The Art in Artificial Intelligence - Part 2

Part of the reason we decided to explore this topic on Technovation Talks was the claim made earlier this Summer that an AI had finally passed the Turing Test. So what's the Turing Test? It was a very basically described metric by which any sort of true machine intelligence might be assessed or otherwise verified. Here's the basic premise of the test - if an AI can engage in normal conversation with multiple human participants without the humans realizing that they were conversing with a machine (obviously it would be a remote conversation of some sort) - then the machine could be considered intelligent.

Alan Turing - BTW: a movie about him will be hitting theaters soon...

According to Turing's predictions in 1950, we should have already achieved this level of machine intelligence (by the end of the last century). Yet if you look at the story about this Summer's supposed triumph (which might be considered the first time it has in fact been achieved, there are nothing but problems and doubts):

• First off, the answers are screwy and it's clear that much of what the computer heard it misinterpreted. 
• Then they presented the AI as if it were an adolescent from war-torn Ukraine.
• And they also used the lowest possible threshold to gauge success - this threshold which represented a part of Turing's paper on the subject - suggested that success be declared if on average at least 30% of humans judging the AI would be fooled into thinking it was a human. So, the AI named Eugene, scored a 33% - but that is only because judges lowered the bar thinking he was a semi-illiterate teen. 

More important than all of this of course is the central question as to whether or not the metric or test is actually an accurate way to assess machine intelligence anyway? In a way, every system that has ever tried to compete in one of these tests to date has been purpose-built to pass the test. But does that make it intelligent (if it were actually to pass it)? The technology necessary for a machine to "think" through a conversation the way a human does simply does not exist - nor are we even close to understanding what that model would even look like. The systems trying to pass the Turing Test are simply conversational "hacks," in other words they include built-in tricks like responding to a question with a question or trying to work off of keyword cues. What's missing of course is any continuity of thought - any consciousness - and even the most simplistic conversation requires that. None of these systems can think and none of them can really learn. 

Now it may be that conversation hacking may become sophisticated enough in coming years so that many of these systems may actually pass the Turing Test threshold of 30% on a regular basis. But that test as it is now defined will never provide us with an accurate assessment as to whether a machine has in fact achieved some innate level of intelligence. There is no way to determine through the conversation if the system has "added value" to the topic rather than simply replied phrase by phrase in rather one-sided dialectics. It will be difficult to assess or acknowledge any growth or change. There is no expectation in a simple conversation to determine if you are in fact conversing with a self-aware entity.

In the movie Her, this guy falls in love with his operating system (and it didn't come from the Apple store!)

The first thing we need to do before we tackle how we might achieve AI is to determine what the appropriate assessment or validation for human-like intelligence really needs to be. We are going to suggest one and explain the rationale for it...

The Technovation AI Test -

AI Test Prerequisites / Expectations

• The Test is not meant to assess acquired knowledge per se, it is meant to assess cognitive ability. In other words, it is not about preparation or repetition of learned information, but is concerned with potential and / or application of any particular knowledge set.
• The Test does not have to occur in one sitting, but take place over any duration (within reason).
• The Test isn't merely concerned with correct answers or maturity in a point of time, but can also assess the ability to grow over time based upon responses to various aspects of the test (or other stimuli encountered within the time-frame of the test).
• The Test is not merely a linguistic exercise - the machine must not merely demonstrate the ability to communicate like a human, it must also demonstrate it can learn. 
• Foremost above all else though, the machine must demonstrate the one trait most closely associated human intelligence (as opposed to raw computing power) - it must demonstrate intuition. In this context, Intuition represents shorthand problem-solving (which we will discuss in much more depth in a future post). 
• On last aspect of the test that must be included is a review of the code to ensure that "conversational snippets" are not allowed to be prep-programmed. This implies that the majority of dialog is generated 'real time' by the machine. Now, that would not prevent the machine from reviewing logs of previously generated dialog (in some database), but that review could not lead to verbatim quoting -  rather must paraphrase or other restate previous points. 

The AI Test 

In a series of panel interviews, the AI must convince the judges or reviewers that it should be hired to perform a complex human role. The type of job and foundational knowledge can cover any number of topics but must be sufficiently complex to avoid "lowering the bar." (so, any job that requires a degree). Also, the interview style must be open (similar to essay tests in written assessments) - the answers must not just be correct, they must demonstrate value added insight from the intelligence conveying them. And the answers may be entirely subjective... (even better as long as the machine can rationalize them)

This test necessarily implies a very high threshold - perhaps in excess of a 90% rating for a very complex set of conversations. Why raise the bar this high? Simple - this is the one way we can force the development of a system that can both learn and apply that knowledge to problem solving and do it on the fly. To have human like intelligence, machines must have the ability to understand nuances of human communication and psychology - thus it must not only be able to interact, it must be able to convince us as well.  

Now that we have a more concrete target to aim for - how do we get there. In our next post, we'll delve into Learning - what works and what doesn't and how human and machine intelligence differ today.

Copyright 2014, Stephen Lahanas

#Semantech
#StephenLahanas
#TechnovationTalks

Read More

The "Art" in Artificial Intelligence - Part 1

Today, we are going to launch a problem-solving exercise on what might be the single most complex topic in Information Science - Artificial Intelligence. The goal here is not to provide any sort of comprehensive survey of current theory or practice; rather our journey begins with the stark realization of how little we've achieved in the field since the term was first coined 58 years ago. This is a problem statement and problem resolution exercise and an excellent case study in technology focused innovation. Let's begin...

We'll start at the beginning with some definitions and a review of key assumptions.

Artificial Intelligence, Defined 
The ability for a machine to consistently demonstrate core cognitive skills generally associated with human intelligence; including, learning, problem solving, intuitive reasoning and contextual memory retention and extraction. These skills generally imply the need to achieve some level of self-awareness.

What We Haven't Achieved, Yet
I said that this problem statement is focused around a lack of success in the AI field to date; let's try to quantify that first.

• No computer can learn like a human.
• No computer can speak like a human (this is deceptive, tools like Siri will provide responses back to you, but is that in fact anything like human speech? The processing that goes on within Siri is a relatively primitive form of pattern recognition as opposed to what even the least capable human mind can produce).
• No computer can handle complexity in the same manner a human can (this warrants much more explanation and we'll come back to it).
• No computer can problem-solve the same way humans can. (there are types of problem solving where computers are of course far superior to humans, yet even with all that power they still fail to solve relatively simple questions that humans can handle naturally).
• No computer has achieved anything coming close to consciousness or self awareness (despite the endless slew of sci-fi stories where this is a common fixture). 

Anybody trying to solve the AI problem is a mad scientist, right? 

Now, there's another ethical or moral side to this topic which we won't jump into until the end - the question as to whether we should even try to endow a machine with these traits - but then again it is likely that someone will do this regardless of the ethical objections. Part of the human learning process seems to require learning through our mistakes - a trait we may eventually end up passing to artificial entities, someday. But back to the problem.

Challenging Assumptions

As with most problem spaces, the set of initial assumptions associated with it tends to drive all else until or unless those assumptions evolve in some fashion. For Artificial Intelligence, there have been a number of assumptions that have helped define what's its become to date and also help to explain its limited success - they include the following:

• The notion that brute force computing power will eventually resolve many issues and push through various AI barriers. This is partially true, but then again we sent Apollo to the Moon with the computing power of a standard calculator (by today's standards), how much computing power do we really need to mimic human thought? Nature got us here through elegance, not waste.
• The notion that we fully understand how the human mind creates consciousness or exercises cognitive capability. We don't, yet.
• The very flawed notion that machine learning can or should have any connection to the current methods we use to teach each other. 
• The lack of sensory input associated with most AI paradigms. AI is IO dependent and that usually means keyboard and mouse although images and video (and audio) have now begun to play an important role. We'll get to this in more detail later.
• The notion that simulation can perform like the real thing; building to the simulation ensures a thing always remains a simulation (thus never achieving actual reproduction). This has some interesting implications which will eventually get us into a discussion of genetic engineering. 
• A lack of focus on natural language. This at first proved too difficult and now natural language does factor into much of the AI research going on. However, natural language hasn't been looked at as the core logic for an AI system - but it should be - instead we tend to view in terms of how an AI system (or other type of system for that matter) can interact with humans (or just capture human speech accurately). 

Watson actually showed us what he/she/it was thinking - if only we could
have seen into the minds of the humans, then we could have seen whether they think the same way...

A Brief Chronology of AI

• Ancient Greece - those happy go lucky philosophers in togas considered the notion of thinking machines. 
• About 150 years ago or longer - Artificial Intelligence (although not explicitly identified as such) becomes a popular topic in Science Fiction.
• 1930s - The golden age of Science Fiction novels includes more than a few stories about artificial brains and super robots.
• 1955 - The Term "Artificial Intelligence" is invented by John McCarthy and within a year the first conferences on the topic are held.
• 1968 - The Arthur C. Clarke novel 2001 becomes a box office hit and HAL, the emotionally unstable AI computer on board the Jupiter, becomes a celebrity. Take your stress pill, Dave.
• 1980 - Matthew Broderick makes it big playing games with an AI pentagon computer - the game is Thermonuclear War but of course you can't win it, right? 
• Early 80's - LISP is introduced.
• Mid-1980's - Expert Systems become popular.
• 1984 - Skynet becomes 'self aware' and Terminators start trying to kill John Conner.
• 1987 - The pasty-faced Commander Data steals the show on Star Trek the Next Generation.
• 1990's - An IBM computer wipes the floor with the world's chess masters.
• 2001 - Stanley Kubrick's last movie becomes Stephen Spielberg's tribute to him as Joel Haley Osment becomes one of the last humans playing a robot (as opposed to CGI animations with voice actors playing humans, robots and everything else).
• 2010 - Stanford opens up an AI online course to the general public, several hundred thousand sign up - few finish the course. 
• 2011 - Siri shows us that smart phones can indeed become somewhat intelligent.
• 2012 - IBM's Watson wipes the floor with the pantheon of Jeopardy champions.

Don't imprint unless you're serious...

The expectations for AI have thusfar radically outstripped the progress made. Now in the same 50 or so years we've taken the Star Trek communicator & tricorder fictions and made them reality in the form of smart phones (some of which you can now load with apps that when used with sensors can measure one's vital functions much like a tricorder).

A lot of smart people have tried for decades to make AI a reality and across the globe billions or hundreds of billions of dollars have been spent on related research. What's going wrong here? It can be only one of two possibilities:

1. Human like intelligence cannot be artificially created and maintained or
2. We've approached the problem wrong

I content that the second possibility is in fact what has happened. As we will progress with the series, it will become clear that the "Art" I referred to in the title is the ability to pick the right path for problem resolution.  In part two of this series, we will examine the question; How can machines learn?

Copyright 2014, Stephen Lahanas

#Semantech
#StephenLahanas
#TechnovationTalks

Read More