To follow this blog by email, give your address here...

Tuesday, November 23, 2010

Making Minds from Memristors?

Amara Angelica pointed me to an article in IEEE Spectrum titled MoNETA: A Mind Made from Memristors

Fascinating indeed!

I'm often skeptical of hardware projects hyped as AI projects, but truth be told, I find this one an extremely exciting and promising project.

I think the memristor technology is amazing and may well play part in the coming AGI revolution.

Creating emulations of human brain microarchitecture is one fascinating application of memristors, though not the only one and not necessarily the most exciting one. Memristors can also be used to make a lot of other different AI architectures, not closely modeled after the human brain.

[For instance, one could implement a semantic network or an OpenCog-style AtomSpace (weighted labeled hypergraph) via memristors, where each node in the network has both memory and processor resident in it ... this is a massively parallel network implemented via memristors, but the nodes in the network aren't anything like neurons...]

And, though the memristors-for-AGI theme excites me, this other part of the article leaves me a bit more skeptical:

"
By the middle of next year, our researchers will be working with thousands of candidate animats at once, all with slight variations in their brain architectures. Playing intelligent designers, we'll cull the best ones from the bunch and keep tweaking them until they unquestionably master tasks like the water maze and other, progressively harder experiments. We'll watch each of these simulated animats interacting with its environment and evolving like a natural organism. We expect to eventually find the "cocktail" of brain areas and connections that achieves autonomous intelligent behavior.
"

I think the stated research program places too much emphasis on brain microarchitecture and not enough on higher-level cognitive architecture. The idea that a good cognitive architecture is going to be gotten to emerge via some simple artificial-life type experiments seems very naive to me. I suspect that, even with the power of memristors, designing a workable cognitive architecture is going to be a significant enterprise. And I also think that many existing cognitive architectures, like my own OpenCog or Stan Franklin's LIDA or Hawkins' or Arel's deep learning architectures, could be implemented on a memristor fabric without changing their underlying concepts or high-level algorithms or dataflow.

So: memristors for AI, yay!

But: memristors as enablers of a simplistic Alife approach to AGI ... well, I don't think so.

7 comments:

Toby said...

Yep, they're quite naive. You can't evolve an AGI without the correct selection pressures to evolve against. If at all.

Anonymous said...

Memristor tech is pretty good IFF it causes better FPGA chips. If it does, well, the applications surely aren't limited to AGI. However, I suppose AGI would greatly benefit from that. I wish the evil FPGA firm and I hadn't parted ways.

-- examachine

Tim Tyler said...

Re: "But here's the really interesting thing about a memristor: Whatever its past state, or resistance, it freezes that state until another voltage is applied to change it. Maintaining that state requires no power. That's different from a dynamic RAM cell, which requires regular charge to maintain its state."

That sounds a lot like flash memory.

flamoot said...

"The idea that a good cognitive architecture is going to be gotten to emerge via some simple artificial-life type experiments seems very naive to me."

A 'mind' produced using artificial life simulations would be a binary blob. Even an evolved logical controller (a computer program you could trace and analyze) wouldn't serve as a configurable, stock cognitive architecture that could be repurposed or extended

Nonetheless the cognitive engines so produced can be used as-is, in robots or electronic spaces used for their evolution. Imagine if cycles-per-second boomed to the point where we could evolve-on-demand an intelligent agent for any task. The notion of a cognitive architecture would be summarily defenestrated, entirely! That day is of course not foreseeable but I felt I should say something for a-life. Tom Barbalet had me on the Biota podcast a week ago and that recording should be up in maybe another week... if you're interested in my further thoughts

Cheers flamoot

Anonymous said...

Yeah this seems too naive for me to. I think it's important to understand that "from nothing nothing comes".
So there's got to be something that makes the magic happen.

Tom said...

I think that the point about a useful a-life being basically a binary blob is important.

The article conflates the purpose and utility of digital and artificial-synapse systems; digital systems are, e.g., easily serializable to storage medium and easily reproduced and duplicated. Artificial-synapse systems, however, don't have this benefit, and must indeed be "grown" to purpose, unless some way can be devised to mass-read and mass-set memristor states in a digital fashion. And then you hit problems of storage capacity, etc.

Part of the allure of AI has always been the combination of both worlds -- the adaptiveness and sheer power of wetware/artificial-synapse brains, and the programmability, reproducibility, and storability of digital media. If you throw away one part of this equation -- if you just keep the synapse-like behaviors, but don't have the easy control and replication of digital media, I think that you'll find that applications outside of academia and certain one-off purposes are few and far between.

Just like with the meshing of real neurons and digital circuitry, I believe that research must proceed in parallel to learn how to effectively use this massive memristor-based processing/learning power in real world, commercial applications.

In short, who's going to want to have an artificial cat that's as unique, poorly behaved, and quirky as a real cat, and take just as long to "train?" Why do that when you can have... a real cat? Now, if that "cat" could be duplicated, reprogrammed, and "tweaked" to do things like find your car keys... and then you could copy that "program" to other "cats"... well, that's when things would start to get interesting, but that's also when you start to deviate from memristor architectures back towards traditional digital architectures.

Dave said...

I don't understand the last comment. The first uses of memristors will be memories. Of course you can mass read and set their states.