Project with Volvo Cars

Project description

This project consists of a student groups work with a problem provided by Volvo Cars. The considered problem is an open set recognition problem for neural networks (NN). Volvo has a NN that is supposed to get data from two distributions; In-distribution and Out-of-distribution. The goal is to separate correctly classified data points from the In-distribution, Inliers from the joint set of points from Out-of-distribution, (OOD) and Missclassified In-distribution, Outliers. The predictors for this classification problem is the softmax output from the NN. The problem is attended with the three methods: Density-based spatial clustering of applications with noise (DBSCAN), Mixed discriminant analysis (MDA) and Isolation forest (IF). The findings are that identifying OOD data points is hard to do with the utilized methods. However, the methods performs well in separating between Inliers and Missclassified In-distribution. It is concluded that the latter separation is possible due to well defined limits in feature space and that the first problem needs to be attended with methods that does not directly use distance in feature space as means for classification. It is proposed that further work investigates methods that use deeper layers from the NN or mappings of the feature space to enable identification of OOD data points.

Link to the report

Link to the presentation

Link to my code