Xavier Boix
(Neuro)science of deep learning for
safe and fair intelligent learning machines.

Research Scientist @ MIT
Xavier Boix
(Neuro)science of deep learning
Current Research
We are addressing deep learning's lack of interpretability and data inefficiency, that is, requiring large amounts of training data, lack of robustness and poor generalization outside the training distribution.

To do so, we study deep learning with the lens of a neuroscientist by articulating hypotheses and testing them as if deep nets were another brain: theories are empirically-grounded, experiments and datasets are controlled, and individual neurons are at the core of the theories.

Group
Meet my fabulous group of students and postdocs:
Ece Özkan (postdoc)
Ian Mason (postdoc)
Anirban Sarkar (postdoc)
Amir Rahimi (postdoc)
Hojin Jang (postdoc)
Avi Cooper (research assistant)

The core values of my group and I are improving transparency, reproducibility and integrity of research. We are also consciously committed to equity and justice. More here.

Xavier's Bio
I currently work as a research scientist at MIT. I am grateful to have received training both in machine learning and neuroscience as a postdoc at MIT in the Sinha's lab and Poggio's lab, as well as the multidisciplinary NSF Center for Brains, Minds and Machines. I obtained a doctorate from ETH Zurich (2014) in computer vision and completed a postdoc at the National University of Singapore (2015).
Recent Projects
Our most recent work revolves around these two topics:
1) Understanding and steering the behaviour of deep nets through their individual neurons.
2) Understanding the principles of modular architectures to faciliate data efficiency.
When and How do DNNs Generalize to Out-of-distribution Category-viewpoint Combinations

1. Data diversity significantly improves OOD performance, but degrades in-distribution performance.
2. Separate architectures significantly outperform shared ones on OOD combinations, unlike in-distribution.
3. Neural specialization facilitates generalization to OOD combinations.
Opening the "black-box" of DNN-based fake new detectors

Our results show that the emergent DNNs' representations capture subtle but consistent differences in the language of fake and real news: signatures of exaggeration and other forms of rhetoric.
Publications and Code
Check out the full list of publications and code here:
Sponsors
Much gratitude to our sponsors and industrial partners -without them this research will have never been possible!
MIT 46-4079, 43 Vassar Street, MA 02139, Cambridge, USA

xboix@mit.edu


Made on
Tilda