Hacker Read top | best | new | newcomments | leaders | about | bookmarklet login

>In the past three years we have dropped error rates by a factor of 3

What are you referring to here?



sort by: page size:

I would welcome more recent data, but I doubt we are talking about a 4 orders of magnitude change to get to /year vs /hour error rates.

thats not the kind of progression you like to see - that is, error rates increasing over time instead of decreasing.

Not even xlose. This was a small incremental gain (maybe 10% relative reduction in errors). It's cool work but if you're an outsider this specific announcement shouldn't lead to any conclusions.

And this is close talk speech. The holy grail (20 feet away at a party with error rate below human) is decades away.


It would have been more honest to talk about how they increased their conversion rate from N% to M%, this we increased our good rate by Y% or decreased our error rate X% is the oldest trick in the book for academic papers (one that is likely to earn one a well deserved beat down by the PC).

> A nitpick, perhaps, but isn't that three orders of magnitude?

Perhaps the example was a best-case, and the usual improvement is about 10x. (That or 'order of magnitude' has gone the way of 'exponential' in popular use. I don't think I've noticed that elsewhere, though.)


it's an improvement (iirc, it makes the error-growth linear), but does not 'fix' the problem.

'Increased Error Rates' is a bit harsh, couldn't they call it 'Sub-prime Success Rates'?

So how would this show up in statistics since 2012? Fewer Errors?

I am talking about the accumulated error part near the end of the article, so your (uncharitable) assumption is not correct.

The whole point of the article is that they've (hopefully) improved the accuracy by using recent developments and combining different types of counts, and that method dramatically increases the count.

> is supposed to addrss that precise limitation, but the results show that it just doesn't do very well at all.

they demonstrate improvement from previous 5% to 7%.


So, if I understand you correctly, the average error decreases but the best error actually increases?

"It's not that their variance is decreasing it's just that their vectors are becoming more aligned." -- Ian Faith

Relative decrease in WER is not so significant for lower percentages. How about "we make 6 errors on 100 words but Kaldi makes 7".

“So, to create a more useful comparison point, we're excluding those three years from our long-term average.”

Sure, look at a few years while excluding ~1/4 of the data and I am convinced your analysis is just great.


> The breakthrough didn’t crack 80%, so three cheers for wide credibility intervals with error margin, but I expect the predictor might be feeling slightly more nervous now with one year left to go.

Good call on that by the author; according to this summary paper [1], a model reached ~79-93% accuracy on various Winograd data sets.

[1] https://arxiv.org/pdf/2004.13831.pdf


> What about the next question – how did the models do? Amazingly well. [...] This means that since 1992, the models have been within 3% of the measurements.

>or corrects 30% in a few days,

Didn't it just do that?


Wow! I wonder though how they account for over fitting of the data. Is it a real solution or a statistical anomaly? I ask because it seems that the progress over the last year(s) has been small increments within the 9-10% range.
next

Legal | privacy