Error-controlled DNN Interaction Detection

May 2021 - Present

Despite their outstanding empirical performance, deep neural networks (DNNs) have been mostly treated as black box models, preventing their adoption in many low-error-tolerant domains, such as healthcare, finance, robotics, and cybersecurity defense. In such contexts, interpreting DNN to reason about why certain decisions are made is crucial.

Inspired by existing works which interpret a DNN model by selecting a subset of explanatory features subject to a controlled error rate. We are interested in detecting higher-order information captured by a DNN model, i.e., interactions between features. Modelling the fact that features often have joint effects with other features is especially useful for scientific discoveries and hypothesis validation. For example, gene-gene, gene-environment, gene-drug and gene-disease interactions are key elements in explaining genetic mechanisms in biomedical applications.

This project aims at combining a recently descibed DNN feature interaction detection algorithm (NID) with the knockoff filter to produce error-controled DNN interaction prediction results.

Proposed DNN architecture for error-controlled interaction detection

As shown in the diagram above, our approch invovles adding a pairwise coupling layer (inspired by DeepPINK) to the DNN models that takes the original data and knockoff data as input. By computing the interaction scores for both original feature interacitons and knockoff feature interactions, we can use knockoff interactions as a negative control to filter out insignificant interactions and find cutoff for interaction predctions that achieves a desired False Discovery Rate (FDR).

Related Links: