Competition and Language

A path towards strong AI

OpenAI recently created a machine learning algorithm that could play the game of Hide-and-seek. Agents were assigned a team and were given a limited rule set in a simple 3D environment. Not only did they learn the the game, but learned to create barriers, use ramps, and build fortresses among other strategies. The agents played against each other for hundreds of millions of rounds evolving one strategy after another in an arms race of tactics and counter-tactics. This is called self-supervision and in my opinion could be the key to intelligence.

Not so long ago there was another huge win for machine learning when AlphaGo beat Lee Sedol, the world’s best Go player. It too played millions of rounds against itself but was supplemented by knowledge of human strategy, by being fed histories of human vs human games. Alphazero, DeepMind’s latest Go, Chess, and Shogi playing system outshines its predecessor and was not augmented with known strategy. It simply evolved to be a world-champion in multiple non-trivial games by only playing against itself.

Many, many years ago humans may have evolved in a similar manner. Some estimate the likelihood of dying by the hands of another human 10,000 years ago at around 50%. Imagine a world in which a coin flip determines if you will be murdered. It’s a scary place where tight cooperation and communication with your tribe was essential for survival. The battles were fierce, forcing all combatants to acquire strategy and count-strategy. Those intelligent enough to think of new weapons and battle tactics were rewarded with survival and propagation of their genes while those too slow, weak, or less creative did not. Whole tribes were wiped out, but the stronger, smarter groups survived.

So what does this have to do with strong AI? First you need to think of language. What is language exactly? It’s a method of tying together discrete symbols that can transfer a complex thought from one person to another person or group. Why do we have language at all? Many animals do not or have an extremely limited vocabulary. What about language led sapiens to conquer the known world? Cooperation. Specifically, complex and dynamic cooperation that allowed a large swarm of humans to defeat an adversary. Chimpanzees, our closest cousins, do not cooperate like we do. For example, you will never see two chimps carrying a log together. Even in battle they don’t seem to be using group tactics. It’s every chimp for himself.

What came first? Language or complex thought? One theory is that they evolved together in an intelligence explosion. This makes sense if you think about it. “Let’s surround them”. This phrase contains only a few bytes of information and can be communicated in about a second. It’s one exemplar in an infinite realm of ideas describing a 3D scene, geometric descriptions of entities in the scene over time, and a way to position oneself in relation to known friends and enemies. It’s an idea that could tip the balance of survival of one tribe over another without it. Where did an idea like this come from? It’s likely that it was stumbled upon like virtually all evolutionary leaps. Perhaps a primitive group of people randomly performed the action of surrounding while in battle and it worked. They noticed their actions were beneficial, and created a sequence of sounds to symbolize the action to be used again later. Repeat this process a hundred million times and you may get something like intelligence.

Now where does this leave AI or lack thereof? Current natural language processing, while impressive, is primarily statistical analysis of words and how they relate to other words. No matter how many words a program analyzes it will never understand a word like surround like we do. It will never be able to focus or rotate its (visual) attention, recognize pixels as “friend”, and keep a proper distance from them based on the location of other pixels it recognizes as “enemy”. It takes embodied AI for that. That is, AI that controls agents in a 3D environment. Facebook research has recognized this and created Habitat, a realistic 3D environment in order to anchor words to a moving agent.

Back to Hide-and-seek. OpenAi’s agents learned concepts such as barricade and shelter, not in the statistical sense of current NLP, but in the human-like sense of geometric relations between agents and objects given a specific point of view (that of the individual). What’s missing is language or discrete symbols that can be communicated. This, in theory, could amplify their abilities as it did with us.

Our minds are 3D scene generators. Think of a dream. Have you ever had one so vivid that you thought it was real? This is evidence of a very powerful and dynamic scene generator. Children learn something like a physics/game engine the first 18 months of their lives. They learn to predict the future, which means they need to recreate the present and the recent past to calculate trajectories. It has been shown that highly dimensional images can be generated from small (32x32 and 64x64) quantized vectors with DeepMind’s variational autoencoders. These vectors are tiny and discreet (think language) yet they hold the key to generate higher quality images than that of a GAN. Soon they may even be able generate 3D scenes containing agents and their interactions. This is getting close to an actual human thought.

So where to go from here? A first and seemingly obvious path is to continue to create self-learning agents in more and more complex environments and situations. Then give them a means of communication via discrete symbols along with episodic memory. Finally, add an expressive autoencoder that can capture the agent’s recent past, present and desired future in order to create symbols. These symbols (words) will represent their idea and can be communicated to the rest of their tribe. It is the evolution of these discrete ideas and their corresponding scene generating systems that seems a plausible path to intelligence as it was with us.