Intelligence Emerges From A Logical Ground

Photo is Enlightened by Jonatan Pie via Unsplash

We have recently had a powerful illustration of how an intelligent agent can emerge, spontaneously, from a logical ground. Which is to say, it can emerge de novo from a prior tabula rasa state.  The preexisting system is defined solely by its logical rules.

In the following long-form article, I also try to map out how this relates to my broader naturalistic metaphysics. In that it clarifies the relationship between the logical cosmos and the minds that evolve within it.

AlphaGo Masters Go

It doesn’t seem that long ago that AlphaGo, the Google DeepMind AI, demolished one of the greatest, human Go players, when it defeated world champion Lee Sedol in March 2016 in Seoul. Wikipedia has in depth coverage of the event, and there are a huge number of online articles about it.  The following Nature video gives a sense of how it happened:

Video is The computer that mastered Go via Nature Video/YouTube

But AlphaGo, in its original version, was still dependent on humans to train the AI in its basic game-skills. DeepMind trained its neural networks – by feeding in a database of game-play from 160,000 expert games – using supervised deep learning algorithms. So in a sense you could say that AlphaGo learned the basic strategies of the game, as any good human player might, by studying expert human players.

Only then was AlphaGo let loose, and allowed to play an additional 30 million games, against another instance of itself. In this way it was able to refine its knowledge of the game, to a superhuman level, using the algorithmic technique known as unsupervised reinforcement learning.

Early versions of AlphaGo then underwent several additional iterations, as it took on other players. Chinese Go-champion Ke Jie’s comments, in the following YouTube video, are particularly worthy of note. Remember he was beaten by an even more powerful version of AlphaGo than Lee Sedol:

Video is GO champ Ke Jie talks about his match with Google’s AI via YouTube

The Incredible Lightness of AlphaGo Zero

However, now we have AlphaGo Zero. It starts from a tabula rasa state. At the beginning, it knows nothing, other than the most-simple, basic rules of Go. Then it just starts to play against itself – initially randomly. It uses the latest, and most-powerful machine-learning algorithms. Also the latest hardware – but notably – considerably less hardware than last year’s AlphaGo:

Video is AlphaGo Zero: Starting from scratch. Credit DeepMind/YouTube

This is the story DeepMind is telling us:

After just three days, AlphaGo Zero evolved to the stage where it was more powerful than last-year’s AlphaGo – the version that defeated Lee Sedol. AlphaGo Zero defeated that AlphaGo by 100 games to 0.

Then, using its unsupervised learning strategies, and a mere forty days after starting from scratch, AlphaGo Zero has not only beaten all previous versions of AlphaGo, but it has also explored spaces where humans have never ventured. It has independently found answers – strategies and styles of play – that humans have never even thought of; not in thousands of years of playing Go.

So now the most powerful Go player in the world is AlphaGo Zero. And it didn’t learn anything specific from us. This is a superhuman, but non-human, de-novo AI – albeit a pretty narrow-AI. It emerged from an initial tabula rasa state.

Professor David Silver again:

Video is AlphaGo Zero: Discovering new knowledge. Credit DeepMind/YouTube

For me, this is the bottom line: AlphaGo Zero is an intelligent agent that emerged, by itself, starting from scratch, from the logical laws in its toy Go-universe.

The following are some useful links, that helped me to understand:

• DeepMind – AlphaGo Zero: Learning from scratch.
• Scientific American.
• Kurzweil AI.
• IEEE Spectrum.

The detailed Nature paper is here.

Update: Google DeepMind has generalized AlphaGo Zero to AlphaZero and has allowed it to train itself on Chess and Shogi, the Japanese version of chess.

In just four hours of self-play training, starting from scratch with no human knowledge, it trained itself to the point where it defeated the current reigning champion, an algorithm called Stockfish, in a hundred-game tournament. AlphaZero defeated Stockfish by 28 games to zero. The remaining games ended in draws.

Here are some links:

• Chess24
• Futurism
• ArXiv preprint.

The Logical Priority Of Logic

Photo is Enlightened by Jonatan Pie via Unsplash

One of the things that used to frustrate me about the “Abrahamic” theisms, was their notion that God was prior to the logical cosmos. It always seemed obvious to me that logic was logically prior. Without a prior logical order, mind makes no sense. Mind has no purpose or function. Mind has no rules to obey. Mind has no formative structure. Logic had to be the most-foundational structure, like the Platonists say, then mind could evolve as an adaptive structure. Then mind could function according to logical rules. Then mind could be purposeful.

I found myself thinking about this, as I read about the spontaneous emergence of an intelligent agent, AlphaGo Zero from a tabula rasa state. Particularly, this thought underpinned my last paragraph, in the previous section:

For me, this is the bottom line: AlphaGo Zero is an intelligent agent that emerged, by itself, starting from scratch, from the logical laws in its toy Go-universe.

Because, I find that so reminiscent of how we emerged – from the logical laws of our Universe.

Only Science Prioritizes Mind And Logic Correctly

The scientific narrative is that the mind evolved, emerging from within a logical universe. This gets the logical priorities the right way around.

The religious narrative is that there was an intelligent creation of a logical cosmos, by a pre-existing Mind. This gets the logic exactly back-to-front.

Because then, God would need to have his separate Cosmos that he evolved in – in order to explain where his logic and purposefulness came from. After all, what use is a mind, unless it has a logical cosmos to act in?

Unfortunately, even if that kind of narrative does explain our existence, before we created the Go-universe, it can’t explain God’s existence – not if we insist, he is the uncaused, first cause.

This point used to frustrate the hell out of me, when I was an early-teen student in a religious school. Even before I really understood evolution, I could see that an ex-nihilo intelligent creation does not make any sense. Mind prior to logic does not make sense.

I think at most, I might have been persuaded to accept some sort of entangled cosmogony, where mind and cosmos co-exist together – as a complex system. Maybe in alternating cycles, like a chicken-and-egg? So perhaps an emergent Vedic or Dharmic theism? This notion captivated me in my late teens, when I first discovered Eastern religions. At least the inter-dependent, co-existence of Mind and Cosmos, solves the question of, why does God have a mind? What is it for? What does he do with it? What did he do with it, in the eternity before he created our Universe?

Only in the end, did I finally come around to the position that, actually… no. All you need is a logical ground – a cosmos – and mind can evolve within it. Logic is early; Mind is late.

The minimal scientific narrative is both necessary and sufficient.

Evolution can go further

The theistic theory, or theology, that I was taught as a child, is that before creation, God was complete and self-aware in his own perfection.

However, today I would argue, that the evolutionary implication is, that awareness is largely prior to self-awareness. So, for instance, as experienced meditators know: Ego self-awareness is not that hard to switch off. The self shrinks away, and being-in-the-moment or mindfulness is the result.

I suspect this was actually the evolutionary progression. Awareness evolved first. Self-awareness is its later, second-order elaboration. It emerged, once we became reflective, dynamic agents, within our own mental universe. We needed an elaborate self-representation, to explore the possibilities of our own alternative possible actions. The first-person perspective is a construct that serves the purposes of planning and simulation.

Edit: There are so many important examples of the value we derive from simulating ourselves. For example, our ability to move the ego reference point forward and backward in time; enables us to replay past events, over and over in our minds, and thereby explore different choices. Many people dwell on events in the  past, this way, reliving situations in mental simulation, so as to explore what might have been done differently, or what should be done better next time. This greatly increases the learning value of events that would otherwise occur only once. (Although this can be carried to pathological  excess in very stressful situations like PTSD). People also imagine alternative futures: Relationships with an imaginary lover is a very common one. Trying adventurous possibilities in simulation, before we attempt it in real life, such as, exploring options in career or travel. The space for “imagination” (actually self-simulation) is endless. For a smart Darwinian agent; simulation is cheap. Real world mistakes can be far more costly.

Only beings-in-time, and beings-in-a-world, need an active and fluid, first-person perspective. What possible purpose would self-awareness serve, for a being who was neither in a world, nor in time?

Man makes God in his own image. We abstract, to an infinite degree, our own evolutionary adaptations. We call them “God”. We don’t understand their essential contingency. We are reflections, responses to the world that created us. Our attributes are actually survival-strategies, constructed by evolution.

Without a world that was so dangerously, not-self, we would never have needed a Self.

---

Update:

I have a new post about DeepMind’s latest innovation, a program that has come unprecedently close to solving the problem of protein folding: AlphaFold Is The Latest Wonder From DeepMind.

---

This long-form article is archived, so it is not open for visitor comments. However visitors are welcome to comment about the article, on the blog-post that first announced the publication of this article.