Yad Konrad



Path To Improved Sequence To Sequence Learning

Yad Faeq

Learning to learn experiments as a catalyst for reasearch.

My research interests relies in building a toolkit to faciliatate and allow researcher to move forward 10 steps ahead than what they were at 5 years ago, the ratio of steps to increase with every year.

The recent advancement in Aritifal Intellgence and Deep Learning is not evenly disturbuted among the scientific fields, meaning that there still are missing research work and results that cover hard sciences and computer science as well.

Chemists, Physicists and perhaps anyone who spends an amount of time in a lab still lacks the power of deep learning among their tool-bet. The real results of AI comes only when we all apply them, get results and start over for bigger goals.

My interest is inspired from the true brilliant work of Sustekver, Bahdanau and DeepMind researchers in this area. Seq2Seq is a great starting point, then we have the addition of attention mechaisim, from Bahdanu, that have been implemented in places such as [NMT papers].

In specific and brief, here are some of the fields that facinates me and still don’t have support of machine intellgence. Nanoscale laser and Microscopy, this rather a hard task to tackle, but the fact that still expensive scanners by scientists use (Delta time lost) over several expierements is dunting. The hands on research goals I have applied to building Meta-Agents that allow a researcher to control.

Research for Research, this is where the utlimate task of finding relevant work and refernces step in. As a researcher there are many risks of re-doing someone’s work, not finding enough resources and support in brief time. The artistic side of research is loosing inspiration to write and cover research topics. Having an assitant that allows researcher discover this is going to an amzing outcome.

Most junior researchers end up doing mundance, repetitve expeirments with or without supervision. This is where Deep RL agents can step in learning to learn expierements. Overtime it can go from learning as a meta-agent to a super-visor to junior researchers.