Towards Explainability of UAV-Based Convolutional Neural Networks for Object Classification

f autonomous systems using trust and trustworthiness is the focus of Autonomy Teaming and TRAjectories for Complex Trusted Operational Reliability (ATTRACTOR), a new NASA Convergent Aeronautical Solutions (CAS) Project. One critical research element of ATTRACTOR is explainability of the decision-making across relevant subsystems of an autonomous system. The ability to explain why an autonomous system makes a decision is needed to establish a basis of trustworthiness to safely complete a mission. Convolutional Neural Networks (CNNs) are popular visual object classifiers that have achieved high levels of classification performances without clear insight into the mechanisms of the internal layers and features. To explore the explainability of the internal components of CNNs, we reviewed three feature visualization methods in a layer-by-layer approach using aviation related images as inputs. Our approach to this is to analyze the key components of a classification event in order to generate component labels for features of the classified image at different layers of depths. For example, an airplane has wings, engines, and landing gear. These could possibly be identified somewhere in the hidden layers from the classification and these descriptive labels could be provided to a human or machine teammate while conducting a shared mission and to engender trust. Each descriptive feature may also be decomposed to a combination of primitives such as shapes and lines. We expect that knowing the combination of shapes and parts that create a classification will enable trust in the system and insight into creating better structures for the CNN.