Jul 15, 2017 · 21m · a16z
Chris Re
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Stanford Professor Christopher Ré presents DeepDive, an open-source dark data system that unifies data extraction, integration, and cleaning using probabilistic factor graphs and hardware-accelerated lock-free optimization.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 0.0 Guest teaching 0.0 Guest disagreement 0.2 The host pushing back 0.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 10:15 Calling publishers criminals
Chris playfully critiques publication paywalls by calling publishers criminals for legally restricting machine access to scientific data.
Hardest push from the host ▶ 8:30 Audience prompt checkIn a monologue talk with no host pushback, a brief audience check at 8:30 serves as the only non-guest utterance.
Biggest teaching moment ▶ 10:45 Human vs machine accuracy breakdownChris explains how volunteer annotators make systematic errors 16 percent of the time, demonstrating why probabilistic machine inference outperforms human manual data entry.
The host holds their own ▶ 0:08 Host absence throughout presentationBecause this episode is a solo academic talk without host dialogue or questioning, no host expertise display occurs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Introducing DeepDive & The Team | 0 | 0 | 0 | 0 | Chris Ré delivers an uninterrupted solo presentation introducing DeepDive and dark data systems. The host is entirely absent from the audio, requiring zeroed host scores while the guest maintains a collaborative academic tone. | |
| The Macroscopic Knowledge Challenge in Science | 0 | 0 | 0 | 0 | Chris outlines the macroscopic knowledge challenge in scientific literature across climate and biodiversity fields. The segment is a monologue lecture with no host participation or pushback. | |
| Inference Across Complex Text and Tables | 0 | 0 | 1 | 0 | Chris details Paleo Deep Dive table inference and compares machine accuracy against human volunteer annotators. He makes light jokes about publishers restricting data access, but no host dynamic is present. | |
| Broader Applications and DARPA MEMEX | 0 | 0 | 0 | 0 | The guest presents real-world applications of DeepDive, including DARPA MEMEX human trafficking detection and prostitution pricing models. The monologue continues without host intervention. | |
| Stochastic Gradient Descent and Going Hogwild! | 0 | 0 | 0 | 0 | Chris concludes his presentation explaining hardware parallelism, SGD, and lock-free execution in Hogwild! The talk finishes with audience applause without any host Q&A or discussion. |
Statements from this episode (12)
Automated dark data pipelines can outperform human data extraction quality
“There are a number of systems that have shown that it's actually possible to build these ETL pipelines, these pipelines that extract, transform, and load information with higher quality than humans in some very simple settings.”
Traditional data pipelines over-extract and over-clean due to unmeasurable impact
“So in contrast, if you think about the way that people build these extraction and integration and cleaning systems, they tend to over-extract, over-integrate, and over-clean Because they have no idea if their cleaning or extraction is actually gonna prove the …”
Probabilistic inference can reduce data pipeline engineering effort by orders of magnitude
“In contrast, we can build a pipeline first, which may not be very high quality, and then use probabilistic inference as a way to say, where should we spend our effort next? And the reduction in effort can sometimes be orders of magnitude, as we've seen in some…”
Modern scientific literature is universally accessible but impossible for humans to read
“For even the narrowest of scientific questions, They couldn't possibly read all the information that was relevant to them, or even a significant fraction of it. So this is why I say that scientific knowledge is accessible in a way like never before, but it's n…”
DeepDive factor graphs can reach hundreds of terabytes in size
“In some of our applications right now, these factor graphs become hundreds of terabytes in size.”
Human volunteers achieved 84% accuracy on PaleoDB data extraction tasks
“When we assess this, we found that the PaleoDB volunteers were accurate about 84% of the time.”
PaleoDeepDive matches or exceeds human data extraction accuracy across all tasks
“In some predicates, it's basically a wash. There's information that humans and machines are equally good at. But we're never any worse, and in some cases, we're substantially better.”
DeepDive human trafficking detection pilots are expanding to law enforcement agencies
“So, I'll skip over the details of the application, but I will say that it's actually in active use and we can actually get high precision values and we're in Extending, ah, pilots to more law enforcement agencies over the next couple months.”
Hardware parallelism enables orders of magnitude speedups over competing systems
“SIMD parallelism, multicore parallelism, NUMA, those are all things we need to take advantage of to go, in some cases, orders of magnitude faster than competitor systems on these tasks.”
Classical models of sequential evaluation can be completely thrown away
“That is, why it's a new trade-off space and classical trade-off space don't sort of suffice, and to do that, I'm going to show you one result which just illustrates the classical models of sequential evaluation you can completely throw away.”
Microsoft adopted Chris Ré's Hogwild algorithm for its Project Atom
“A while ago, last year, Microsoft released their Project Atom, and they had a great demo where they were showing they were releasing, you know, recognizing Shih Tzus with Windows phones. Using deep learning. And the reason I will love Microsoft forever is the …”
Relaxing hardware memory consistency yields up to 1000x speedups over competitors
“We've basically done a couple of parlor tricks, and those parlor tricks have allowed us to get, you know, 10, a hundred, a thousand times faster than competitor systems, because we're actually taking advantage of what the hardware is giving us.”