Neurips 2020 Review: Biology Edition

December 2020

Sabera Talukder Neurips 2020 Review Biology Neuroscience
  Last week I had an amazing time at NeurIPS 2020! Despite the need to be physically distanced, the flow of ideas though zoom, gather, and rocket chat intellectually connected us. Inspired by these new connections, I wanted to distill some of biology + machine learning thoughts at NeurIPS into one place. Let's jump straight in!

  Here's a roadmap of what I hope to discuss: I'll first highlight one talk in a little more depth, and dive into some of my outstanding questions. Then I'll share a curated a list of talks, orals, spotlights, expos, and workshops that I enjoyed in the spaces of Covid19, Healthcare, Biology, and Neuroscience & Neuroengineering. This list is of course incomplete, but hopefully will be a resource that you can revisit. Definetly check out the complete conference proceedings for more related work.

  If you are logged into your NeurIPS account you can click the links below and be taken straight to the recordings! If you aren't logged or don't have an account you'll get either a "Login" page or an "Access Denied" page. For those of you that make it to the end of this post, I have a surprise for you! It's cheating if you just scroll through though!

  During the Cognitive Neuroscience talk at Where Neuroscience Meets AI, and What's in Store for the Future, Kevin Miller discussed some common cognitive neuroscience themes and how they relate to machine learning. Specifically, he dove into the ideas of innateness, modularity, methods for studying cognition, vision in the dorsal and ventral streams, episodic memory, planning, cognitive neuroscience in rodents, and coputational cognitive neuroscience. It was a fully loaded talk!

  The first idea that intrigued me was the notion of the "intangible mind" (slide 51), perhaps because I am not a cognitive neuroscientist! I wonder how this framework of thinking can constructively be appiled in machine learning, and would love to read some papers that convincingly demonstrate this concept.

  Another interesting point that Kevin made in the Cognitive Neuroscience in Rodents section, was that because rodents perform complex tasks, like multistep planning, we can study their cognition. As an aside, perhaps by studying rodent behavior we can understand how they implement algorithms to make machine learning algorithms better.

  There is no doubt rodents can perform complex tasks, like multistep planning. If you haven't see the rats driving mini cars, stop everything you are doing and check out this link from the BBC. However, many complex behaviors have no ethological relevance in the animals trained to perform them. In other words, just because we can get "cognitive" behavior it doesn't make the behavior relevant for the animal in its natual environment or day to day life. This makes me wonder more generally if it is meaningful for us to study biology with the hope of improving machine learning if the biology isn't even optimized for the tasks we test it on. If you are interested in diving deeper, this tutorial's page is open sourced so anyone should be able to access it!

  One of the unique things about virtual NeurIPS, is that most all events were recorded! Hypotheically, for the first time ever, someone could watch all the talks, orals, spotlights, workshops, etc. if they were really determined. I'm not advocating we try it, but there were some talks that warrant a watch or a rewatch! Below I've linked to many of the talks I found interesting pertaining to Covid19, Biology, Healthcare, and Neuroscience & Neuroengineering. The talk with ★s on either end of the title denote one I detailed above. I hope you enjoy this curated list - happy learning!