Don’t Get Tricked By Fake ‘Continual Learning’

A new holy grail has emerged in the AI industry in recent months: continual learning, or the promise of AI that’s able to learn from real-world experience the way humans do, without having to undergo long formal training processes that require tons of computational power and data.

And as you’d might expect with the rise of an ill-defined, buzzy term in the tech sector, investors are already telling me that they’re being pitched by startups claiming that they’ve solved the problem of how to get models to continually learn. 

“As continual learning has come into focus in recent months as one of the most important unsolved research problems in AI, it has caused a lot of startups to orient their narratives and pitches around addressing continual learning,” said Rob Toews, a partner at Radical Ventures. “More than once recently I have been pitched by startups who have claimed that they have ‘solved’ or are ‘near solving’ continual learning.”

There are a couple of techniques these startups are framing as ways to get models to do continual learning, even if they’re closer to clever workarounds than real research breakthroughs. One is to give models a scratchpad of sorts that they can take notes on to reference later. For instance, the model might write down the results of a sports game or instructions on how its user prefers it to sign off on emails.