Towards Robust and Understandable Neuromorphic Systems

Description

Our research team, lead by Martial Mermillod and Sophie Achard is developping bio-inspired algorithms to improve the behaviour of deep neural networks (DNN).

Deep Neural Networks (DNN) have recently shown important limitations for different computational tasks. Our aim is (1) to use innovative mathematical algorithms to understand how DNN solve these computational problems and (2) to provide new architectures to address these limitations in a more efficient manner than the state-of-the-art.

About our research


Researching bio-inspired ways to improve the robustness of convolutional neural networks (CNN) to adversarial perturbations

CNN suffer from adversarial perturbations (Goodfellow et al., 2015), i.e. small shifts in input images which trick the CNN into misclassifying the image. One explanation of this problem could be found in the oversensitiy of these networks to texture information (Geirhos et al., 2019), and more generally, because the features these networks use to classify an image seem to strongly differ from humans.

Recent studies seem indeed to point out that CNN use a wide range of different features to classify images, many of them being irrelevant to humans, and possibly great targets to craft adversarial perturbations. (Ilyas et al., 2019). For example, high spatial frequencies (which include texture information) are vastly used by naturally trained CNN, to the point where images filtered to a degree where only the very high frequencies remain still remain correctly categorized by state of the art CNN on datasets such as ImageNet. (Yin et al., 2019).

Why are CNN fooled by these perturbations, while human vision remains untouched ? Psychological and neuroscientific experiments have shown that human vision system relies heavily on spatial frequency processing of visual information, as well as recurrent, proactive processing (Bar, 2007) of the visual input, where higher level brain areas guide the lower level vision processing zones toward recognition of an object or a scene.

We currently focus on two main perspectives following these leads namely (1) developping algorithms able to mimick the recurrent, spatial filtering processes which take place in the human brain, and(2) finding way to highlight the "non-robust" features of a CNN which are the most vulnerable to adversarial perturbations.