Unsupervised estimation of domain applicability of models

Abstract With growing applications of machine learning across many domains, it is increasingly important to assess whether or not a given machine learning model is applicable to a domain. It is well-known that accuracy of a deployed model is highly dependent on the “similarity” of the deployed environment to the training environment where data was collected and pro- cessed. Present methods for assessing the applicability of a given model on a new dataset rely on the dataset having the “ground truth” labels. However, in many scenarios, especially those involving coalition edge operations, labeling the new datasets is costly or not feasible. As a result, model predictions based on anomalous data inputs may lead to erroneous predictions and faulty downstream decisions. This paper presents a novel unsupervised algorithm for detecting whether or not a given input to a deployed model is anomalous, without requiring the input to be labeled. Further, a practical approach for discovering new emerging classes in the edge environment is presented. The algorithm is evaluated on the MNIST character recognition dataset. The results show that the anomaly detection algorithm catches 93% of the anomalies when the model is deployed in a new domain. Further, only 1.5% of the normal inputs are falsely identified as anomalies. Via existing clustering techniques, we also show the effectiveness of our algorithm in discovering new classes from the anomalous inputs.
Authors
  • Nirmit Desai (IBM US)
  • Raghu Ganti (IBM US)
  • Heesung Kwon (ARL)
  • Ian Taylor (Cardiff)
  • Mudhakar Srivatsa (IBM US)
Date Sep-2018
Venue 2nd Annual Fall Meeting of the DAIS ITA, 2018
Variants