Is my Deep Learning Model Learning more than I want it to?

Naveen Panwar, Tarun Tater, Anush Sankaran, Senthil Mani

TL;DR: Can we trust our deep learning models? A framework to measure and improve a deep learning model's trust during training.

Abstract: Existing deep learning approaches for learning visual features tend to extract more information than what is required for the task at hand. From a privacy preservation perspective, the input visual information is not protected from the model; enabling the model to become more intelligent than it is trained to be. Existing approaches for suppressing additional task learning assume the presence of ground truth labels for the tasks to be suppressed during training time. In this research, we propose a three-fold novel contribution: (i) a novel metric to measure the trust score of a trained deep learning model, (ii) a model-agnostic solution framework for trust score improvement by suppressing all the unwanted tasks, and (iii) a simulated benchmark dataset, PreserveTask, having five different fundamental image classification tasks to study the generalization nature of models. In the first set of experiments, we measure and improve the trust scores of five popular deep learning models: VGG16, VGG19, Inception-v1, MobileNet, and DenseNet and demonstrate that Inception-v1 is having the lowest trust score. Additionally, we show results of our framework on color-MNIST dataset and practical applications of face attribute preservation in Diversity in Faces (DiF) and IMDB-Wiki dataset.

Code: https://github.com/dl-model-recommend/model-trust

Keywords: model trust, disentangled representation, colored mnist, face attribute preservation, new dataset