February 2015
Ivan Obolensky

Artificial Intelligence (AI) has come into the news recently as a result of comments from such luminaries as Stephen Hawking, who warned that AI could spell the end of the human race.¹

Nick Bostrom from Oxford University remarked about a hypothetical AI device whose goal is to make as many paperclips as possible. He thought it conceivable that it might consider humans as standing in the way of that goal and thereby might rationally decide to take out our entire species.²

Elon Musk, CEO of Tesla and SpaceX, went so far as to say that AI is potentially more dangerous than nukes.³

Should we be concerned?

After all it was Lord Kelvin, one of the greatest physicists of the late 19^th century who said that X-rays would prove to be a hoax, radio had no future, and that heavier-than-air flying machines were impossible. Closer to the present, the former Chairman of IBM, Thomas Watson, said in 1943 that the world market for computers amounted to five. ⁴

Future predictions even by those who we consider in-the-know have not always turned out in the way they thought. So why should AI be any different?

Granted, computers have come a long way since 1943, and those mentioned above I admire and are each brilliant in their fields.

Predicting the future is easy, but being correct is more difficult. Life has its own complexity and, in this, the future of computing and Artificial Intelligence may be no different.

One of the forks in the road that Artificial Intelligence faces is: will AI remain an extension of the human mind in the same way that a robot is an extension of what a human can do, only faster and more accurately? Or, will AI undergo a metamorphosis whereby machines are transformed into self-aware units that will not only surpass our abilities, but have the power to destroy its creators if they so choose?

Perhaps more challenging is the entire question of self-awareness. Given a certain level of complexity, are all networks above that specific threshold self-aware? Is the only reason we don’t know the answer because computers or networks haven’t reached that point? The consequences of such a threshold are profound. It might mean that all suitably complex networked entities, which would include most living things, would have self-awareness characteristics in varying amounts and with varying abilities to express that awareness. How should we view the animal kingdom if this is true, and how would we know it was? There is no language bridge so we can have that conversation.

One of the increasingly important subheadings of computer science is the field called Natural Language Processing. NLP is an area of study having to do with human-computer interactions. It has become more important as processing power has increased on the computer side of the human-computer divide, and humans have tried to make computers more useful to humans. In this we are not just referring to simple interfaces that enhance human senses such as implants and prosthetics, but real dialogue between humans and machines.

The bridge that spans the two is language.⁵

In the debate about whether humans will be able to build a super-intelligent self-aware entity the question remains: will such a creation remain a pliable and effective servant, or will it rise up in open rebellion against the humans? In either case, communication and language will play a fundamental part.

In the first instance, language may be the bridge to better coordination and collaboration between man and machine.

In the second, language might just be the single best tool to prevent an outcome where humans are no longer relevant to its creations. (After all the only way a super intelligent entity will be able to appreciate mankind’s sense of humor is through language, and just to be clear, the ability to tell a good joke has kept more than one human being alive when confronted with certain death. Of course, it is a last resort option but don’t knock it ‘til you’ve tried it.)

But even today, language is a puzzle.

In 1957, Chomsky published Syntactic Structures which started a new look at the subject. Since then a great deal of work has been done to understand what underlies language. Much of the later work is based on mathematical ideas as to how in abstract terms a compact finite structure can generate an infinity of results. The intention is to discover the simple computational rules that underlie language ability. For instance how can a few grammar rules, some sounds strung together into a modest vocabulary create an infinity of concepts that can be understood by one and all? It should be simple, but it isn’t. Much work needs to be done, but when these rules can be expressed and turned into code, given sufficient processing power, machines and humans can have an actual dialogue.

The alternative to computational language generation is that of statistical machine learning. In this, one simply gathers huge amounts of statistical data in terms of word order and frequency, and processes it with the result being some sort of translation and interpretation of what is trying to be expressed in both human and non-human terms. With the entire Worldwide Web available as a data set and massive processing power, this brute force method has potential but will likely prove limited.⁶

To understand this more clearly, simply record what you say for thirty minutes in a spontaneous conversation with someone else on a subject that needs a great deal of explanation. Listen to the result. The number of false starts, repetitions, “you knows”, other remarks, gestures, and mannerisms are embarrassing. How can a semi-rule-based, statistically-oriented machine cut through all that?

The truth is we may know what we want to say, only we can’t seem to talk that way.

This points to the heart of all questions concerning true Artificial Intelligence: Does the one doing the talking fully comprehend what it is saying? Can it hold a conversation with itself, as well as others, and make itself understood? Additionally, can it talk about what it just said?

Self-reference is one of the more tricky elements of language.

This not just a surface manifestation. A creature may look like a horse and act like a horse, but it may not necessarily be a horse. It could be a zebra, and just because somebody looks like they know what they’re doing, it is not conclusive proof that they actually know what they’re doing.

It is a thorny issue that comes with teaching all machines even those with advanced neural network programs.

There are two ways to go about getting a machine to behave and act as if it knows what it’s about:

The first is to program it based on a number of rules such as: If A, do B. Machines built in this way can be very powerful; particularly, if there are large numbers of options, and they can search for and follow this programming rapidly. The key point is there is no creative thinking. What to do has already been programmed in. This can even include an ability to react to fuzzy inputs. In a simple instance, the program might be able to utilize an object even if it had a rounded corner rather than a square one.⁷

The machine could mimic super-intelligence, but ultimately, it could only do what it was programmed to do. If this machine was programmed to take destructive actions against humans and other machines then it would be an extension of human programming. That this machine would be dangerous and potentially lethal there is no doubt. The result might be the same as a super AI unit deciding we are expendable, but the cause would ultimately rest with ourselves.

A second way of programming is with neural networks. These are mathematical constructs that mimic the behavior of layers of neural connections in the brain. They can be represented in a sense as several layers of little calculators which feed their answers forward to calculators in the next layer or back to calculators in an earlier layer.

These networks can be taught to recognize certain inputs and then respond according to what is presented. Speech recognition and handwriting recognition are some of the early uses for this type of programming. Visual recognition is now the area of emphasis.

Neural networks, just like synapses and neurons in the brain, must be trained.

Suppose one is trying to create a neural network that can predict the next number in a sequence. One feeds in the first value. Suppose the first number in the training set is 5. The first neuron takes that number and gives it a weight such as multiplying the 5 by 2.  If that number (10) is greater (or lesser) than the number it should be, the weight used in the next iteration is lessened (or increased) according to a rule. With each running of the program, feedback is used to modify the weights of all the neurons in a network and zero in on the correct value. The network is in effect learning through feedback.⁸

What each neuron does individually is simple. With modern computers there can be massive numbers of stored networks using thousands of neurons. Given enough layers, enough neurons, and enough training, a network can learn and recognize very subtle differences in handwriting, colors, almost anything where inputs and outputs can be translated into numbers, but there are limitations.

Once a network has been trained, it is given new data (a test). It processes this new data correctly, or not. Imagine there are hundreds of these networks. It is like choosing the best in the class. Only those networks that are correct with an accuracy greater than 95% on the new data are kept. The rest are discarded.

There are two key points here.

Firstly, learned behavior using neural networks even when thousands of them are pitted against each other results in few, if any, that are 100% accurate. Why? How many human students do you know that got 100% on every exam they ever took? Maybe that student only took one exam in his life, so how well do you think they would do on five more? In fact, you might find that given a lot of training and testing a really good network might be only 85% accurate in the task assigned.

Secondly, given a network that has many layers, how do we know what’s going on inside that network exactly? We don’t. Imagine a single chip that contains one million of these neurons. How we can tell exactly what all those weights are in each neuron and what the output will be given those weights? We can’t. It is just too many.

And as to that million-neuron chip, it already exists.

It is IBM’s new chip, code name: True North. The chip contains 5.4 billion transistors and is able to do 46 billion of the neural operations described above each second. It draws only a tiny amount of power and is supposed to emulate the processing power of a bee’s brain. True North’s forte is in facial recognition. Its intended use is to automate the process of looking at drone surveillance footage.⁹

Neural networks are a start in the development of a super AI entity. The problem remains however, that just because a network, or a machine, has been trained to do a certain thing, one is never sure that what has been learned is the thing the teacher taught. We do not have the language capability to find out.

I recall a story that circulated in the high tech world of target recognition software that used neural networks extensively. A missile firing program was supposed have learned to target the silhouettes of tanks which it seemed to do remarkably well. One day during a field test the array targeted the observers who were there to evaluate it. Apparently after a great deal of analysis, what the networks had learned was not tank silhouettes at all, but the angle of certain shadows the tank silhouettes made on the ground. The observer station was targeted because it had a similar shadow.

In the above we have an inkling of what might be the real issue with AI.

We can assume that the growth in sophistication and complexity of computers will continue exponentially. At some point, we will have to be able to communicate with the machines that we build using human language. To understand what a network, even a very sophisticated network is doing, or did, cannot only occur in some after-action report. It may be way too late for that. We need to know in real time not after the fact, but to do so we need to understand language sufficiently to be able to build machines with that language capability. True, they could lie. But then that isn’t so new either.

Chances are it will not be some evilly-intentioned, morally-challenged, super-intelligent machine that decides our fate, but what we didn’t know we didn’t know hidden in a network that didn’t know either.

 

---

1. Cellan-Jones, R. (2014) Stephen Hawking warns artificial intelligence could end mankind. BBC News. Retrieved February 12, 2015 from http://www.bbc.com/news/technology-30290540.
2. Leonard, A. (2014) Our weird robot apocalypse: How paper clips could bring about the end of the world If you dread a robot revolt, stop worrying about killer computers, and start worrying about … paper clips?. Salon. Retrieved February 12, 2015 from http://www.salon.com/2014/08/17/our_weird_robot_apocalypse_why_the_rise_of_the_machines_could_be_very_strange/
3. Vincent, J. (2014) Elon Musk says artificial intelligence is ‘more dangerous than nukes’. The Independent. Retrieved February 12, 2015 from http://www.independent.co.uk/life-style/gadgets-and-tech/elon-musk-says-artificial-intelligence-is-more-dangerous-than-nukes-9648732.html
4. Zimmer (N.D.) Retrieved February 12, 2015 from http://zimmer.csufresno.edu/~fringwal/stoopid.lis
5. Dyer, M. (2014) Superintelligence Threat vs Robot Society Threat (and an Oversight Concerning NLP). Retrieved February 12, 2015 from http://www.amazon.com/review/R3EGMJ28UD35YN/ref=cm_cr_dp_title?ie=UTF8&ASIN=0199678111&nodeID=283155&store=books
6. N.A. (N.D.) Natural Language Processing. Microsoft Research. Retrieved February 12, 2015 from http://research.microsoft.com/en-us/groups/nlp/.
7. Hall, J.S. (2007) Beyond AI: Creating the Conscience of the Machine. Amherst, NY: Prometheus Books
8. Cengiz, O. (N.D.) Celebi Tutorial: Neural Networks and Pattern Recognition Using MATLAB. Retrieved February 12, 2015 from http://www.byclb.com/TR/Tutorials/neural_networks/
9. Markoff, J. (2104) IBM Develops a New Chip That Functions Like a Brain. The New York Times. Retrieved February 12, 2015 from http://www.nytimes.com/2014/08/08/science/new-computer-chip-is-designed-to-work-like-the-brain.html?_r=0

---

If you would like to sign up for our monthly articles, please click here.

Interested in reprinting our articles? Please see our reprint requirements.

© 2015 Ivan Obolensky. All rights reserved. No part of this publication can be reproduced without the written permission from the author.