Research

Over the years, I’ve had many interesting opportunities to perform (independent) research. Currently my field of interest is Machine Learning, whereas previously I did some work in Micro- and Nanoelectronics.

2018

Profit-Driven Machine Learning Models for Credit Scoring

The recently proposed expected maximum profit (EMP) measure for credit scoring allows to select the most profitable classifier. Taking the idea of the EMP one step further, it is desirable to integrate the measure into model construction, and thus obtain a profit maximizing model. The proposed implemented technique shows a significant improvement compared to regular maximum likelihood based logistic regression models on real-life data sets in terms of total profit, which is the ultimate goal for most businesses.

Reference: A. Devos, J. Dhondt, E. Stripling, B. Baesens, S. v. Broucke and G. Sukhatme, “Profit Maximizing Logistic Regression Modeling for Credit Scoring” IEEE Data Science Workshop (DSW), Lausanne, Switzerland, 2018, pp. 125-129.

2017

Humanoid Imitation Learning with GANs

At USC, we implemented a system which learns locomotion skills for humanoid skeletons from imitation, and all of the supporting infrastructure and data processing necessary to do so. We are extending previous work by DeepMind to imitate directly from video, using Generative Adversarial Imitation Learning.

Reference: Humanoid Imitation Learning From Diverse Sources

2016

Multiphase 34 GHz Oscillator for 5G

5G telecommunication hardware will make use of beamforming transceivers for high frequency (> 30 GHz) carriers. This creates a need for efficient multiphase frequency generators. I designed a digitally controlled multiphase oscillator in 90nm CMOS for this purpose, exploiting a distributed architecture, which had a better than state-of-the-art figure of merit in accurate simulation.

Reference: Devos A., Vigilante M., Reynaert P., “Multiphase Digitally Controlled Oscillator for Future 5G Phased Arrays in 90 nm CMOS.” in Proceedings of NORCAS, pp. 10-14, Copenhagen, 2016.

Low-Power Low-Area Transmitter for Neural Implants

When measuring neural brain signals, implantable devices require a battery which is often too big or low-power designs require a lot of chip area resulting in a very expensive product. I co-designed a low-power low-area transmitter enabling for more accurate and comfortable neural signal measurements.

Reference: K. Ture, A. Devos, F. Maloberti and C. Dehollain, “Area and Power Efficient Ultra-Wideband Transmitter Based On Active Inductor,” in IEEE Transactions on Circuits and Systems II: Express Briefs..

2015

Photosensitivity of Silicon Junctionless Transistors

Due to aggressive downscaling of conventional silicon MOSFET transistors problems arise between different doping regions known as short channel effects. As a side-effect of this scaling new devices such as the junctionless transistor (JLT) become functional. In simulation, I analyzed the photosensitivity features of this emerging device.

Reference: Devos A., Jazaeri F., Sallese J., “Photosensitivity of silicon based junctionless field effect transistors: a simulation study” (not published)