438 karma · joined February 4, 2016
We’ve built this index for autonomous driving datasets (https://scale.com/open-datasets) and are building that out for other domains right now.
Open source data has been a pillar to progress in ML (starting with ImageNet). It should continue to be the case that data that enables researches is sufficiently democratized.
One thing I’ll mention is that this is true both at the very early stages of a ML project, and even when an ML project is scaled up and in production. Oftentimes, the data pipeline is the true way in which a model will improve versus anything else, so it’s pretty critical that these data pipelines are setup to get an initial dataset but also to scale properly.
It’s one reason I started Scale (scale.com). It was viscerally clear that the real bottleneck to ML was getting the needed data, and in our case, annotating that data appropriately. It is very heartening to hear it echoed in this whole thread that data is very clearly what “matters” for ML.
2. Stereo depth estimation is quite unreliable in practice because it requires you to match up pixels between the two images very precisely (1-2px difference can be a large disparity in distance), so it is not reliably used.
It’s a small thing, but it’s surprising easy to spot once you look for it. And it really matters—startups are the business of building something from nothing. You need people who believe they can bend the earth.
Overcome many challenges, but per my last answer, building a team of the best people has been the most important and most challenging. That, and learning how to do sales ;)
Too many mentors. People in Silicon Valley are incredibly helpful. To name a few: Dan Levine, Mike Volpi, Nat Friedman, Adam D’Angelo, Ilya Sukhar, Jonathan Swanson, Albert Ni, Jeff Arnold, Charlie Cheever, and Drew Houston to name a few. I’m very very lucky.
Second, self-driving as a problem space will need labels for a very long time. In an application where (1) verifiable model performance is paramount, and (2) the models need to be extremely robust for cars to be safe, the need for labeled data is only magnified.
In the meantime, check out our open source datasets:
https://scale.com/open-datasets/nuscenes https://scale.com/open-datasets/pandaset https://level5.lyft.com/dataset/
Re 2—As with most companies working on ML these days, our stack is not fully proprietary. We don't take too strong an opinion on ML framework and use both Tensorflow and Pytorch currently. We generally use neural network architectures from the literature and then iterate on top of them to suit our unique problem requirements.
You can see some videos of what this looks like in this Twitter thread: https://twitter.com/BW/status/1158407524216909826
One difference from "Not Hotdog" is that our data is used to power the algorithms of other AI/ML companies like OpenAI, Waymo, Lyft, etc., so it's imperative that we have impeccable quality. That necessitates humans to ensure accuracy, particularly in safety-critical applications like self-driving cars.
I just wanted to chime in that we're a YC company as well (S16), and I'm thankful to the HN community for having been supportive through our whole journey.
This comment does not represent the company's viewpoint, and cardigan is not speaking on behalf of Scale.
We are very excited to have been able to work with Lyft in open-sourcing this dataset and advancing the research community. We are also very grateful to Lyft for choosing to leverage our point cloud viewer and have credited the annotations to us on their launch page.
In general, you probably should have at least something in your stack which reads all changes from your DB, at the very least for backup reasons.
We've worked with a number of self-driving companies like GM Cruise, nuTonomy, Voyage, Embark, and more to build high-quality training datasets quickly leveraging our API. Scale is the perfect platform for this work—we're focused on really high quality data produced by humans via API.
Solving problems like DropTurk with stronger guarantees around turnaround/quality/confidentiality is exactly what we do at Scale. Would love to hear more about DropTurk’s experiences around building this internally and how it compares to our product: