Hacker News new | ask | show | jobs
by robius 2900 days ago
There are many, just ask any unhappy scientists.

Funding goes to popularity, not new research. This is even true at DARPA where they ignore new technology because it doesn't fit some preconceived notion or don't have a framework to evaluate it.

Case in point for XAI, explainable artificial intelligence. The algorithms we use today give us black box models we can't interpret directly. So instead of fixing the algorithms, they focus on modeling the models and "guessing" which ones come close enough via simpler more intuitive stacks of models. Guesses upon guesses.

There has been research in new algorithms that generate open models where the weights make sense and are editable. There is one company working on this, but it's not nearly enough.

There's another set of research that has managed to convert black box models into open ones, giving full transparency.

Then there's asynchronous circuits research which do not require a clock. These can reduce power usage and boost efficiency on low power devices. Not much going on here.

There's one group building a RISC5 architecture with these, based on 30+ year old research with the inventor who still has not seen his life's work commercialized.

Then there's various types of imaging and tracking with signals we use every day, such as BT, Wi-Fi and Cellular among others, and being able to locate devices or people. You can find several universities doing this, none have made it commercially.

2 comments

Can you provide more references on XAI and converting current black-box models into more open ones?
The Economist had a relevant article a few months ago: https://www.economist.com/science-and-technology/2018/02/15/...
Google finds XAI easily: https://www.darpa.mil/program/explainable-artificial-intelli...

As for the conversions, I think these guys are working on a service or with early customers. https://optimizingmind.com

I actually think that's a pretty hot topic right now. Tons of people are working on commercializing it. Heck, a couple years ago I even did a contract with a major government where I helped them fix their recommender and it was built with explainability in mind because explainability allows politicians to understand the risk and into install human-review safeguards.
What group is building an asynchronous RISC5 architecture?
Karl Fant and company: http://www.theseusresearch.com/