Making AI work with less data

In a Recode Podcast Interview, Kai-Fu Lee, author of “AI Superpowers” ventures a thought on why China may provide a more promising locale for entrepreneurs working on AI products and services. He argues Chinese consumers are more willing to part with personal data than their siblings in the US/Canada/Europe. Personal data is a source of vital “nutrition” for AI algorithms. The most promising examples of these are almost all built on Deep Learning.

But the EU and data privacy advocates here in the States make a compelling case for tightly managed personal data, via a growing set of restrictions on who/what/where touchpoints available to data collectors (Google, Facebook, Amazon, even Apple & Microsoft). The problems stemming from otherwise indiscriminate use of personal data are well documented.

Is the US/Canada/Europe caught in lose/lose scenario? China takes the clear lead in AI development and efforts to regulate how personal data is handled in the strongholds of data privacy champions still fail to work as advertised?

Not necessarily.

In a story titled “Algorithms That Learn with Less Data Could Expand AI’s Power”, Gary Marcus’ announcements at MIT Technology Review’s Emtech Digital Conference, 2016 are reported: “… [Professor Marcus’] XProp software requires significantly fewer examples than the dominant form of machine-learning software, known as deep learning, to learn a new visual task.”

“XProp” evolved from “Backprop”, the concept driving much of Professor Geoff Hinton’s foundational work on neural networks and, later, their next generation: Deep Learning systems.

Gary Marcus started a business called Geometric Intelligence to market products and services built on XProp. By December of 2016 Uber had purchased Geometric Intelligence. Darrell Etherington wrote a story for Tech Crunch about the accquisition. Mr. Etherington reports Uber’s AI team was also wanted to “ [look] at the other side of the problem; making systems smarter with limited input. That could be a huge help in quickly helping ramp the effectiveness of products at Uber that don’t necessarily have an equivalent data set to draw upon.” (quoted from

It should also be noted Professor Hinton told the Axios News Service back last year Backprop should be “thrown out” and AI developers should “start from scratch” working in an entirely different direction.

Building data “lean and mean” AI systems would actually be a win/win. Data collectors might be able to satiate their voracious data appetites with a lot less of our personal information and more efficient AI systems can move into production.

We can dream, can’t we?

Source: Deep Learning on Medium