Hillel

Summary includes advertisement for @NoRedInk, who is hiring, if you want to work in Elm, Haskell, and possibly Roc

Next up: @marielpettee, "Dancing With Myself". Mariel is a particle physicist, currently doing her postdoc at Lawrence Berkeley National Lab. She's here today to talk about her work as a performing artist and dancer, which she analyzed with motion capture and ML. #strangeloop

Problem: "I find myself falling into a creative rut, where I return to the same ideas over and over again." Wanted to use ML to generate new innovative movements still in her own style. Success metrics: authenticity, realism, surprise, "envie cinetique" (kinetic desire)

Long history of tech in dance. Many choreographies based on algorithms, randomness, or motion capture. Using neural nets is relatively recent, but by no means radical. Early 2017 work were based on 2D captures of movement, which is poor for modeling dances

Mariel was able to get full 3d captures by using a motion capture suit with 53 points, high enough resolution to even capture rolls. First project was "variational autoencoder". Information is "bottlenecked" to force NN to reconstruct input from a lower dimension. #strangeloop

NN was able to make diverse, realistic movements, but they weren't very interesting. Pushing to the fringes of the latent space gives incredibly weird movements. Showed us a demo. Model was better at generating variations of a movement that was known to be interesting.

They could take a real movement and add an amount of variation, and see what comes out of that. During the pandemic she used the model to make duets. While working at Intel AI, she "looked inward": using model to understand the relationships inside her body during dance

This used a "graph neural network", with NN nodes and edges between them. Edges belong to different types. Discussion of the overall architecture. Diff edge types are different relationships, like "left hand | upper body" #strangeloop

Underscored the difficulty of dance preservation, since there's no written language. The models feel like a way of "preserving creativity". Hoping to use this to do historical preservation. Also glitches are hilarious

Most excited about using models to critique her own creative practice.

Q: "First of all, great job nailing a strangeloop talk." [Totes true] "Would it be interesting to train on a more structured dance like ballroom?" A: Trained on just myself, modern & ballet, want to try other movement styles Q: Were any edges isolated movements? A: Feet

Q: [missed the question. Think it was "do people find this authentically captures dance?"] A: Scientists and engineers love this, dancers don't find it that interesting. ML can't capture the dance preservation

Q: Could this capture the accuracy of history? A: Could be useful for teaching choreography, maybe in a gamified way: here's the latent path, do movements that match it. Q: How does input space map to output space with the live feed? Is it just current position? A: Pose space

Q: Would it be more robust with absolute vs relative positioning? A: One of the prepocessing step was removing x-y positions so model can focus on just body movements, which gave better results, but lead to "ice skating" effect.

Closing keynote: @cristalopes, "The future of conferences." Brief history of modern conferences: come from the renaissance as a means of knowledge exchange. [being shared as a screen crawl] In 2019, business tourism was 1.3 trillion dollars #strangeloop

According to https://www.references.net/societies/history/ … there were 30 scholarly societies in 1323, and 230 by 1850. Conferences aggregate famous people and make them talk to each other. Conferences are a major part of many people's professional identity

Crista did her first conference in 1992. Since then, the WWW has exploded, which changes the value prop of conferences. Originally, it was about knowledge sharing. Nowadays the web makes that less important. Networking remains valuable, but less so— can meet easily online

Value 3: "paid breaks in nice places". Still valuable! [Couple of mine: marking out space, and creating *dialogue* beyond acquisition] Conferences have costs, too: time away from family, microharassment, CO2 emissions budget. Air travel has environmental impact #strangeloop

Crista estimates 330 tons of CO2 from her air travel over 3 decades. A tree stores 1 ton of CO2 over 40 years. She was on the ACM SIGPLAN Climate Change Committee from 2016-2020. Conclusion from committee: carbon tax!

We should voluntarily tax ourselves because governments won't do it [Govs don't do it because it's a democratic nonstarter. See Yellow Vest riots] So what's the carbon footprint of Strange Loop? No exact value, but we can estimate it

Carbon footprint is lowest for conferences in eastern US - western Europe. Pandemic made all conferences virtual for 18 months, so the SIGPLAN studied the virtual conferences. Some results: #strangeloop

1. [Prerecorded?] Video presentations do better for knowledge sharing, they have better production and fewer technical glitches. Also more accessibility, more reach 2. Livestreaming presentations have variable value. Big issue: timezones 3. Live Q&A can be better than in person

Virtual conferences are terrible for networking. Some ways to make it better, but at best it will be okayish "Paid breaks" benefit: "total fail" OTOH, you get to spend time w/ your family, fewer opportunities for harassment, vastly reduced CO2 emissions

Organization design decisions matter a lot to making a good virtual conference Case study: the OpenSimulator Community Conference (OSCC) was always virtual, even before the pandemic #strangeloop

The OSCC has a carbon footprint of about 750 kg per year Predictions: in-person events will not go away, look how eager we all were to get back to Strange Loop. But virtual participation will continue, so many conferences will be "hybrid", whatever that means

And that's the end of #strangeloop for today! I'll be out eating and juggling and stuff, see you all tomorrow