Google AI’s Health division has published a new end-to-end lung cancer screening system. The model, published in Nature Medicine, takes a 3D CT scan as input and outputs a prediction of whether it is a potential cancer case; it does this very effectively:
When using a single CT scan for diagnosis, our model performed on par or better than the six radiologists. We detected five percent more cancer cases while reducing false-positive exams by more than 11 percent compared to unassisted radiologists in our study.
It’s very cool to see advanced ML technology being published outside of the usual AI conferences: besides the obvious scientific and medical value of this specific contribution, hopefully publications like this will help scientists in other fields understand more deeply what types of problems AI is good at tackling. More:
Also from Google: Translatotron, an end-to-end speech-to-speech translation model. Previous models for translating spoken text would take three steps: (1) automatic speech recognition to turn the spoken source sentence into text, (2) machine translation to turn translate into the target language, and (3) text-to-speech synthesis to “speak” it. Translatotron does this all in one go, improving translation speed, reducing compounding errors between steps, and making it easier to retain the voice of the original speaker in the translation (which is really, really cool). Read more here:
Rob Story wrote about Railyard, an API and job manager that payments processor Stripe uses for scalable machine learning. Stripe uses machine learning for everything from fraud detection to retrying failed credit card charges, and it has teams training hundreds of new models every day. Railyard streamlines this process by exposing an API to which data scientists can submit jobs from any ML framework. Running on top of Stripe’s Kubernetes cluster, Railyard then trains, evaluates and saves the model. This takes away the cognitive load of having to think about infrastructure, operations, model state, etc., so data scientists can focus just on building and testing their models. Story’s full post explains how they implemented this at Stripe:
Railyard: how we rapidly train machine learning models with Kubernetes.
The Wolfram Engine is now free for developers to encourage its use in production systems; this engine is what powers the Wolfram Language and
Wolfram|Alpha. I’ve always wanted to get more into the Wolfram Language because it has implementations of lots of machine learning models, plus an enormous set of built-in “computational knowledge” about the real world:
There are now altogether 5000+ functions in the language, covering everything from visualization to machine learning, numerics, image computation, geometry, higher math and natural language understanding—as well as lots of areas of real-world knowledge (geo, medical, cultural, engineering, scientific, etc.).
I never thought of a project that I could build end-to-end in the Wolfram Language ecosystem, but this release now makes it possible to run the Wolfram Engine on a server and call into it from other programming languages, opening up a world of possibilities. More on Stephen Wolfram’s blog:
Launching Today: Free Wolfram Engine for Developers.
Quick ML resource links ⚡️ (
see all)
- Philip Guo shared ten “research design patterns” for finding research project ideas in technology-related fields, plus what to watch out for when applying them. Link: Research Design Patterns
I’ve now also made a
Notion page with all the quick ML resource links I’ve shared so far, which I’ll keep updating as I send out new issues of Dynamically Typed. Check it out here:
Machine Learning Resources.