Inside | Real news, curated by real humans
Inside AI

Inside AI (Aug 19th, 2018)

Happy Sunday and welcome to the weekend edition of InsideAI.  I'm Rob May, CEO of Talla, and an active AI angel investor (52 companies and counting).  I also recently interviewed Steven Peltzman of Forrester on our AI at Work podcast, so be sure to check that out if you are interested in his take on AI in the Age of the Customer.

Below are the most popular links of the week:

Evan Patterson, a fellow at IBM Science for Social Good, led a team that is using machine learning to analyze computer code. Machine learning is often applied to images, video, audio, and natural language text, but this new system focuses on data science software. Patterson's technique captures the functions in the code — data transformation, analysis, modeling, and interpretation operations — and connects them to a data science ontology which presents a semantic flow graph representing the program. — PHYS.ORG

Two startups are using flash memory for AI edge processingIrvine, Calif.–based Syntiant and Austin, Texas–based Mythic are embedding flash memory to reduce the energy needed for deep-learning computation. Syntiant uses analog circuitry for smaller applications to identify text and speakers. Mythic is using a combination of analog circuitry and programmable digital circuitry for applications that need more complex networks for high resolution video in drones or smartphones. — IEEE SPECTRUM (Disclosure:  I'm an investor in Mythic)

AI technology is being used to level the playing field between small law firms with fewer associates and big firms with more staff. AI tools like Casetext for research, LawGeex to review contracts, Veritone to assess video and audio content for compliance purposes, Everlaw to evaluate documents, and Gavelytics for judicial analytics are helping to automate tasks that typically take a lot of manpower and time. These algorithms and services are also becoming more affordable as they become more ubiquitous. — ABOVE THE LAW

Google has announced more details about its new AI developer tools, which are now available in beta. The tools — AutoML Vision, AutoML Natural Language, and AutoML Translation — were first unveiled in July and are a part of Google's Cloud AutoML suite. According to the company, the new products offer customers more specialized machine-learning models. AutoML Vision allows developers to upload their own image datasets and create custom models, AutoML Natural Language helps developers build text analysis into their applications, and AutoML Translation leverages Google's translation functions. — EWEEK

A group of students from Fast.ai, an organization that offers free machine-learning courses online, created an AI algorithm that outperforms code written by Google researchers. Performance is measured by DAWNBench, a tool that tracks the speed of a deep-learning algorithm per dollar of computing power. Google previously held the top ranking, but Fast.ai trained an algorithm using the ImageNet database and AWS (Amazon Web Services) tools in just 18 minutes at the cost of $40 — 40 percent more efficient than Google's effort. "State-of-the-art results are not the exclusive domain of big companies," said Fast.ai co-founder Jeremy Howard. — TECHNOLOGY REVIEW

  • Email gray
  • Permalink gray

-- Commentary --

One of my favorite book's is Robert Pirsig's "Zen and the Art of Motorcycle Maintenance."  I first read it 20 years ago, and since then have read it 4 more times.  Every time, I take something a little bit different away from the book.  I was explaining this to a friend recently when I began to wonder how ideas, when consumed at different times, affect mental models, and whether this would affect how AIs are trained.

When you train machines at low level tasks like image identification, seeing a specific image for the first time as your 37th image or your 37,895th image doesn't really affect how you classify images.  But complex ideas are different.  There is a hierarchy to them.  As a simplistic example, if you learn about the concept of the invisible hand in economics, after you have already studied the biological concept of evolution by natural selection, you may think about economic equilibrium differently.  You may think of it as a temporary state that will shift again as underlying things in the world change, rather than as a long term position of stasis once you settle into the right equilibrium.  As you climb the idea hierarchy as a human, there is a "path dependence" to idea exposure where the ideas you consume at one point influence the ideas you consume later.  In our formative years, we call this teaching, and we know you need Algebra first to properly understand Calculus.  But later in life, or for concepts we don't teach in school, humans often show variety in how they think about things, based on this path dependence of learning and idea exposure.

My question then, is whether or not there will be a path dependence to training machine models as they advance into more hierarchical forms of intelligence?  This will be testable actually, in that we could have 1000 different AIs read a series of books in a different order to see if they then think about the world differently.  Or, maybe the fact that we humans are affected by this path dependence is a sign of a cognitive flaw - one that machines won't possess.  Perhaps they will assimilate information into their knowledge graphs of the world with more accuracy.

That's what I've been thinking about this week, so I wanted to share.  Thanks for reading, and see you all next Sunday.

@robmay

  • Email gray
  • Permalink gray

 

Many thanks to Inside AI's corporate supporters.  Please go check them out!

Leading global communication agency, advocating for the power of true. Read our recent AI report.

 
 
Ascent Reg Tech
At Ascent we use NLP in novel ways to help our customers to meet their compliance obligations. If you're interested in building new kinds of language models and infrastructure, check out our Data Science and Engineering positions or reach out to spencer@ascentregtech.com

If you were a corporate sponsor,
your logo would be here!

Learn more

Thoughtfully designed sit-stand desks, exceptional customer service. Try us risk-free for 30 days!

Subscribe to Inside AI

MORE NEWSLETTERS

The news, trends and tech that is reshaping the rapidly changing world of retail

Inside Retail

The news, trends and tech that is reshaping the rapidly changing world of retail

DAILY
The news, trends and tech that is reshaping the rapidly changing world of retail

Inside Retail

DAILY

SUBSCRIBED!

Share via

A hand-picked selection of products, deals, and ways to save money.

Inside Deals

A hand-picked selection of products, deals, and ways to save money.

DAILY
A hand-picked selection of products, deals, and ways to save money.

Inside Deals

DAILY

SUBSCRIBED!

Share via

David Strom's in-depth cybersecurity news and analysis

Inside Security

David Strom's in-depth cybersecurity news and analysis

DAILY
David Strom's in-depth cybersecurity news and analysis

Inside Security

DAILY

SUBSCRIBED!

Share via

The best source of in-depth news and analysis about Amazon

Inside Amazon

The best source of in-depth news and analysis about Amazon

DAILY
The best source of in-depth news and analysis about Amazon

Inside Amazon

DAILY

SUBSCRIBED!

Share via