Saturday, 3 March 2018

U of A Lecture – Demystifying Artificial Intelligence – Part 2



U of A Lecture – Demystifying Artificial Intelligence – Part 2

Part 2

I have split this blog into two parts, as it is over 2000 words in its entirety.  Here’s a link to Part 1:

Varieties of AI

The analytics problem solving stream goes something like this:

  • Description
  • Diagnostics
  • Prediction
  • Prescription

It is something like what a medical doctor does: look at symptoms, diagnose disease, predict course of disease, prescribe a treatment.  That general process works for a lot of problems in business (and life in general).  I think most people would place different types of AI at different points on this stream, with a heavy emphasis on Prediction.

A big division point in the spectrum of data science and AI methods is between Unsupervised and Supervised methods.

Unsupervised methods

Unsupervised methods include methods such as cluster analysis and social graphs.  Essentially, you have a lot of data and you are trying to see what patterns or commonalities exist in that data.  You might have a lot of data on customers, and want to group them into meaningful sets, to aid in marketing and such (e.g. early adapters, mainstream tech users, tech phobics, etc.).  Or you might have Facebook like data, and want to see who connects with whom in your network, who is influential, and so forth.  In scientific fields, that might help in categorizing galaxies or life forms.
I have done some work with these methods, especially K-means clustering, on such things as student surveys.  They can be interesting, but it can be difficult to persuade decision makers that they are all that useful.

Supervised methods

Supervised methods include methods most of those based on neural nets, commonly called machine learning, though Decision Trees would also fall into this category.  The idea here is that you have a lot of data as input, which is fed through various layers, using an algorithm rather similar to what neurons do in the brain, tweaking coefficients that weight inputs, so that the output comes closest to matching some “ground truth” that you have also fed into the algorithm.

Input layer => hidden layer => output layer
It is interesting to see the code for a simple feed-forward network.  These implementations can be done with surprisingly short programs.  However, there are several extensions and refinements of this basic technique, such as reinforcement learning, transfer learning and deep learning.  They obviously can become extremely complicated, using additional layers and various feedback mechanisms to increase the system’s problem solving ability.

A good example is the Alpha Go program, which beat the world’s best at an immensely difficult game of strategy.  As a computer, the program had the advantage of being fed much historical data on games that humans have played, then augmented that with millions of games that it played against itself.  This reinforcement learning could far outstrip what any human could learn, limited as we are by our inability to play more than a few tens of thousands of games in a human lifetime.

Such methods don’t always scale well or transfer well to different problems, though.  To do some might require a lot of programming maintenance and tweaking, which is expensive and requires human intervention, which defeats the purpose of AI, at least to some extent.

An Example, and the Notion of Explainable AI

A fairly simple example that I have used is a multilevel perceptron neural net, to predict the likelihood of graduate students completing their degrees.  A number of data elements pertaining to graduate student characteristics were fed into the system (age, gender, grades, subject, financing, etc.), which then came up with a prediction model that most closely fitted the actual graduation outcomes that were given the system, as a sort of training set.  That model could then be used to predict the likelihood of a student graduating, who was not in the training set.


This exploratory analysis was much like a logistic regression model that I had written to solve the same problem.  In my example, the two methods had almost exactly the same prediction success.  The machine learning model might have done better if I had fed it a lot more variables, and/or had a lot more cases.  It’s hard to say for sure, as I simply didn’t have more data to experiment with.  It was interesting that neither model actually improved very much on the decisions actually made by admissions people, overall.  In a sense, crowd-sourcing the decisions seemed to be about as effective as statistical or machine learning modelling.

Perhaps the most important difference between the machine learning technique and the older statistical technique (logistic regression) was that the machine learning method wasn’t very useful for explanation.  It could make accurate predictions, but it wasn’t possible to say which variables that were fed into the model were the important ones.  Old-style logistic regression, on the other hand, did give estimates of the relative importance of the variables.  That’s important for understanding a problem and therefore, understanding how to apply a solution to that problem.

This brings up the concept of “explainable AI”, that some governments and corporations are insisting upon.  In a sense, machine learning can be thought of as a form of pattern matching, or profiling.  This has obvious political and ethical implications.  A machine learning model might be excellent in predicting credit-worthiness, for example, but part of that excellence might be based on what would be considered illegal discrimination, if done by a human being.  Obviously, that’s a serious problem.

For example, in the graduate student problem, my logistic regression might show students from certain backgrounds do less well than students from other backgrounds.  That knowledge might be used to filter out those students, or alternatively to assist them to succeed (with extra funding, for example).  But, with a pure machine learning model, one would never know just why the system is making the predictions that it is, so these students would be filtered out, without anyone even knowing about it (unless humans noticed a problem, via their own pattern matching abilities).

Even in something as seemingly objective as self-driving vehicles, these problems could sneak in.  For example, a self-driving automobile AI might discover that young women were much less likely to j-walk as young men, and therefore give them less of a safety margin.



Work is being done on explainable AI.  In fact, my SPSS multi-level perceptron routine had a “relevance factor” option, that seemed to go some way towards this, for my graduate student model.  But, this could take a long time, and never be perfect.  After all, we don’t really know why our own human neural nets give us the usually reliable goal directed behaviours that they do.    

Conclusion

Some final comments:


  • Machine learning will become more and more common.
  • Actually implementing it in a business or industry is always a form of Research and Development.
  • It takes time and money to:
  • o   define your problem.
  • o   know your data, and prepare it to help solve the problem.
  • o   define your solution.
  • You will need to maintain the system.
  •  AI should be thought of as helping human experts, not replacing them.
  • Unless, of course, Skynet becomes self-aware, in which case all bets are off.

----------------------------------------------------------------------------------------------------------------
Now that you have read of some cutting edge science, you should consider reading some Science Fiction.  How about a short story, set in the Arctic, with some alien and/or paranormal aspects.  Only 99 cents on Amazon.

The Magnetic Anomaly: A Science Fiction Story

“A geophysical crew went into the Canadian north. There were some regrettable accidents among a few ex-military who had become geophysical contractors after their service in the forces. A young man and young woman went temporarily mad from the stress of seeing that. They imagined things, terrible things. But both are known to have vivid imaginations; we have childhood records to verify that. It was all very sad. That’s the official story.”




“The Zoo Hypothesis”, an Alien Invasion Story

Here’s a story giving a possible scenario for the so-called Zoo Hypothesis, known in Star Trek lore as the Prime Directive.  It’s an explanation sometimes given to account for a mystery in the Search for Intelligent Life, known as The Great Silence, or Fermi’s Paradox.

Basically, Enrico Fermi argued (quite convincingly, to many observers), that there had been ample time for an alien intelligence to colonize the galaxy since its formation, so where are they?  The Zoo Hypotheses says that they are out there, but have cordoned off the Earth from contact, until we are sufficiently evolved or culturally advanced to handle the impact of alien contact.

This story takes a humorous tongue in cheek approach to that explanation.  It also features dogs and sly references to Star Trek.  Talk about man’s Best Friend.

Amazon Canada: https://www.amazon.ca/dp/B076RR1PGD


No comments:

Post a Comment