Bernardino Romera-Paredes

Research Scientist at DeepMind

Motivation: To leverage the power of AI to bring light to important scientific problems.

My story, going backwards is as follows:

Since mid 2016, I am working in DeepMind as a research scientist

• Currently working on using RL for finding more efficient matrix multiplication algorithms.
• Previously working on AlphaFold for solving protein folding.
• When joining DeepMind I worked on applying deep nets for health purposes, like retinal disease and breast cancer screening.

Before that, during 2014 to 2016 I was in the Torr Vision Group as a Postdoc at the University of Oxford, researching about:

• Semantic segmentation and instance segmentation .
• The simplest ever approach to zero-shot learning.

During late 2010 to mid 2014 I did a PhD at University College London, where I researched multi-task learning and tensor decomposition methods applied to recognition tasks aimed at helping chronic low-back pain patients.

• Within this period, in 2013 I did an internship in Microsoft Research, Redmond, where I researched facial expression tracking based on head-mounted, inward looking cameras.

Before that, in 2009 I did the MSc Computational Statistics and Machine Learning at UCL, just after completing my bachelor degree in CS at the University of Murcia.

What I believe:

• Composability (hierarchies/structures) is not a property of nature/universe, but a property of our brain. Anything that cannot be explained by the composability framework becomes a Somebody Else's Problem field.
• There is a parallelism between the progress of AI and the childrens' book Zog: dragons (ML models), Zog (transformer), Madam Dragon (AI labs/big corps), Princess Pearl (scientific knowledge), Gadabout (humanity), and Gadabout's horse (human intellect).

The views here are not only not those of my employer, they are not even the opinions I held in the past or may hold in the future, as they tend to change with new evidence.

Collaborative long thorough outputs are available in my Google Scholar, whereas individual short cursory outputs can be seen in my Twitter.