Human Learning Needs in the Age of Machine Learning

Yesterday I experienced quite a coincidence that helped me crystallize my thoughts regarding machine learning; as I was scanning my Twitter and Google+ timelines and my Inoreader feeds each site had a post about machine learning. This post is about 2 common questions I get asked and how these posts helped me reflect on my responses to these questions.

Due to my work with the IBM Watson research team, I often get asked these 2 questions by other physicians:

•  The first question is if/when computers are going to make physicians redundant.  My response is that at present the vision is to create Artificial Intelligence systems that would help healthcare providers provide more efficient or better quality care and not to replace them.  Clearly, no one has a crystal ball and the expert opinions run the gamut from fear of AI to tremendous optimism.

From Newsweek

Peter Diamandis and Steven Kotler

• The second question is how to train our current generation of learners to prepare them for the future where they will work closely with artificial intelligence systems.  My response here is that students need to be part of building these systems in their disciplines.  Just like in high schools students take immersive experiences to learn foreign languages, our learners need to start doing electives in computer sciences and  data analytics departments or companies to leverage these technologies to solve problems in their disciplines.  Thus medical students should work with computer science students to use big data from wearable devices to improve health of a population for example.  This will not only help solve problems but they will learn first hand the limitations of these tools and recognize these in the future rather than just blindly relying on their recommendations.  

So what were the three posts?

1.  The Machine Vision Algorithm Beating Art Historians at Their Own Game

This is a post about a paper by researchers at Rutgers University who developed machine learning algorithms to help recognize styles and artistss of fine art paintings with a great deal of accuracy.  But the most important lesson was that when the algorithms failed to identify an artist correctly, there was a lesson to be learned.  There was a similarity in the paintings that was evidence of how the two artists (the correct one and the one wrongly identified by the algorithm) were similarly influenced.  Something that an art historian may not even have been aware of.

From: MIT Technology Review (Click image for web page)

2. Image Scaling Using Deep Convolutional Neural Networks

This amazing post describes in an (relatively speaking) easy to understand manner how a neural network was designed to process low resolution images to "fill in the missing pixels" to produce high-res images for Flipboard posts.  The post is a great one for our high school and college math students to see how concepts that they are learning have tremendous practical implications.

Courtesy Normal Tasfi via Flipboard

3.  a16z Podcast: Making Sense of Big Data, Machine Learning, and Deep Learning

This is a terrific 27 minute podcast the quotable quote being,

"Machine learning is to big data as human learning is to life experiences" 

It has a great story of how Larry Page talking with Google employees exhorts them to shorten the latency between entering a search term and getting the results.  When asked if what the final goal should be, whether it should be zero, he responds, "Why should we stop at zero?".  The goal should be that machines should be able to anticipate our needs not just respond to our requests.  Nguyen goes on to discuss why machine learning needs to be part of every app.

Sonal interviewing Christopher Nguyen CEO of Adatao

Given enough data, machine learning can identify patterns that humans cannot and will be able to predict problems before then happen.  Thus humans know that driving a car with poor brakes and wipers, very fast in rain on a curvy road often leads to disasters. 

We have an increasing torrent of data from medical literature, genomic analysis, electronic health records and wearable devices. When we start making this data available to appropriately programmed machines, patterns will emerge that may help predict or prevent heart attacks, strokes and cancers.  

This is what our learners need to get comfortable with so they know when to rely on these predictions and when to spot errors and learn from them.