jump to navigation

Bias and Truth and AI, Oh My October 4, 2017

Posted by Peter Varhol in Machine Learning, Software development, Technology and Culture.
Tags: ,
add a comment

I was just accepted to speak at the Toronto Machine Learning Summit next month, a circumstance that I never thought might happen.  I am not an academic researcher, after all, and while I have jumped back into machine learning after a hiatus of two decades, many more are fundamentally better at it than me.

The topic is Cognitive Bias in AI:  What Can Go Wrong?  It’s rather a follow-on from the presentations I’ve done on bias in software development and testing, but it doesn’t really fit into my usual conferences, so I attempted to cast my net into new waters.  For some reason, the Toronto folks said yes.

But it mostly means that I have to actually write the presentation.  And here is the rub:  We tend to believe that intelligent systems are always correct, and in those rare circumstances where they are not, it is simply the result of a software bug.

No.  A bug is a one-off error that can be corrected in code.  A bias is a systematic adjustment toward a predetermined conclusion that cannot be fixed with a code change.  At the very least the training data and machine learning architecture have to be re-thought.

And we have examples such as these:

If you’re not a white male, artificial intelligence’s use in healthcare could be dangerous.

When artificial intelligence judges a beauty contest, white people win.

But the fundamental question, as we pursue solutions across a wide range of applications, is:  Do we want human decisions, or do we want correct ones?  That’s not to say that all human decisions are incorrect, but only to point out that much of what we decide is colored by our bias.

I’m curious about what AI applications decide about this one.  Do we want to eliminate the bias, or do we want to reflect the values of the data we choose to use?  I hope the former, but the latter may win out, for a variety of reasons.

Advertisements

What Brought About our AI Revolution? July 22, 2017

Posted by Peter Varhol in Algorithms, Software development, Software platforms.
Tags: , , ,
add a comment

Circa 1990, I was a computer science graduate student, writing forward-chaining rules in Lisp for AI applications.  We had Symbolics Lisp workstations, but I did most of my coding on my Mac, using ExperList or the wonderful XLisp written by friend and colleague David Betz.

Lisp was convoluted to work with, and in general rules-based systems required that there was an expert available to develop the rules.  It turns out that it’s very difficult for any human expert to described in rules how they got a particular answer.  And those rules generally couldn’t take into account any data that might help it learn and refine over time.

As a result, most rules-based systems fell by the wayside.  While they could work for discrete problems where the steps to a conclusion were clearly defined, they weren’t very useful when the problem domain was ambiguous or there was no clear yes or no answer.

A couple of years later I moved on to working with neural networks.  Neural networks require data for training purposes.  These systems are made up of layered networks of equations (I used mostly fairly simple polynomial expressions, but sometimes the algorithms can get pretty sophisticated) that adapt based on known inputs and outputs.

Neural networks have the advantage of obtaining their expertise through the application of actual data.  However, due to the multiple layers of algorithms, it is usually impossible to determine how the system arrives at the answers it does.

Recently I presented on machine learning at the QUEST Conference in Chicago and at Expo:QA in Spain.  In interacting with the attendees, I realized something.  While some data scientists tend to use more complex algorithms today, the techniques involved in neural networks for machine learning are pretty much the same as they were when I was doing it, now 25 years ago.

So why are we having the explosion in machine learning, AI, and intelligent systems today?  When I was asked that question recently, I realized that there was only one possible answer.

Computing processing speeds continue to follow Moore’s Law (more or less), especially when we’re talking floating point SIMD/parallel processing operations.  Moore’s Law doesn’t directly relate to speed or performance, but there is a strong correlation.  And processors today are now fast enough to execute complex algorithms with data applied in parallel.  Some, like Nvidia, have wonderful GPUs that turn out to work very well with this type of problem.  Others, like Intel, have released an entire processor line dedicated to AI algorithms.

In other words, what has happened is that the hardware caught up to the software.  The software (and mathematical) techniques are fundamentally the same, but now the machine learning systems can run fast enough to actually be useful.

AI: Neural Nets Win, Functional Programming Loses October 4, 2016

Posted by Peter Varhol in Software development, Software platforms, Uncategorized.
Tags: , , ,
add a comment

Today, we might be considered to be in the heady early days of AI commercialization. We have pretty decent speech recognition, and pattern recognition in general.  We have engines that analyze big data and produce conclusions in real time.  We have recommendations engines; while not perfect, they seem to be to be profitable for ecommerce companies.  And we continue to hear the steady drumbeat of self-driving cars, if not today, then tomorrow.

I did graduate work in AI, in the late 1980s and early 1990s. In most universities at the time, this meant that you spent a lot of time writing Lisp code, that amazing language where everything is a function, and you could manipulate functions in strange and wonderful ways.  You might also play around a bit with Prolog, a streamlined logic language that made logic statements easy, and everything else hard.

Later, toward the end of my aborted pursuit of a doctorate, I discovered neural networks. These were not taught in most universities at the time.  If I were to hazard a guess as to why, I would say that they were both poorly understood and not worthy of serious research.  I used a commercial neural network package to build an algorithm for an electronic wind sensor, and it was actually not nearly as difficult as writing a program from scratch in Lisp.

I am long out of academia, so I can’t say what is happening there today. But in industry, it is clear that neural networks have become the AI approach of choice.  There are tradeoffs of course.  You will never understand the underlying logic of a neural network; ultimately, all you really know is that it works.

As for Lisp, although it is a beautiful language in many ways, I don’t know of anyone using it for commercial applications. Most neural network packages are in C/C++, or they generate C code.

I have a certain distrust of academia. I think it came into full bloom during my doctoral work, in the early 1990s, when a professor stated flatly to the class, “OSI will replace Ethernet in a few years, and when that happens, many of our network problems will be solved.”

Never happened, of course, and the problems were solved anyway, but this tells you what kind of bubble academics live in. We have a specification built by a committee of smart people, almost all academics, and of course it’s going to take over the world.  They failed to see the practical roadblocks involved.

And in AI, neural networks have clearly won the day, and while we can’t necessarily follow the exact chain of logic, they generally do a good job.

Update:  Rather than functional programming, I should have called the latter (traditional) AI technique rules-based.  We used Lisp to create rules that spelled up what to do with combinations of discrete rules.

What Are We Doing With AI and Machine Learning? February 12, 2016

Posted by Peter Varhol in Software development, Uncategorized.
Tags: , ,
add a comment

When I was in graduate school, I studied artificial intelligence (AI), as a means for enabling computers to make decisions and to identify images using symbolic computers and functional languages. It turned out that there were a number of things wrong with this approach, especially twenty-five years ago.  Computers weren’t fast enough, and we were attacking the wrong problems.

But necessity is the mother of invention. Today, AI and machine learning are being used in what is being called predictive analytics.  In a nutshell, it’s not enough to react to an application failure.  Applications are complex to diagnose and repair, and any downtime on a critical application costs money and could harm people.  Simply, we are no longer in a position to allow applications to fail.

Today we have the data and analysis available to measure baseline characteristics of an application, and look for trends in a continual, real-time analysis of that data.  We want to be able to predict if an application is beginning to fail.  And we can use the data to diagnose just what is failing.  In that the team can work on fixing it before something goes wrong.

What kind of data am I talking about?  Have you ever looked at Perfmon on your computer?  In a console window, simply type Perfmon at the C prompt.  You will find a tool that lets you collect and plot an amazing number of different system and application characteristics.  Common ones are CPU utilization, network traffic, disk transfers, and page faults, but there are literally hundreds more.

The is a Big Data sort of thing; a server farm can generate terrabytes of log and other health data every day.  It is also a DevOps initiative.  We need tools to be able to aggregate and analyze the data, and present it in a format understandable by humans (at the top level, usually a dashboard of some sort).

How does testing fit in?  Well, we’ve typically been very siloed – dev, test, ops, network, security, etc.  A key facet of DevOps is to get these silos working together as one team.  And that may mean that testing has responsibilities after deployment as well as before.  They may establish the health baseline during the testing process, and also be the ones to monitor that health during production.