A Methodology to Identify Cognition Gaps in Visual Recognition Applications Based on Convolutional Neural Networks
This addresses the challenge of irrational behavior in CNNs for developers in safety-critical applications like autonomous driving, but it is incremental as it builds on existing adversarial and augmentation methods.
The paper tackles the problem of opaque cognitive behavior in convolutional neural networks (CNNs) for visual recognition, such as in autonomous driving, by presenting a methodology that uses adversarial search to generate worst-case images via augmentation techniques, identifying potential cognition gaps when performance drops on these images, as evaluated with AlexNet on driving scenarios.
Developing consistently well performing visual recognition applications based on convolutional neural networks, e.g. for autonomous driving, is very challenging. One of the obstacles during the development is the opaqueness of their cognitive behaviour. A considerable amount of literature has been published which describes irrational behaviour of trained CNNs showcasing gaps in their cognition. In this paper, a methodology is presented that creates worstcase images using image augmentation techniques. If the CNN's cognitive performance on such images is weak while the augmentation techniques are supposedly harmless, a potential gap in the cognition has been found. The presented worst-case image generator is using adversarial search approaches to efficiently identify the most challenging image. This is evaluated with the well-known AlexNet CNN using images depicting a typical driving scenario.