DLRL Summer School

The Deep Learning Reinforcement Learning Summer School was fun and a great opportunity to learn from different perspectives. Some highlights of the panel by Rich Sutton and Yoshua Bengio, moderated by Martha White.

• All agree that doing a PhD should be fun. A PhD can still be fruitful even if you end up doing a job in industry.
• On reason for success: Determination, networks. Being motivated about the problem you work on is really important. RS says it’s important to be able to communicate (both write and speak) to audiences who do not know about the work.
• On research habits: YB recommends setting aside nonnegotiable time every week to read e.g. 10 hours. This happens naturally at the beginning of the PhD but tends to decrease as a graduate student takes on more projects over time. Read papers that are interesting. RS recommends keeping a research journal and writing it regularly, maybe a page a day. Read not just papers from the past year, but five years, ten years, even one hundred years earlier.
• On state of ML: RS reminds us that we are incredibly fortunate. We get to pick the best of many good options.
• On work-life balance: YB says it’s important but he doesn’t do a good job leading by example. Nor does RS. YB wishes he could make up for lost time with his kids. MW says that we often higher expectations for ourselves than others do. We can say no to things and still do a good job. YB says this is potentially an institutional problem (possibly due to conference system, you can’t just tell researchers to work 40 hour weeks).
• On ethics: YB recommends asking the question: “how does this affect the world?” regularly. RS reminds us that our actions send signals about what we believe to be right or wrong. Applying machine learning to areas such as education, medicine, ane the environment are all good. Panel unanimously agrees not to do something unethical. Sure someone else could do it, but you can only control your own actions.
• On ideas: YB says it’s good to take a step back and check to make sure that ideas make sense from perspective of learning in humans and animals. RS is happy to hear YB say that. Something that is not working is often an opportunity to learn more. MW echoes the sentiment that a failed algorithm often gives insight into the underlying problem that you care about. Talking through your ideas out loud is important. As a community, we should strive to be collegial and look for the best in each other’s ideas. We are also often harder on our own ideas because we are aware of the failure nodes.
• On publishing: Just too many papers to read. Even conference papers are not necessarily good. We need better systems for aggregating progres. Word of mouth and your network is the best way to get new research. YB says we have this unhealthy trend now of having six conferences a year with intense deadline pushes and small breaks in between. No clear way to fix and leads to short-term research. YB suggests that we could potentially have a journal system with quality reviewers, and accepted journal papers could present at the next coming conference. MW says perhaps systems for evaluation should focus on top 2-3 publications, rather than something like h-index.

Other misc.

Most intuitive explanation

<!–