#27: Google's Teachable Machine 2.0, OpenAI's GPT-2 XL, and capturing solar energy with AI
Hey everyone! It was bound to happen at some point—this is the first Dynamically Typed I haven’t sent on time. But with a good reason: I was in Edinburgh last weekend to receive my MSc Artificial Intelligence degree, and I spent all my time there with family and friends. :)
Anyway, here’s what I’ve got for today’s edition. On the productized AI side, I’m covering Google’s updated Teachable Machine tool; for ML research, there’s an update to OpenAI’s staged GPT-2 release and a new reinforcement learning dataset with IKEA furniture. For climate change AI, there’s Heliogen’s computer vision-powered concentrated solar technology and CCAI’s startup webinar; and for cool stuff I found a project that’s using neural networks to “unlock” the contents of ancient Japanese documents. Let’s dive in!
Productized Artificial Intelligence 🔌
Teachable Machine can be used to classify images, sounds, and poses.
Google launched Teachable Machine 2.0, a web-based no-code tool for gathering data, training an AI model, testing and tweaking it, and deploying it to a website or app. It can classify images (from files or a webcam), sounds (recorded from a microphone, with file support coming soon), and poses (from files or a webcam). Here’s a few example projects from the Teachable Machine website:
Tiny Sorter: A DIY experiment connecting Arduino and Teachable Machine.
Project Euphonia: Steve Saling is using Teachable Machine to communicate in new ways, such as using facial gestures to trigger sounds.
Teachable Sorter: A physical machine that you can teach to rapidly recognize and sort objects using your own custom machine learning models.
- Kyle Phillips for Google’s The Keyword blog: Teachable Machine 2.0 makes AI easier for everyone
- Try it out yourself: Teachable Machine
Machine Learning Research 🎛
"[Full-size GPT-2] fed the opening line of neuromancer has created a novel premise i really want to read?” (@tambourine on Twitter.)
OpenAI has finished the staged release of GPT-2 with their publication of the 1.5 billion-parameter model. GPT-2 is OpenAI’s language model capable of generating text that convincingly looks like it’s written by humans. Out of fear for abuse, the lab has been releasing progressively bigger (and more capable) versions of the model slowly over the past nine months—see my coverage of the initial 117M release (DT #8); the 345M release (#13); and the 774M release (#22). OpenAI’s five big takeaways from the release were the following:
1. Humans find GPT-2 outputs convincing.
2. GPT-2 can be fine-tuned for misuse.
3. Detection is challenging.
4. We’ve seen no strong evidence of misuse so far.
5. We need standards for studying bias.
For details on each of these see OpenAI’s blog post from earlier this month.
In the weeks following the release, a lot of interesting work related to this full-size GPT-2 release popped up on Twitter. I’ve compiled a few of them below:
- Talk to Transformer, which uses GPT-2 to generate some text based on a user-specified seed sentence (see DT #13), has been updated with the 1.5B model. (Link to @tambourine’s tweet from above.)
- Hugging Face (an AI lab in NYC and Paris) has created a GPT-2 Output Detector Demo that predicts whether a piece of text was generated using the model. It classified something I generated using Talk to Transformer as 99.47% fake. Pretty good.
- Another detector, wrapped in a Firefox and Chrome extension. Code on GitHub: thesofakillers/GPTrue-or-False.
- Akshat Bubna wrote about fine-tuning GPT-2 for the Scale AI blog. Interesting read: How to label 1M Data Points/Week.
- Jay Alammar wrote a really good visual explainer on how GPT-2 works: The Illustrated GPT-2 (Visualizing Transformer Language Models).
Robots attempting to assemble IKEA furniture. (Lee et al.)
There’s a fun new reinforcement learning environment for IKEA furniture assembly:
The IKEA Furniture Assembly Environment is one of the first benchmarks for testing and accelerating the automation of complex manipulation tasks. The environment is designed to advance reinforcement learning from simple toy tasks to complex tasks requiring both long-term planning and sophisticated low-level control. Our environment supports over 80 different furniture models, Sawyer and Baxter robot simulation, and domain randomization.
Each task in the environment can be observed as a camera view of the scene, a depth view, a segmentation view, and a goal view of the assembled piece of furniture. A smaller beta version of the environment is currently available, with the full version set to launch next month. More:
- Paper by Lee et al. (2019) on arXiv: IKEA Furniture Assembly Environment for Long-Horizon Complex Manipulation Tasks
- Code on GitHub: clvrai/furniture
Quick ML resource links ⚡️ (see all 48)
- Cortex is an open source platform for deploying machine learning models—trained with nearly any framework—as production web services. GitHub link: cortexlabs/cortex
Artificial Intelligence for the Climate Crisis 🌍
One of Heliogen’s concentrated solar power plants.
Heliogen is using artificial intelligence to increase the efficiency of concentrated solar power. Unlike the more common solar PV setups, where each panel converts sunlight to energy, concentrated solar power uses arrays of mirrors to concentrate sunlight on a single small point. It then warms up a transport liquid to create a “super-heated system” that can be used to generate electricity or power industrial processes—like steel and cement production—that require higher temperatures than other renewable energy sources can provide.
Concentrated solar power itself is not a new technology, but it has previously been expensive to use because of how precisely the mirrors had to be manufactured and installed. Heliogen’s AI approach removes the need for this costly precision:
[To] address the calibration and control challenges of small heliostats, we’ve developed and patented an innovative computer vision closed-loop tracking control system. This allows heliostats to be installed and manufactured in a relatively imprecise manner because the control system can detect and correct their actual tracking position in real-time during operation.
Heliogen has commercialized this HelioMax technology into two products: HelioHeat (heat up to 1500°C for industrial processes) and HelioFuel (green hydrogen production). Their overall mission is to reduce our reliance on carbon-intensive energy sources by “replacing fuels with sunlight.” Read more about Heliogen here:
- Heliogen’s website: heliogen.com
- Heliogen’s blog post explaining concentrated solar and the role of computer vision in HelioMax: Concentrated Solar and its Role in Solving Climate Change
- Matt Egan for CNN Business: Secretive energy startup backed by Bill Gates achieves solar breakthrough
Climate Change AI released a start-up webinar co-sponsored by Better Ventures. They pitch it as:
A webinar for current and prospective entrepreneurs on commercializing innovations in the area of AI and climate change. Topics covered include funding milestones, early team formation, pitch deck structure, and fundraising best practices.
You can watch the full webinar on YouTube: Start-up Webinar with Better Ventures and CCAI.
Cool Things ✨
A page of Kuzushiji text transcribed by KuroNet.
Machine learning is helping unlock millions of ancient Japanese documents written in Kuzushiji:
From 800 until 1900 CE, Japan used a writing system called Kuzushiji, which was removed from the curriculum in 1900 when the elementary school education was reformed. Currently, the overwhelming majority of Japanese speakers cannot read texts which are more than 150 years old. The volume of these texts—comprised of over three million books in storage but only readable by a handful of specially-trained scholars—is staggering.
Alex Lamb wrote about his work with collaborators to create a KutoNet, a neural optical character recognition model to make these documents legible to more people. The post digs into the challenges involved with transcribing Kuzushiji—including the fact that most documents were created using artful woodblock printing—and also covers the Kaggle competition they launched with their dataset. It’s a cool read about the perhaps surprising application of AI to history. Check it out over at The Gradient: How Machine Learning Can Help Unlock the World of Ancient Japan.
Thanks for reading! As usual, you can let me know what you thought of today’s issue using the buttons below or by replying to this email. If you’re new here, check out the Dynamically Typed archives or subscribe below to get a new issues in your inbox every second Sunday.
If you enjoyed this issue of Dynamically Typed, why not forward it to a friend? It’s by far the best thing you can do to help me grow this newsletter. 👨🎓