Hacker News Reader

Comment by visarga | Hacker News Reader

Direct link to paper: https://www.nature.com/articles/s41598-022-15245-z

> Using datasets from four different industries (healthcare operations, transportation, finance, and weather) and four standard machine learning models, we identify and describe the main temporal degradation patterns.

So they focus on shifting distributions and discover that models degrade in time, it should be an obvious thing, if you don't retrain on new data.

It's not about LLMs or diffusion models degrading. Time series prediction is a classical branch of ML with techniques like Moving Average (MA), Autoregressive Integrated Moving Average (ARIMA) or LSTM.

View on HN

Data drift is a problem regardless of the model/prediction task. Which is why models should be monitored in production, something everyone agrees is important, but nobody actually does.

It's a new era for "computer says no" types to revel in.

Sometimes though, "computer says no" is in fact the goal for some people. That some people dont realize so just demonstrates naivety.

It doesn't mean I have to accept that it's the computer's fault, though. Forcing them to take responsibility for their own decision is not naive, it's the counter to that strategy.

To be clear, I'm not saying that you specifically are naive. I do believe though that this naiveity exists here in tech amongst some proponents.

I strongly believe that for stable tasks not only models should be monitored in production, but there has to be a business process which takes any detected mistakes of the system and feeds them back as training data, creating a "human-in-the-loop" system even if the vast majority of the data is handled automatically - as these mistakes have a much stronger useful signal for updating the weights as any random sample.

it's obvious that they do, but exactly how they do it and when to retrain are critical for optimal execution. having a quantitative analysis like that is very interesting and useful.

Tangential, but I'm surprised to see a paper published in a Nature-affiliated journal that misuses the notation "dT" like that.

It's like they've combined 3 different conceptual-errors into one, where there're several obvious reasons to object to the notation, but no-one caught it.

---

Actually, is this an AI-written paper? Or maybe AI-edited?

It's got a strange number of errors and inconsistencies, and these errors seem biased toward superficial associations.

Although LLMs and diffusion models should also degrade over time as language evolves and users reference new phenomena in their prompts.

Didn't current LLMs already gas light their users over the current year? These models don't just degrade, they are already outdated and wrong the moment they get deployed.

I think a lot of the gaslighting wasn't because of outdated info it was just a limitation in the abilities to begin with.

It's interesting they chose those 4 domains, those are notoriously hard problems to begin with. And financial markets in particular self-correct themselves all the time. I'm pretty confident that whatever ML models they tried to test on those datasets would degrade over time. But title feels buzzword-ish for sure.

Reply on news.ycombinator.com