DeepMind AGI safety researcher Rohin Shah recently published an interesting paper on how agents can learn the wrong goal, described in this post: Goal Misgeneralisation: Why Correct Specifications Aren’t Enough For Correct Goals By Rohin Shah, Vikrant Varma, Ramana Kumar, Mary Phuong, Victoria Krakovna, Jonathan Uesato, and Zac Kenton. For more…deepmindsafetyresearch.medium.com Then @DavidSKrueger QT’d a related paper from his group “Goal Misgeneralization in Deep Reinforcement Learning” https://arxiv.org/abs/2105.14111 (Langosco et al 2021), with a similar focus.

NeurIPS 2018 in Montreal was my first experience with the NeurIPS series of conferences. I was there primarily for Deep RL workshop and as an entrant to the Pommerman competition. Overall it was a fantastic experience, and I met wonderful, brilliant people who are pushing the limits of RL in…

A common metaphor to explain exploit/explore tradeoff in bandit problems is that you are in a new town (say Montreal), and have tried 2 of the cafes so far. For your next meal, you could either: Exploit: go to a cafe you know you like, with reward=0.5, or Try a…

Recently at NeurIPS 2018 in Montreal, I witnessed Uber’s Jeff Clune present Go-Explore, their solution to Montezuma’s Revenge, the Atari game famous for posing a very difficult exploration problem to the current generation of RL algorithms. But Uber’s claim was met with controversy. What was the controversy? Afterwards at the poster session, I saw…

Before posting your question, first try getting to YES to all these questions: Did I post the exact error? If it is error related, don’t make people ask for this. Pasted error text is 10x better than screenshot (you can search easily). …

I originally shared this presentation with Vancouver’s Learn Data Science group on Nov 3 2017, which meets at the VentureLabs space in SFU’s Vancouver campus. This is a close look at a recent predictive modelling competition on Kaggle.com. The challenge was to predict grocery re-orders on Instacart.com, …

I love R and R markdown. But when I went to produce HTML email reports using Rmd using something simple like this: rmd="my_report" Rscript --vanilla -e "require(knitr); rmarkdown::render('$rmd.Rmd',params=list() )" cat $rmd.html | mail -a "From: fromemail@servername.com" -a "MIME-Version: 1.0" -a "Content-Type: text/html" -s "$subject" $recip …I ran into some roadblocks. …

Urmila Nadkarni, a friend and software engineer I used to work with at Microsoft, recently asked her network what career advice we wish we were given early on. This is my response, based on both my successes and my mistakes. …

Tension-driven. Distributed authority. Nested circles vs stacked pyramids. Overview Tactical meetings Links Holacracy Organizations are the most powerful force of change on the planet - yet they're held back by outdated operating models…www.holacracy.org Holacracy Holacracy is a social technology or system of organizational governance in which authority and decision-making are…en.wikipedia.org