Agent Plus Environment

Meandering along a street in Georgetown recently, I looked out at the throngs of people, walking in pairs and groups and singly, ethnicities as varied as their unseen motivations in walking out that day. I saw them, and felt apart; not disconnected or of some different species, an outsider looking in, but in the way of feeling that I was more self-aware at that moment. A single cell in a much larger organism, looking out and seeing the workings of all other cells. It was a knowing that I could see us all, self-organizing sacks of meat composed of so many tiny molecules, marvelously complicated and simple at the same time.

Hierarchies of behavior, heuristics for selecting actions, emotions and motivations influencing them all. Somehow self-aware, conscious of being, conscious of our consciousness. And I, in that moment, more conscious than the rest, looking at the curve of the road, the brick buildings squeezed up beside white-painted shops and little restaurants, the cars and trucks speeding by, and the people, all blissfully ignorant of everything at my level of awareness.

It was a feeling of awe. If I were a religious being, perhaps I would say I felt the hand of a god, touching me then, showing me the vast oneness of the universe. But I am not, and so I interpreted the feeling otherwise: not a part of any great unity, no; merely one agent existing in a world and conscious of that existence. Marveling at that existence — that any clump of matter could build a society, could conquer the land and sea and air, could walk briskly down a cement sidewalk thinking or not thinking of the weird complexities of the animal brain that made any of these accomplishments possible. That we could organize ourselves such that societies are possible, and so are streets, and restaurants, and cars.

Sometimes it strikes me like that: The realization that all we are, all we ever are, are clumps of molecules bumping around. We have motivations, emotions; we have what feels like intentionality and we speak to each other, creating stories and lies and truths. But we are not special. We have no meaning. The most marvelous thing of all is that we have no more meaning than the simple biologically-inspired, behavior-based robots I created for my thesis. Agents, existing in and interacting with a world. Fascinating, brilliant agents.

Moments like these, wondering at how any of us can exist and marveling at how we do and are conscious of it, I know I picked the right thing to study. That I can be here, writing about the meanings we create and believe, is stupefying and wonderful.

Billions and trillions

Step by slow, supercomputed step, we approach singularity.

This step: Two massively parallel cortical simulations, run at the Lawrence Livermore National Labs by Rajagopal Ananthanarayanan, Steven Esser, and Dharmendra Modha of the IBM Almaden Research Center, and Horst Simon of the aforementioned labs–these are the guys who previously simulated at the scale of mouse and rat cortices. They used a Blue Gene supercomputer (with a whopping 456 CPUs and 144 TB of main memory–just wait, ten years from now I’ll look back on this sentence and laugh at how little computing power and memory that is). The first, and larger, simulation included 1.6 billion neurons and 8.87 trillion synapses. Human brains still dwarf these numbers: roughly 20 billion neurons and 200 trillion synapses. But it’s a cat-sized step with the complexity and scale of a feline brain.

The first simulation used experimentally-measured gray matter thalamocortical connectivity from a cat’s visual cortex–the simulations neurons were connected in a biologically plausible fashion. Phenomenological spiking neurons, individual learning synapses, axonal delays, and dynamic synaptic channels were all included in the software. The second simulation, with 900 million neurons and 9 trillion synapses, had probabilistic connectivity.

Speed-wise, the researchers report that their simulation runs 2-3 orders of magnitude slower than real-time, when compared to a human cortex. With near perfect weak scaling (doubling the memory resource doubles the model size that can be simulated), human-scale models may be just around the corner… well, relatively speaking; the researchers predict it’ll happen in less than ten years. Just as soon as there’s a supercomputer super enough.

The research paper is also available at researcher Dharmendra Modha’s blog [PDF].

But bigger isn’t necessarily better

We may have to wait ten years for human-scale simulations, but we may not need a human-scale platform to be able to build intelligent AI. Researchers at Queen Mary, University of London suggest that bigger may not necessarily be better, when it comes to brains. A lot of complexity can be found even in tiny insect brains. Maybe it’ll be a swarm of honeybee robots that takes over the world!

The complexity of models

For a time, I was convinced that every model out there would not be an adequate model of what a human brain could do because every model out there had to simplify, and thus, that no model or computer software would ever be able truly intelligent until we had the computing power to make an electronic human. I knew there was value to models, but deep down, I retained the conviction that no model, no simulation, no AI would ever manage the same level of complexity or intelligence as a human without being, simply put, a human.

Fortunately, I was relieved of this notion around the same time I started taking Cognitive Science classes: Humans aren’t the only intelligent creatures, the point of a model is not to create the thing you are modeling, all models simplify some aspect (it’s just a matter of choosing which aspects are most important to get exactly right). The world may be its own best representation, as Rodney Brooks so aptly said, but that should not preclude us from simplifying the world to better understand how it works, nor should that, in return, prevent us from trying to simulate ourselves in software.

I, for one, am looking forward to watching the intelligent honeybee robots and the supercomputer human brains band together to overthrow the government.