Jul 28, 2017 · 22m · a16z
Michael Jordan
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
Professor Michael I. Jordan presents a framework for unifying computational thinking from computer science with inferential thinking from statistics to solve modern Big Data challenges. He demonstrates this synthesis through privacy-aware statistical inference and the Bag of Little Bootstraps, a highly scalable resampling algorithm.
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 7.3 Guest disagreement 3.5 The host pushing back 0.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 13:05 Rejection of standard Bayesian error bar usage
Jordan forcefully critiques standard CS practice, arguing that practitioners use Bayesian error bars blindly without knowing what priors or tail behavior should be.
Hardest push from the host ▶ 0:10 Absence of host pushbackThe host does not offer any pushback because this monologue transcript contains no host dialogue.
Biggest teaching moment ▶ 6:40 Differentiating computational execution from inferential thinkingJordan re-educates the audience on statistical principles, explaining that merely running machine learning algorithms on software does not constitute true inferential thinking.
The host holds their own ▶ 0:10 Absence of host expertise demonstrationThe host does not demonstrate expertise or intervene because the transcript consists entirely of a keynote presentation by the guest.
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 |
|---|---|---|---|---|---|---|
| Historical Perspective on the Big Data Phenomenon | 0 | 6 | 3 | 0 | In this monologue segment, Michael Jordan outlines the history of big data across physics, genomics, and modern tech, pointing out that current systems just build software and hope it works. Host-side scores are zero because the host does not speak. Jordan demonstrates deep domain history while mildly criticizing industry superficiality. | |
| The Engineering & Intellectual Challenges of Big Data | 0 | 7 | 4 | 0 | Jordan explains the mismatch between manager expectations and theoretical reality regarding statistical personalization and compute time constraints. He notes that industry is decades away from solving these fundamental trade-offs cleanly. Host scores remain zero due to host silence. | |
| Blending Computational Thinking and Inferential Thinking | 0 | 8 | 4 | 0 | Jordan breaks down the conceptual divide between computational thinking and inferential thinking, illustrating why differential privacy must be viewed through population-level inference. He dismisses simple database lookup perspectives as non-inferential. | |
| Minimax Privacy Rates and the Need for Frequentist Error Bars | 0 | 8 | 5 | 0 | Jordan challenges the suitability of classical Turing complexity for statistical risk and pushes back on the widespread reliance on Bayesian error bars with unexamined priors. He presents minimax privacy equations as a preferred framework. | |
| The Bag of Little Bootstraps (BLB) Framework | 0 | 8 | 3 | 0 | Jordan delivers a detailed technical tutorial on frequentist error bars and explains why standard bootstrap resampling fails at terabyte scale. He introduces the Bag of Little Bootstraps (BLB) architecture and gently clarifies an audience question regarding subsampling. | |
| BLB Performance Results and Presentation Conclusion | 0 | 7 | 2 | 0 | Jordan highlights empirical benchmark results from Amazon EC2 showing BLB outperforming the standard bootstrap by orders of magnitude in runtime and accuracy. He concludes his talk on a collaborative note. |
Statements from this episode (9)
Jordan says tech industry builds software without addressing underlying statistical issues
“We're building lots of software, hoping it works, and we're not really thinking about the harder intellectual issues at play.”
Jordan says personalization business models fail due to statistical limits
“A lot of these business models are failing. People actually can't personalize very well, and it's Because of statistical issues. You've got huge amounts of data about some people, and very little about lots of people, and you don't know how to transfer the sta…”
Jordan predicts principled personalized big data systems remain decades away
“So I think we're decades away from being able to do what this boss is asking us to do in some principle way. You can occasionally build a one-off system that does some of these things, but we're decades from having the real principles.”
Jordan notes core statistical decision theory ignores computational runtime
“If you look at core statistical theory, it's statistical decision theory, I teach it all the time, you never see the word run time.”
Jordan argues differential privacy is rarely analyzed for statistical inference
“Differential privacy has mostly not been thought about inferentially”
Jordan warns high-dimensional Bayesian inference is overly sensitive to unknown priors
“And a lot of times you have no idea what the prior should be. You don't know what the tails should be in particular. And you're in high dimensions, you really have no idea how the tail behavior should be. And the whole inference is highly sensitive to the tail…”
Jordan argues meaningful real-world decisions require statistical error bars
“And real life decisions that mean something, you need error bars. If they don't mean anything, you're just trying to serve customers and hope that everybody comes to your website, you know, I don't know, who cares? But in real life, you need error bars.”
Jordan explains traditional statistical bootstrap fails to scale on terabyte datasets
“Okay, but, gotcha, big gotcha, which is you can't do this on a terabyte of data, alright, because each resampling of the original data set on, if you have a terabyte, it's about 632 gigabytes. So you're sitting there on your terabyte of data at a central compu…”
Jordan claims Bag of Little Bootstraps vastly outperforms traditional bootstrap
“Here's the new algorithm, you know, again, implement on Spark. It's that little red box there. It took about a couple hundred seconds to get it, And the answer quality is better than the bootstrap after 15,000 seconds.”