Explainability for machine learning models has gained considerable attention within the research community given the importance of deploying more reliable machine-learning systems. In computer vision applications, generative counterfactual methods indicate how to perturb a model’s input to change its prediction, providing details about the model’s decision-making. Current methods tend to generate trivial counterfactuals about a model’s decisions, as they often suggest to exaggerate or remove the presence of the attribute being classified. For the machine learning practitioner, these types of counterfactuals offer little value, since they provide no new information about undesired model or data biases. In this work, we identify the problem of trivial counterfactual generation and we propose DiVE to alleviate it. DiVE learns a perturbation in a disentangled latent space that is constrained using a diversity-enforcing loss to uncover multiple valuable explanations about the model’s prediction. Further, we introduce a mechanism to prevent the model from producing trivial explanations. Experiments on CelebA and Synbols demonstrate that our model improves the success rate of producing high-quality valuable explanations when compared to previous state-of-the-art methods. Code is available at https://github.com/ElementAI/beyond-trivial-explanations.

MOTIVATION: Fluorescence imaging has become a commonplace for quantitatively measuring mRNA or protein expression in cells and tissues. However, such expression data are usually relative-absolute concentrations or molecular copy numbers are typically not known. While this is satisfactory for many applications, for certain kinds of quantitative network modeling and analysis of expression noise, absolute measures of expression are necessary. RESULTS: We propose two methods for estimating molecular copy numbers from single uncalibrated expression images of tissues. These methods rely on expression variability between cells, due either to steady-state fluctuations or unequal distribution of molecules during cell division, to make their estimates. We apply these methods to 152 protein fluorescence expression images of Drosophila melanogaster embryos during early development, generating copy number estimates for 14 genes in the segmentation network. We also analyze the effects of noise on our estimators and compare with empirical findings. Finally, we confirm an observation of Bar-Even et al., made in the much different setting of Saccharomyces cerevisiae, that steady-state expression variance tends to scale with mean expression. AVAILABILITY: The data are all drawn from FlyEx (explained within), and is available at http://flyex.ams.sunysb.edu/FlyEx/.