Flomenberg: Synthetic data gets self-driving cars 99.9% there but misses edge cases
Jake Flomenberg · 20VC: The Value Chain of Machine Learning, Is There Really An Incumbency Advantage in ML & Will The Rise In Cyber Remain For the Long Term with Jake Flomenberg, Partner @ Accel · Dec 6, 2017 · at 14:36
Jake Flomenberg, Partner at Accel, discusses the limitations of using simulated and automated synthetic data environments for training AI models in high-stakes applications like autonomous vehicles.
0:00 / 0:29exact quote · 29.9s
720p mp4 · rendered on demand · StarZero watermark
“In the instance of self-driving cars, It could make a lot of sense, and I think it can get you 99.9% of the way there, but the question is, how fully reflective is the underlying data set of the real world, and in the instance where it's not 100% fully reflected, can you actually generate all possible edge cases, and are you willing to accept that level of risk if there's point oh one percent chance that they didn't get this right? It's very, very different when we're playing a video game versus when a human life may actually be at stake.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Jake Flomenberg
Flomenberg: AI startups cannot build durable differentiation solely through algorithms
“A lot of the companies that I meet, they come up to me and they picked a spot that's very close to that algorithm dot on the top, and my concern with that is it's very hard to imagine an algorithm that's durably differentiable.”
Flomenberg: AI startups must build workflow software before building models
“It comes back to this notion of workflow first, data capture, followed by model building and iterations.”
Flomenberg: Prime vertical SaaS opportunities require zero initial data
“That's really where I think the bulk of the opportunity is going to be going forward, particularly in areas like vertical SaaS, where I can improve the current experience without any data, and any data that I can collect to improve your life is just gravy on t…”
Flomenberg: Data feedback loops provide stronger moats than initial dataset size
“That's where I think a lot of this feedback loop and sort of supervised learning around the data is Actually becomes a little bit more of an advantage than the original data set itself.”
Flomenberg: Google will compete with ad-data startups but unlikely cybersecurity ones
“If you want to use some Google-like or Google-specific dataset for an advertising product, it's a really uncomfortable situation because at any point in time, Google may very well decide to build that same product. Now, for instance, in security, could Google …”