3 ms·
Yesterday here on HN there was a top thread about how a open dataset used for training self-driving cars is rife with missing labels and mislabeling. [1] Quite
by crazygringo 7y ago
Yesterday here on HN there was a top thread about how a open dataset used for training self-driving cars is rife with missing labels and mislabeling. [1]
Quite a few commenters insisted the issue was a non-issue -- that all datasets are noisy and mislabeling occurs all the time, that nobody's building actual self-driving cars from it, that surely if an object isn't labeled in one frame it'll be labeled in the next.
Now obviously we don't what the cause of this particular crash was.
But I will say that I found people's willingness to defend widespread sloppy labeling in training sets used for literal high-speed life and death situations rather shocking.
And that, hopefully, crashes like this serve to remind us of the far greater responsibility we have with regards to quality and accuracy when it comes to building ML models for self-driving cars, rather than when we're merely predicting how likely a credit card customer is to pay their next bill on time, or which ad is likely to be most profitable.
[1] https://news.ycombinator.com/item?id=22298882 https://news.ycombinator.com/item?id=22298882
- bagacrap 7y agoThe consensus was that the training set in question was not used for any real life situations. The only player in autonomous vehicles that doesn't understand the gravity of the technology is Tesla/Musk.