The Automated Model Evaluation (AutoEval) framework entertains the possibility of evaluating a trained machine learning model without resorting to a labeled testing set. Despite the promise and some decent results, the existing AutoEval methods heavily rely on computing distribution shifts between the unlabelled testing set and the training set. We believe this reliance on the training set becomes another obstacle in shipping this technology to real-world ML development. In this work, we propose Contrastive Automatic Model Evaluation (CAME), a novel AutoEval framework that is rid of involving training set in the loop. The core idea of CAME bases on a theoretical analysis which bonds the model performance with a contrastive loss. Further, with extensive empirical validation, we manage to set up a predictable relationship between the two, simply by deducing on the unlabeled/unseen testing set. The resulting framework CAME establishes a new SOTA results for AutoEval by surpassing prior work significantly.1

Multi-label Image Classification Systems (MICSs) developed based on Deep Neural Networks (DNNs) are extensively used in people's daily life. Currently, although there are a variety of approaches to test DNN-based systems, they typically rely on the internals of DNNs to design test cases, and do not take the core specification of MICS (i.e., correctly recognizing multiple objects in a given image) into account. In this paper, we propose ATOM, an automated and systematic black-box testing framework for testing MICS. Specifically, ATOM exploits the label combination as the testing adequacy criteria, hoping to systematically examine the impact of correlations between a fixed number of labels on the classification ability of MICS. Then, ATOM leverages image search engine and natural language processing to find test images that are not only common to the real-world, but also relevant to target label combinations. Finally, ATOM combines metamorphic testing and label information to realize test oracle identification, based on which the ability of MICS in classifying different label combinations is evaluated. To evaluate the effectiveness of ATOM, we have performed experiments on two popular datasets of MICS, VOC and COCO (each with five state-of-the-art DNN models), and one real-world photo tagging application from our industrial partner. The experimental results reveal that the performance of current DNN-based MICSs remains less satisfactory even in recognizing correlations between only two labels, as ATOM triggers a total number of 6,049 such label combination related errors for all MICSs studied. In particular, ATOM reports 587 error-revealing images for the industrial MICS, in which 92% of them are confirmed by the developers.