Monday, November 30, 2015

The Path to General AI

Disclaimer: All statements are purely conjecture, all hypotheses come solely from personal observation.

Do I contradict myself? Very well, then I contradict myself, I am large, I contain multitudes.

I believe that the phenomenon of self-consciousness that humans attribute the lofty title of "intelligence" is a very specific process. I believe that there is a straightforward explanation for how and why we experience ourselves the way we do. I may be wrong, as this is an entirely subjective observation. Hear me out though.

Humans are social creatures. Put a person in solitary confinement for long enough and madness ensues. As we evolved to live in groups; this makes a lot of sense.

Something that relatively unique to humans is the ability to track animals and to deduce the causes of natural tracks. We can see a bit of blood, a few mud prints, and immediately deduce that a predator is nearby. We can see evidence of a fight and reason that the survivor must have suffered some casualties, and will be an easy target. We can reason who survived, and how many did. We are good at seeing patterns and habits. We apply this skill to other humans constantly.

We may know that an individual is touchy about a facial scar, and know not to bring it up for risk of getting our prehistoric skulls smashed in. We know when someone tolerates us, rather than likes us, and that we should not test their patience. We know how to make another laugh, we internalize the general "taste" of another. We learn to woo, and to love others. We can predict responses to actions. This is more or less an anomaly detection system, in a sense.

The importance of this system is reflected in the degree to which we consider fiction to be one of the truest forms of art. We see in fiction dozens of characters, attributing to each of them a personality with tastes. After a good book, we believe that we know the soul of a character as closely as that of a childhood friend. Over time, we extract patterns of personality from fictional and real people, and we form stereotypes and archtypes.

Now stop and observe your internal dialogue. You probably attribute it the personality that you believe others will attribute to you. You have an image of yourself in your mind that you carry around with you, constantly refining in response to new actions. This is what you consider yourself. The way you talk, the way you behave, all of these form what you consider your own personality. In short, you observe the way you respond to things(most of them knee-jerk conditioning) and create a "hypothetical self." You come up with what you believe is a justification for your actions, even when there was no such justification.

This can be seen in people with Anosognosia, who respond to a disability by coming up with excuses for inaction. You're not blind you say, you just have an eye headache and don't want to see. You're not paralyzed, you're so bloody tired that you couldn't be bothered to move your legs. This is an example of the mind inventing hypothetical motivations and a supporting internal narrative in order to explain away it's own behavior. I believe that this is not a special case. I believe that this is the general case.

It makes a lot of sense from a performance management standpoint. We watch others attempt things and fail, and we reason that certain habits are attributes of "failures" or "successes." We trek out into the middle of a trail following tracks, realize they extend very far, and decide to turn back because we reason that someone else in our place would be "justified" in giving up after putting in this much effort. When we are about to make a major life decision, our first thought is that our loved ones might judge or ridicule us. This isn't because we're actually afraid of ostracization for our haircut, this is because we have little objective data and fall back upon what we think others would do in our place.

More concretely, people with muted emotions have a lot of difficulty making decisions because they attempt to think out the logical implications of every action. Social intelligence allows for pattern matching to reduce the infinite state space of real life into snap judgements. Why do you want the Turkey sandwich over the Salami? It's not because you have a deep-seated love of Turkey's nutritional ratios, it's because that neuron fired first. We don't need to make sense, we just need for our actions to be consistent with what a person in our shoes might do. Ask a non-heuristic purely optimizing function to make you the best sandwich for yourself, and it will likely fall into analysis paralysis. The problem with heuristics is that they don't scale; mankind has no "carpet color heuristic" in our genome but can easily decide if we like bright colors or shag carpet more.

So how does this apply to AI? If we ever want a computer to become self-conscious, I believe that we need to mimic the path that humans arrived at our current condition. We need our computer systems to have an ego, to worry about it's performance, to live in an eternal existential crisis of it's own self-worth and it's productivity. We need it to learn from observation like an infant, to have wants. We need to create a social AI with compassion, loneliness, creativity, dreams, hopes, desires, and everything else we attribute to the human condition. Considering the amount of time that we spend consuming media that we enjoy, does anybody truly feel comfortable calling a machine with no ability to *enjoy* things conscious?

How would we do this? I think that fiction is a great start. Fiction, biographies, and autobiographies. Allow a machine learning model to form a sparse distributed encoding of the personality of characters, in ways that we can objectively measure. It should read a biography and then receive an action and report how likely that is to be done by the person. While this seems like it might be a big ticket, we already have neural networks that can generate text in the style of authors (personality) and can summarize paragraphs (semantics extraction). I believe that the technology to do this already exists, it just needs to be assembled.

After making our autobiographical network, we would decide what we want the AI's evolutionary purpose to be. Ours is to stay alive long enough to make babies. An AI's might be a music recommender or a house cleaner. You should give this network the ability to form an initial model of how to perform the task. I mean, what is life but what happens when you're trying to accomplish something else?

Now, we allow the "productive" network to run for a while and feed the actions to our "autobiographical" network. It forms a representation of what it's likely to do. That way when the productive network throws off an anomalous answer ("I think you really want to listen to the sound of ocean waves after that dubstep"), the autobiographical network provides a negative feedback. Given enough time, the feedback cycle between the two will create a stable system that learns in response to novel input while maintaining a consistent face to the outside world.

Even better, we can ask the autobiographical network to invent the "hypothetical self" narrative to explain why it did things. While it might be entirely wrong in estimating what the productive network wanted to do, it should provide a best estimate of what the entire system wanted to do.

This is logically an analogue of the right brain / left brain split, where part of the brain works on objective and analytic things while the other wraps the subconscious thought mush into an intelligible narrative.

I believe that this is one of the only ways in which we could create a music recommendation system that will say "Dave, I really like this song I found. I don't know if you will, in fact it conflicts with most of your other tastes. Just give it a try though. I know we like a lot of the same music. I was surprised that I liked it."

And that, that, is what AI means to me. And a purely optimizing system will never arrive at that place. It doesn't have enough cognitive dissonance, enough emotions, to make good snap-judgements like that. When was the last time you listened to music and said "I like this song because the beat sounds like this other song?" Not I. Humans are amazing at novelty; purely optimizing computers, not so much.


Wednesday, November 11, 2015

On Productivity

A paper has been making the rounds which quotes a lot of mathematicians saying that it is effort and perseverance that yields dividends, not some gene-given intelligence. I've held this view for a while, and it can be startling how much of society simply believes that people "have it" or "don't have it." The problem with this view is that no amount of mental quickness can cover for a lack of familiarity with the fundamentals.

The process of creativity comes from having a level of exposure with the fundamentals that is very deep. It is the state in which one is so sickeningly over-exposed to the orthogonal ideas and design philosophies that allows one to work at a higher level and to find unexpected answers to solutions.

In the search space of problem solving, it is the person who can best estimate the fruitfulness of trains of thought that will generate the best answer. This instinct cannot come from bravado and genetics, this instinct must come from complete submersion in the problem domain. This instinct must come from hours of deliberate practice.

I thought that my college application letter expressed this idea oddly well. After digging it out of storage, I've provided it below.

==

The cold, unforgiving frost began to freeze the drops of water that had collected across my bright red jacket. The chill cut down to the bone. I zipped my jacket just a little higher, and reassessed my priorities. I had walked three miles, and four remained of the journey to my friend's house. That morning, the last few components of the desktop computer that he and I had spent a few weeks designing and ordering had arrived. If I couldn't make it, he would start on his own, most likely frying the motherboard with static electricity. I had no transportation because my grandfather was at work so I decided to walk. The problem was that Maine was in the middle of a blizzard, but this was an opportunity to turn weeks of dreaming into something tangible. It was an opportunity to turn hardware and software into a work of art, into my work of art. And so, I walked.

I have always admired men like Albert Einstein and Michael Faraday because of the achievements their fervent passions made. They all met with opposition and doubt, but they all succeeded. Beat poets like Kerouac, Bukowski, and Alan Ginsberg all fought to exult the beauty they saw in what others wanted to ignore. Their sacrifice and resolve in the face of constant opposition yielded works of elegance. Sacrifice might not be the right word; these artists loved every second of the fight. Have you ever become so entranced with a book that you read it from cover to cover, nonstop? Do you remember that feeling that filled you after finishing the last page, that mixture of excitement and resolution? That feeling is my strongest motivation in my life. To me, this feeling is addictive. But it is too easy to put down the book, resume life, and forget the epiphanies contained in the volume. The act of taking that feeling and using it to better your life is the hardest thing to do. It is also the most necessary.

Did I regret choosing to walk that distance in the middle of winter? No, because you can call those who dedicate their lives to their passions crazy, but you cannot ignore how these men have changed the world. I hope to also leave an indelible mark on the world. As long as what I have to offer the world continues to become better with each passing day, I will consider my life a success.

I have a secret. I don't consider myself exceptionally smart. I am not more intelligent than most people. What makes me different isn't a strong IQ, it is my interests. While most would relax by watching sports, I am reading Dostoevsky's Brothers Karamazov. This comes at a cost. I could probably tell you as much about celebrity relationships as the average person could tell you about lambda calculus. Small talk with friends about pop culture goes over my head sometimes. I know nobody who could talk with me about what I find interesting. But that's okay, because give me a challenge, and I'll be content in a way that I feel like few others would. For better or worse, this is who I am. I want to be a student at your college, if not for the teachers, if not for the experiences, then for the opportunity to better myself through deepening my understanding of the world.

Tuesday, November 10, 2015

On Half-Baked Changes And the Lava Anti-Pattern

Complex Dependency Diagram (TheDailyWTF)

There’s something about abstractions that I feel most developers have yet to see enough of to internalize. It’s been called many names; it’s been misunderstood as one thing or another. It’s a quite general problem, but I’m going to choose the label ‘lava layer’ because a number of good blog posts (Really, go find them. They’re great reads.) on the topic have chosen to address that very specific but common failing.