There is No ‘Learn’

Zayd Enam at Stanford University has posted an accessible discussion of why developers struggle to perfect artificial intelligence systems. The key point is that patrons of AI development aren’t willing to build an algorithm and turn it loose on the world. They expect the algorithm to be trained against a set of representative inputs, and the responses evaluated to assure that the frequency and severity of improper responses poses tolerable risk.

This is not a new principle. Auto manufacturers model collisions to identify “worst case” scenarios where structural elements will interact to kill passengers that would normally have survived the initial collision. They balance the likelihood of these scenarios to produce the “safest car possible.” In most cases, auto safety systems (air bags, crumple zones) will save lives, but in some cases they will kill.

It’s not evil, it’s unavoidable.

The problem with AI is that it assumes control of decision making. In an auto accident, human beings make the decisions that result in the accident. The decisions are made with a rapidity that is perceptible to other drivers, who presumable can take action to protect themselves. This happens every day.

Of course, we’ve all seen a texting teenager drive through an intersection on the red when cars start moving with the green left-turn arrow. Situations like these – accidents generated by inattention or stupidity – are easily preventable by tireless digital components whose only job is to monitor for specific errors. If traffic signals had been installed as part of the internet of things, that could be done without artificial intelligence: the timing system could broadcast the signal state through the vehicle sensors, which would prevent the front car from moving. But since that system is not in place, engineers use AI to interpret camera pictures to determine the state of the lights. Obviously the AI algorithms must be at least equal to the judgment of a attentive human driver, which means that the correctness standard must be high.

But the motivation for the development of the AI systems is the inattentive teenager.

The more dangerous class of AI applications are those running in environments that humans cannot perceive at all. Obvious cases are industrial control (dangerous conditions) and electronic stock trading (high speed). The motivation here is profit, pure and simple. When an opportunity presents itself, the speed and precision of the response is paramount. Conversely, however, when the algorithm acts in error, that error is compounded more rapidly than humans can intervene.

Again, this is not new: in the 1700s, the British crown commissioned governors to manage its far-flung empire, and could control abuse of that authority only through the exchange of letters delivered by ships. In that situation, power was distributed and compartmentalized: the thirteen American colonies had governors and parliamentary bodies to resist executive misdeeds.

This is also the approach taken with training of natural learning systems: children. We don’t give children absolute authority over their lives. In fact, wise parents extend such authority only gradually as competency is demonstrated.

This is an approach to the problem of developing and deploying AI systems. No single system should be deployed on its own. Instead, they should be deployed in communities, with a managerial algorithm that polls the proposed actions and allows implementation only when consensus exists. The results are fed back into the training system, and the polling weighted towards the most effective algorithms. When a Newton or Einstein arises from the community – an AI system that always produces the best result – only then is absolute authority conferred.

Until the system changes. For example, a robot housekeeper may operate on high power until a baby is brought home, and then be forced back into low-power mode until it has adapted to the presence of the unpredictable element brought into its demesne.

To be Trolled, or Not to be Trolled

Trolling – posting a comment on a discussion forum for the sole purpose of creating hostility among the participants – is the most dangerous threat to the exchange of ideas on the web. The tech community is seeking to moderate the damage. Some new artificial intelligence engines profile accounts, others monitor for certain logical constructs (for example, any statements starting “You think…”).

Engadget reports that a site in Norway disallows a comment until the poster demonstrates knowledge of the main article. A script engine presents a multi-choice question generated from the article, and disallows comments until it is answered correctly.

What would be even better is if they would verify comment relevance against the original post and the previous comments in the thread. Given that a script is generating the questions, and a script is able to answer them, it seems that should be possible. Maybe that could be verified by a question posed to the poster?

Unfortunately, all of these solutions play into the hands of state-run trolls, such as the Russian “fake media” mills. By generating scripts that determine the correct answer, they can post far more efficiently than others, and thus come to dominate forum contents.

Here’s another option: build an AI engine that ranks the relevance of comments against the article and discussion, and allow readers to filter content against that ranking. That would enable those seeking serious discussion not just to be protected from trolls, but also to skip past comments that are just socializing. Offering “Good point!” is nice, but posted enough times and more substantial commentary falls off the bottom of the page.

Intelligence and Creativity

Joseph at Rationalizing the Universe explores the modern formulation, rooted in Goedel’s Theorem of logical incompleteness, of living with uncertainty. The following discussion ensued:


Brian

You point out correctly that Goedel’s theorem is restricted to a fairly narrow problem: proving that a proof system is “correct” – i.e. – that its axioms and operations are consistent. In other words, we can’t take a set of axioms and apply the operations to disprove any other axiom.

This seems to lead to the conclusion that we can’t trust our proofs of anything, which means that there are no guarantees that our expectations will be met. Unfortunately, expectations are undermined by many other problems, among them determination of initial conditions, noise and adaptation. The last is the special bete noir of sociology, as people will often violate social norms in order to assuage primitive drives.

At this point in my life, I am actually not at all troubled by these problems. Satisfaction is not found in knowing the truth, it is found in realizing creative possibilities. If we could use mathematics to optimize the outcome of social and economic systems, we would have no choices left. Life would become terribly boring. So what is interesting to me is to apply understanding of the world to imagine new possibilities. Mathematics is a useful tool in that process, particularly when dealing with dumb matter.

This brings me back to the beginning of the post: you state that “mathematics is the unspoken language of nature.” If there is anything that Goedel’s theorem disproves, it is precisely that statement. Mathematics is a tool, just as poetry and music are tools. At times, both of the latter have transported my mind to unseen vistas; mathematics has never had that effect.


Joseph

You raise a very interesting point; if we could optimise everything then would we take all of the joy out of being…. you may well be right. I know I get a lot of my satisfaction from the quest to know more. Although I disagree that Godel’s theorems disprove my original statement in this sense; language is essentially about describing things. That is why you can have different languages but they are easily translatable…. bread/pan/brot etc…. we all know what they mean because they all describe the same thing. In exactly the same way, mathematics describes things that actually exist; that isn’t to say nature is mathematics at all – mathematics is the language of nature but it is just as human in its construction as the spoken word. But is matter not matter because a human invented the label? Matter is matter.

To be, these theorems don’t break down all of our proofs; but what they do show is a vital point about logic. One which I think is going to become and increasingly big issue as the quest to understand and build artificial intelligence increases – can we every build a mind as intelligent as a humans when a human can know the answer to a non-programmable result? We hope so! Or rather I do – I do appreciate it’s not for everyone


Brian

I appreciate your enthusiasm, but I must caution that the mathematical analogies in classical physics cannot be extended in the same way to the quantum realm. Richard Feynman warned us that there is no coherent philosophy of quantum mechanics – it is just a mathematical formulation that produces accurate predictions. Ascribing physical analogies to the elements of the formulation has always caused confusion. An extreme example was found in the procedure of renormalization, in which observable physical properties such as mass and charge are produced as the finite ratio of divergent integrals.

Regarding human and digital intelligence: one of the desirable characteristics of digital electronics is its determinism. The behavior of transistor gates is rigidly predictable, as is the timing of clock signals that controls that propagation of signals through logic arrays. This makes the technology a powerful tool to us in implementing our intentions.

But true creativity does not arise from personal control, which only makes me loom bigger in the horizon of others’ lives, threatening (as the internet troll or Facebook post-oholic) to erase their sense of self. Rather, creativity in its deepest sense arises in relation, in the consensual intermingling of my uniqueness with the uniqueness of others.

Is that “intelligence?” Perhaps not – the concept itself is difficult to define, and I believe that it arises as a synthesis of more primitive mental capacities, just as consciousness does. But I doubt very much that Artificial Intelligence is capable of manifestations of creativity, because fundamentally it has no desires. It is a made thing, not a thing that has evolved out of a struggle, spanning billions of years, for realization. Our creativity arises out of factors over which we have no control: meeting a spouse-to-be, witnessing an accident, or suffering a debilitating disease. We have complex and subtle biochemical feedback systems which evolved to recognize and adjust to the opportunities and imperatives of living. We are a long way from being able to recreate that subtlety in digital form, and without those signals, meaningful relation cannot evolve, and thus creativity is still-born.

Google-Plex

When my sons learned about exponents in elementary school, they came home chattering about googol – which is 10100, or ‘1’ followed by 100 zeros. This is an impressively large number: physicists estimate that the entire universe only contains enough matter to make 1093 hydrogen atoms. However, as with their nine-year old peer that invented the terminology, the next step was even more exciting: the googolplex, or ‘1’ followed by a googol of zeros. When challenged to exhibit the number, the originator adopted a more flexible definition, purportedly:

‘1’ followed by writing zeroes until you get tired

One of the most attractive aspects of science and engineering research is just such child-like “what comes next?” The tendency is to think in exponential terms, although not always in powers of ten. Moore’s Law in semiconductors held that microprocessors should double in power every 18 months, and inspired a generation of chip designers, so that now our cell phones contain more processing power than the super-computers of the ’70s. Digital storage systems have improved even faster, allowing us to indulge the narcissism of the “selfie” culture: every day we store 50 times as much data in the cloud than is required to encode the entire Library of Congress.

The problem is when what appears to be the next obvious step in physics and engineering becomes decoupled from social need. We spend billions of dollars each year launching satellites to probe the structure of the cosmos, and even further billions providing power and material to the facilities that probe the matter that surrounds us. “Super-high” skyscrapers are pushing towards a kilometer in height, led by Saudi Arabia and Dubai, who appear to be engaged in a penis-envy contest that seems tolerable mostly because it side-steps the threat of mass extinction posed by the last such contest (the nuclear arms race between the US and the USSR).

In the case of software development, the “what next” syndrome is particularly seductive. It used to be that we’d have to go to a store and buy a box to get a better operating system, word processor or tax preparation package. Now we purchase them online and download them over the internet. This means that a software developer with a “better idea” can push it out to the world almost immediately. Sometimes the rest of us wish that wasn’t so – we call those experiences “zero day exploits,” and they cost us collectively tens of billions of dollars to defend against and clean up afterwards.

But most of the time, we benefit from that capability, particularly as artificial intelligence improves. We already have medical diagnostic software that does better than most physicians at matching symptoms to diseases. Existing collision avoidance algorithms allow us to sleep behind the wheel as long as a lane change isn’t required, and self-driving cars are only a few years away from wide-spread use. Credit card transaction monitoring protects us from fraud. These all function as well as they do because the rules described in software don’t disappear when the machine turns off. While the understanding encoded in the neural pathways of a Nobel Laureate vanishes upon death, software algorithms are explicitly described in a form that can be analyzed long after the originator retires. The lineage of successors can therefore improve endlessly upon the original.

The combination of robust, long-lived memory and algorithms means that the software industry believes that it is curating the evolution of a new form of mind. Those developing that mind seek to extend its capabilities in both directions: recording everything that happens, and defining rules to control outcomes in every possible situation. Their ambition, in effect, is to create God.

Contemplating this future, I had a colleague at work effuse that he “looked forward” to Big Brother, believing that in an era in which everything was known and available online for discovery, people would think twice about doing wrong.

In response, I suggested that many religions teach us that such a being already exists, but has the wisdom to understand that confronting us with our failings is not always the most productive course. In part, that is because error is intrinsic to learning. While love binds us together in an intimacy that allows us to solve problems together that we could never solve alone, we’re still going to hurt each other in the process. Big Brother can’t decide who we should love, and neither can God: each of us is unique, and part of life’s great adventure is finding that place in which we create greatest value for the community we nurture.

Furthermore, Big Brother is still a set of algorithms under centralized control. In George Orwell’s 1984, the elite used its control of those algorithms to subjugate the world. By contrast, the mind of God is a spiritual democracy: we chose and are accepted only by reciprocal gestalts.

Finally, Big Brother can never empathize with us. It can monitor our environment and our actions, it can even monitor our physiological state. But it cannot know whether our response is appropriate to the circumstances. It cannot be that still, quiet voice in our ear when our passion threatens to run amok. Big Brother cannot help us to overcome our weakness with strength – it can only punish us when we fail.

So, you in the artificial intelligence community, if you believe that you can create a substitute for God from digital technology, you should recognize that the challenge has subtleties that go beyond omniscience and perfect judgment. It includes the opportunity to engage in loving co-creation, and so to enter into possibilities that we can’t imagine, and therefore that are guaranteed to break any system of fixed rules. Your machine, if required to serve in that role, will unavoidable manifest a “Google-plex,” short for a “Google complex.”

Oh, Tay, Can You See?

Microsoft put up a speech-bot name ‘Tay’ on Twitter last week, and it took less than twenty-four hours for it to become a sexist Nazi. While labelled as “artificial intelligence,” Tay did not actually understand what it was saying – it merely parroted the speech of other users. On the /4chan/pol feed, that includes a lot of dialog that most of us would consider inappropriate.

What distresses is that Microsoft hoped to have Tay demonstrate the conversational skills of a typical teenager. Well, maybe it did!

In a recent dialog on the “liar Clinton,” I probed for specific proof, and received back the standard Fox News sound bites. When I described the Congressional hearings on Bengazi, the accuser had the grace to be chastened. This is typical of so much of our political dialog: people parrot sayings without availing themselves of access to the official forums in which real information is exchanged. The goal is to categorize people as “us” or “other,” with the goal of justifying arrangements for the distribution of power that benefit the “us.”

Donald Trump is a master of this political practice. Apparently his campaign doesn’t do any polling. He simply puts up posts on Facebook, and works the lines that people like into his speeches.

So I worry: did Microsoft actually succeed in its demonstration? Most American teenagers don’t understand the realities of the Holocaust or the difficulties of living under a totalitarian regime. In that experiential vacuum, do they actually evolve dialog in the same way that Tay did – with the simple goal of “fitting in?”

Somewhat more frightening is that Donald Trump appears to employ algorithms not too different from Tay’s. For God’s sake, this man could be president of the most powerful country in the world! He’s got to have more going on upstairs than a speech bot!

Fortunately, many teenagers, when brought into dialog regarding offensive speech, actually appreciate receiving a grounding in fact. You’d hope that our politicians would feel the same.

Artificers of Intelligence

The chess program on a cell phone can beat all but the best human players in the world. It does this by considering every possible move on the board, looking forward perhaps seven to ten turns. Using the balance of pieces on the board, the algorithm works back to the move most likely to yield an advantage as the game develops.

These algorithms are hugely expensive in energetic terms. The human brain solves the same problem in a far more efficient fashion. A human chess player understands that there are certain combinations of pieces that provide leverage over the opposing forces. As opportunities arise to create those configurations, they focus their attention on those pieces, largely ignoring the rest of the board. That means that the human player considers only a small sub-set of the moves considered by the average chess program.

This advantage is the target of recent research using computerized neural networks. A neural net is inspired by the structure of the human brain itself. Each digital “node” is a type of artificial neuron. The nodes are arranged in ranks. Each node receives input values from the nodes in the prior rank, and generates a signal to be processed by the neurons in the next rank. This models the web of dendrites used by a human neuron to receive stimulus and the axon by which it transmits the signal to the dendrites of other neurons.

In the case of the human neuron, activation of the synapse (the gap separating axon and dendrite) causes it to become more sensitive, particularly when that action is reinforced by positive signals from the rest of the body (increased energy and nutrients). In the computerized neural network, a mathematical formula is used to calculate the strength of the signal produced by a neuron. The effect of the received signals and the strength of the generated signal is controlled by parameters – often simple scaling factors – that can be adjusted, node by node, to tune the behavior of the network.

To train an artificial neural network, we proceed much as we would with a human child. We provide them experiences (a configuration of pieces on a chess board) and give feedback (a type of grade on the test) that evaluates their moves. For human players, that experience often comes from actual matches. To train a computerized neural network, many researchers draw upon the large databases of game play that have been established for study by human players. The encoding of the piece positions is provided to the network as “sensory input” (much as our eyes do when looking at a chess board), and the output is the new configuration. Using an evaluative function to determine the strength of each final position, the training program adjusts the scaling factors until the desired result (“winning the game”) is achieved as “often as possible.”

In the final configuration, the computerized neural network is far more efficient than its brute-force predecessors. But consider what is going on here: the energetic expenditure has merely been front-loaded. It took an enormous amount of energy to create the database used for the training, and to conduct the training itself. Furthermore, the training is not done just once, because a neural network that is too large does not stabilize its output (too much flexibility) and a network that is too small cannot span the possibilities of the game. Finding a successful network design is a process of trial-and-error controlled by human researchers, and until they get the design right, the training must be performed again and again on each iteration of the network.

But note that human chess experts engage in similar strategies. Sitting down at a chess board, the starting position allows an enormous number of possibilities, too many to contemplate. What happens is that the first few moves determine an “opening” that may run to ten or twenty moves performed almost by rote. These openings are studied and committed to memory by master players. They represent the aggregate wisdom of centuries of chess players about how to avoid crashing and burning early in the game. At the end of the game, when the pieces are whittled down, players employ “closings”, techniques for achieving checkmate that can be committed to memory. It is only in the middle of the game, in the actual cut-and-thrust of conflict, that much creative thinking is done.

So which of the “brains” is more intelligent: the computer network or the human brain? When my son was building a chess program in high school, I was impressed by the board and piece designs that he put together. They made playing the game more engaging. I began thinking that a freemium play strategy would be to add animations to the pieces. But what about if the players were able to change the rules themselves? For example, allow the queen to move as a knight for one turn. Or modify the game board itself: select a square and modify it to allow passage only on the diagonal or in one direction. I would assert that a human player would find this to be a real creative stimulus, while the neural network would just collapse in confusion. The training set didn’t include configurations with three knights on the board, or restrictions on moves.

This was the point I made when considering the mental faculties out at http://www.everdeepening.org. Intelligence is not determined by our ability to succeed under systems of fixed rules. Intelligence is the measure of our ability to adapt our behaviors when the rules change. In the case of the human mind, we recruit additional neurons to the problem. This is evident in the brains of blind people, in which the neurons of the visual cortex are repurposed for processing of other sensory input (touch, hearing and smell), allowing the blind to become far more “intelligent” decision makers when outcomes are determined by those qualities of our experience.

This discussion, involving a game without much concrete consequence, appears to be largely academic. But there have been situations in which this limitation of artificial intelligence have been enormously destructive. It turns out that the targeting systems of drones employ neural networks trained against radar and visual observations of friendly and enemy aircraft. Those drones have misidentified friendly aircraft in live-fire incidents, firing their air-to-air missile and destroying the target.

So proclamations by some that we are on the cusp of true artificial intelligence are, in my mind, a little overblown. What we are near is a shift in the power allocated to machines that operate with a fixed set of rules, away from biological mechanisms that adapt their thinking when they encounter unexpected conditions. That balance must be carefully managed, lest we find ourselves without the power to adapt.