3IA PhD/Postdoc Seminar #11

Binarization for Optical Processing Units via REINFORCE

Abstract: Optical Processing Units (OPUs) are computing devices which perform random projections of input vectors by exploiting the physical phenomenon of scattering a light source through an opaque medium. The random projection operation is particularly useful when approximating kernel functions via random features. OPUs have successfully been proposed to carry out approximate kernel ridge regression with low power consumption by the means of optical random feature approximations. The main limitations on the generality of this approach are that OPUs require input vectors to be binary and that OPU projection matrices are unknown, and can only be retrieved through an expensive calibration procedure. The main difficulty to develop a solution is that the OPU projection matrices are unknown which poses a challenge in deriving a binarization approach in an end-to-end fashion. We propose a novel way to perform supervised data binarization inspired by a reinforcement learning approach and empirically demonstrate its effectiveness comparing to other binarization techniques.

Attention-based Multiple Instance Learning for Histopathology

Abstract: Although radiological images have long been used for computational analysis, the broad study of Whole Slide Images (WSIs), i.e., tissue samples, is much more recent. Histopathology, which is the analysis of said tissue, is the gold standard for tumor diagnosis. Therefore, much of the computational histology work to this day has focused on either cancer diagnosis or prognosis, looking for predictive biomarkers. However, WSIs display several characteristics which can be bothersome for the application of deep learning methods for instance. They are huge images, with billions of pixels, which no existing model can process directly. As a result, WSIs have to be split into images of smaller size, called tiles or patches, which can in turn be processed by the usual models. Theses patches are usually not labeled, whereas the slide is, which is why Multiple Instance Learning (MIL) and weak supervision are the main learning frameworks for computational histology. We will detail the working of one of these models: the attention-based deep MIL (Ilse et al., 2018), and discuss some ongoing improvements we are working on.

Open discussion on the two contributions