropout and augmentation in one network rather than independently trained networks on different data subsets.
  • Temporal ensembling connects to bootstrapping methods for training with noisy labels.
  • The authors identify incorporating a generative component and applying the methods to regression as possible future directions.

A NETWORK ARCHITECTURE, TEST SETUP, AND TRAINING PARAMETERS

The experiments use a convolutional architecture, dataset-specific preprocessing, and scheduled optimization, with additional details concerning extra data and convergence constraints.

  • The network uses weight normalization, mean-only batch normalization, leaky ReLU, and max pooling throughout the tested architecture.
  • Training uses Adam, dataset-dependent maximum learning rates, and a Gaussian ramp-up for the unsupervised loss and learning rate.
  • CIFAR-10 preprocessing applies ZCA, horizontal flips, and independent random translations to the Π-model branches.
  • SVHN inputs are standardized, and experiments use only the official 73,257-image training set without horizontal flips.
  • A slow unsupervised-cost ramp-up is important for convergence, while very low-label SVHN runs sometimes experienced exploding optimization.
  • The training procedure does not guarantee stratified labeled examples across minibatches when label density is very low.
  • Restricted Tiny Images extra data contains 237,203 category-selected images, while CIFAR-100 overlaps substantially with the broader Tiny Images source.
Loading 1610.02242v3…