skip to main content

Title: Test-Time Training with Self-Supervision for Generalization under Distribution Shifts
In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a prediction. This also extends naturally to data in an online stream. Our simple approach leads to improvements on diverse image classification benchmarks aimed at evaluating robustness to distribution shifts.
Authors:
; ; ; ; ;
Award ID(s):
1764033
Publication Date:
NSF-PAR ID:
10249254
Journal Name:
ICML 2020
Sponsoring Org:
National Science Foundation
More Like this
  1. To address the sample selection bias between the training and test data, previous research works focus on reweighing biased training data to match the test data and then building classification models on there weighed raining data. However, how to achieve fairness in the built classification models is under-explored. In this paper, we propose a framework for robust and fair learning under sample selection bias. Our framework adopts there weighing estimation approach for bias correction and the minimax robust estimation approach for achieving robustness on prediction accuracy. Moreover, during the minimax optimization, the fairness is achieved under the worst case, which guarantees the model’s fairness on test data. We further develop two algorithms to handle sample selection bias when test data is both available and unavailable.
  2. n this paper, we use a thermal camera to distinguish hard and soft swipes performed by a user interacting with a natural surface by detecting differences in the thermal signature of the surface due to heat transferred by the user. Unlike prior work, our approach provides swipe pressure classifiers that are user-agnostic, i.e., that recognize the swipe pressure of a novel user not present in the training set, enabling our work to be ported into natural user interfaces without user-specific calibration. Our approach generates average classification accuracy of 76% using random forest classifiers trained on a test set of 9 subjects interacting with paper and wood, with 8 hard and 8 soft test swipes per user. We compare results of the user-agnostic classification to user-aware classification with classifiers trained by including training samples from the user. We obtain average user-aware classification accuracy of 82% by adding up to 8 hard and 8 soft training swipes for each test user. Our approach enables seamless adaptation of generic pressure classification systems based on thermal data to the specific behavior of users interacting with natural user interfaces.
  3. This paper presents a novel zero-shot learning approach towards personalized speech enhancement through the use of a sparsely active ensemble model. Optimizing speech denoising systems towards a particular test-time speaker can improve performance and reduce run-time complexity. However, test-time model adaptation may be challenging if collecting data from the test-time speaker is not possible. To this end, we propose using an ensemble model wherein each specialist module denoises noisy utterances from a distinct partition of training set speakers. The gating module inexpensively estimates test-time speaker characteristics in the form of an embedding vector and selects the most appropriate specialist module for denoising the test signal. Grouping the training set speakers into non-overlapping semantically similar groups is non-trivial and ill-defined. To do this, we first train a Siamese network using noisy speech pairs to maximize or minimize the similarity of its output vectors depending on whether the utterances derive from the same speaker or not. Next, we perform k-means clustering on the latent space formed by the averaged embedding vectors per training set speaker. In this way, we designate speaker groups and train specialist modules optimized around partitions of the complete training set. Our experiments show that ensemble models made upmore »of low-capacity specialists can outperform high-capacity generalist models with greater efficiency and improved adaptation towards unseen test-time speakers.« less
  4. The 2nd Annual WPI-UMASS-UPENN EDM Data Min- ing Challenge required contestants to predict efficient test taking based on log data. In this paper, we describe our theory-driven and psychometric modeling approach. For feature engineering, we employed the Log-Normal Response Time Model for estimating latent person speed, and the Generalized Partial Credit Model for estimating latent person ability. Additionally, we adopted an n-gram feature approach for event sequences. For training a multi-label classifier, we distinguished inefficent test takers who were going too fast and those who were going too slow, instead of using the provided binary target label. Our best-performing ensemble classify er comprised three sets of low-dimensional classi ers, dominated by test-taker speed. While our classi- er reached moderate performance, relative to competition leaderboard, our approach makes two important contributions. First, we show how explainable classi ers could provide meaningful predictions if results can be contextualized to test administrators who wish to intervene or take action. Second, our re-engineering of test scores enabled us to incorporate person ability into the estimation. However, ability was hardly predictive of efficient behavior, leading to the conclusion that the target label's validity needs to be questioned. The paper concludes with tools that are helpfulmore »for substantively meaningful log data mining.« less
  5. Kann, Maricel G (Ed.)
    Biological sequence families contain many sequences that are very similar to each other because they are related by evolution, so the strategy for splitting data into separate training and test sets is a nontrivial choice in benchmarking sequence analysis methods. A random split is insufficient because it will yield test sequences that are closely related or even identical to training sequences. Adapting ideas from independent set graph algorithms, we describe two new methods for splitting sequence data into dissimilar training and test sets. These algorithms input a sequence family and produce a split in which each test sequence is less than p % identical to any individual training sequence. These algorithms successfully split more families than a previous approach, enabling construction of more diverse benchmark datasets.