3 ms·
The answer to this seems obvious: continuous training of live models. No more "cutoff dates": have a process to continually ingest new information and update we
by feoren 2y ago
The answer to this seems obvious: continuous training of live models. No more "cutoff dates": have a process to continually ingest new information and update weights in existing models, to push out a new version every week.
Note that I said "obvious", not "easy", because it certainly isn't. In fact it's basically an unsolved problem, and probably a fiendishly difficult one. It may involve more consensus-based approaches like mixture of experts where you cycle out older experts, things like that -- there are dozens of large problems to tackle with it. But if you want to solve this, that's where you should be looking.
- thatguysaguy 2y agoYeah I think every lab would love to do this and the field has been thinking about it forever. (Search lifelong learning or continual learning on Google Scholar). I don't think a technological solution is likely enough that we should pursue it instead of social solutions.
- carlio 2y agoWhile you might be able to continuously update the model, are you able to continuously update the moderation of it? As the article says, it takes time to tune it and filter it; if you allow any content in without some filtering of outputs you might end up with another Tay. You'd have to think the liability would slow down the ability to simply update on the fly. Also, if the proportion of training data available is larger for more established frameworks, then the ability of the model to answer usefully are necessarily dictated by the volume of content which is biased towards older frameworks. It might be possible with live updating to get something about NewLibX but it probably would be a less useful answer compared to asking about 10YearOldLibY
- zelphirkalt 2y agoModeration is the real reason it will be difficult to have online learning models in production. I think the technical side of how to do it will not be the biggest issue. The biggest one will be liability for the output.
- perrygeo 2y agoTalking to non-technical, but otherwise well-informed people, there is a broad assumption that AIs already "learn" as they're used for inference. IME people are surprised to find training and inference to be entirely separate processes. Human intelligence doesn't have such a stark distinction between learning and applied learning.