Generative AI and incremental learning

Today’s generative AI is amazing and improving rapidly, but despite the hype it still faces major challenges:  a lack of incremental learning, a propensity for hallucinations (plausible guesses), it is easily distracted, and often weak on semantic consistency.

It is fruitless to try to directly compete with the large well-funded AI market leaders such as Open AI. Much better is to work on research challenges they aren’t addressing.  That includes incremental learning and deliberative reasoning.

Understanding and mimicking how humans learn incrementally is really tough and needs to be broken down into smaller more achievable challenges.  It looks like a more complex architecture will be required compared to that used for large language models (LLMs).

One of the hurdles for this work is to identify a means to bootstrap learning, as the more you know the easier it is to learn new things. LLMs learn about everything all at the same time. Incremental learning requires a more structured approach, including the means to try things out. Basic numeracy and simple math look like a promising domain for this as it minimises the dependency on common sense knowledge.

I am still thinking about some suitable initial steps. These could include work on single-shot learning to recognise sequences and to generalise across them, as well as preliminary work on sequential language processing and short term memory.

To ground this a little, imagine watching a drama series with subtitles.  If you step away and return minutes or even an hour or two later to a point before where you paused the video, you will recognise that you’ve already seen the subtitle text. Your ability to do so depends on the time interval and what you were doing in the interim. Your ability to generalise depends on grounding your understanding. This applies to learning a human language and likewise to learning basic math.

I hope to be able to report further progress over the next few months.

All the best,

Dave Raggett <dsr@w3.org>