neurodudes

An extremely interesting trend in neuroscience has been to use the language of Control Theory to explain brain function. A recent paper by Shadmehr and Krakauer does a very nice job of summarizing this trend and assembling a comprehensive theory of how the brain controls the body. Using control theory, they put forward a mathematically precise description of their theory. Because their theory uses blocks that are direct analogues of specific brain regions like the basal ganglia, motor cortex, and cerebellum, they can use brain lesion studies to undergird their ideas about these components. From the paper:

The theory explains that in order to make a movement, our brain needs to solve three kinds of problems: we need to be able to accurately predict the sensory consequences of our motor commands (this is called system identification), we need to combine these predictions with actual sensory feedback to form a belief about the state of our body and the world (called state estimation), and then given this belief about the state of our body and the world, we have to adjust the gains of the sensorimotor feedback loops so that our movements maximize some measure of performance (called optimal control).

At the heart of the approach is the idea that we make movements to achieve a rewarding state. This crucial description of why we are making a movement, i.e., the rewards we expect to get and the costs we expect to pay, determines how quickly we move, what trajectory we choose to execute, and how we will respond to sensory feedback.

This approach of describing brain lesion studies in the context of a well-thought out theory ought to be further encouraged.

Researchers at the University of Nevada, Reno have an interesting and ambitious set-up for doing research in AI that the describe in a recent paper.

From the paper:

We define virtual neurorobotics as follows: a computer-facilitated behavioral loop wherein a human interacts with a projected robot that meets five criteria: (1) the robot is sufficiently embodied for the human to tentatively accept the robot as a social partner, (2) the loop operates in real time, with no pre-specified parcellation into receptive and responsive time windows, (3) the cognitive control is a neuromorphic brain emulation incorporating realistic neuronal dynamics whose time constants reflect synaptic activation and learning, membrane and circuitry properties, and (4) the neuromorphic architecture is expandable to progressively larger scale and complexity to track brain development, (5) the neuromorphic architecture can potentially provide circuitry underlying intrinsic motivation and intentionality, which physiologically is best described as “emotional” rather than rule-based drive.

What’s interesting to me about this is the combination of a embodied robot in a virtual world with a neurally inspired controller for that robot. While there are pros and cons of embodiment in virtual world (some of which have been touched on here before), I think that if your priority is closing the loop from embodiment to research on neural systems, the importance of this kind of approach cannot be ignored.

Neurons come in many shapes and sizes. Frequently, the shape of a neuron is characteristic to its type. Several theoretical papers have demonstrated that the shape of a neuron can crucially determine its pattern of activity, independently of other factors (Mainen & Sejnowski, 1996, for example). Several resources on the web such as neuromorpho.org and the Cell Centered Database are dedicated to maintaining repositories of different neuronal shapes (also known as morphologies).

Any computer scientist worth their salt, noticing this trend, is tempted to say: if neuronal shape is so important, maybe we ought to have good data standards to describe it. That’s just what a paper last year did. It surveyed the popular data standards for modeling, primarily in the NEURON and Genesis simulation packages. The result is a data standard called MorphML, which is part of a larger effort called NeuroML.

Neuronal shape is a weird data type for the computer science world, but I think an incredibly important and fundamental one for deeply coping with the complexity of real brain tissue. It seems to me that many areas of neuroscience research could benefit from the construction of more explicit models of the circuits they study.

The concept that the brain holds maps of the surface of the body in the primary sensory and motor cortex is a fascinating but well known fact to the field of neuroscience since the early work of Wilder Penfield. What is less broadly appreciated is the concept of “peripersonal space”. A new book by Sandra and Matthew Blakeslee describes peripersonal space in the following way:

The maps that encode your physical body are connected directly, immediately, personally to a map of every point in that space and also map out your potential to perform actions in that space. Your self does not end where your flesh ends, but suffuses and blends with the world, including other beings. [...] Your brain also faithfully maps the space beyond your body when you enter it using tools. Take hold of a long stick and tap it on the ground. As far as your brain is concerned, your hand now extends to the tip of that stick. [...] Moreover, this annexed peripersonal space is not static, like an aura. It is elastic. [...] It morphs every time you put on or take off clothes, wear skis or scuba gear, or wield any tool. [...] When you eat with a knife and fork, your peripersonal space grows to envelop them. Brain cells that normally represent space no farther out than your fingertips expand their fields of awareness outward, along the length of each utensil, making them part of you.

What I appreciate about this, besides the stretchy comic book characters that it makes me think about, is that it provides a powerful perspective to begin piecing together a mass of disparate neuroscience data, which the Blakeslee’s capitalize on.

(more…)

From the department of “we are living in a sci-fi movie”, here’s a video of a bionic arm that uses “fluidic muscles”.

The original slashdot article is here.

(more…)

If there’s one lesson to be learned from almost 60 years of AI research it’s almost certainly to be skeptical of anyone who says they have found THE ANSWER to producing human-level intelligence from computers. Even in the face of this, however, I am intrigued by a new company’s approach to developing Artifical General Intelligence (AGI), a term which is meant to indicate Strong AI rather than Weak AI. That’s probably because its founder, Ben Goertzel, manages to skillfully walk the tightrope between staying conservative about how much they can realistically accomplish and still managing to inspire hope that their methodology has the potential to get close to AGI.

(more…)

Interview with Steve Grand on building human level artificial intelligence at Machines Like Us. Really interesting. Via Chris Chatham at (the excellent) Developing Intelligence.

In particular, MLU asks why his current project to create an android was done as a physical robot rather than as a simulation. The answer, that you can cheat too much in a simulation, is familiar to those from the Brooksian school of embodied intelligence. He says that simulations still aren’t good enough to provide the kinds of physical constraints, like gravity and friction, etc, that you get when building real robots .

However, with the availability of free 3D simulation environments that handle physics, like Breve, we are getting a lot closer. Building a robot within a simulation like this, particularly where you don’t modify the code of the the simulation environment itself, is a terrific way to balance the competing interests of keeping yourself honest and avoiding the painstaking mechanical engineering required to construct complicated robots. This kind of environment allows you to build a body with primary sensory systems and primary motor outputs in a similar fashion as one would with real robots.

Why there aren’t more who have adopted this kind of “in silico embodiment” philosophy I think is the result of taking Brooks’ a bit too seriously. Brooks idea of embodiment is very well founded, but back in the day when he first made those statements, there really were no good ways to simulate the physics of an embodied creature very faithfully. Today that is not the case. Moreover, building real physical robots is great if you have a lot of time, or an engineering team, but it’s a huge investment that distracts from the real problem of understanding the nature of intelligence. The fact that the world has extremely few labs that can make that investment is one of the many reasons there aren’t more serious strong AI researchers any more.

Update: Steve apparently received a few comments along these lines and replies.

The field of neuroscience naturally focuses its inquiry into neurons. This approach to understanding the brain by studying its parts has been thought to have a greater potential than that of psychology to understand how the brain works, a comment made by no less than Daniel L. Schacter, chair of Harvard’s Department of Psychology, in his book, The Seven Sins of Memory.

However promising the field has been thus far, even the most accomplished neuroscientists will admit that we still do not understand how the brain really works. I would submit that the current reductionist nature of neuroscience has shed much light on the dynamics of how neurons work, but has to a far lesser degree shed light on how neurons process information. The difference between these two lines of inquiry is important for making progress in understanding how the brain works.
(more…)

Entrepreneur-turned-cognitive neuroscientist Jeff Hawkins is distributing a “research release” of their experimental code base implementing his idea of hierarchical temporal memory described in his book, “On Intelligence”. Hawkins drew inspiration for the model from his own reading about the structure and function of the human neocortex and believes that it represents the foundation for developing intelligent machines.

Jeff explains this surprising move to open source the code for the Numenta Platform for Intelligent Computing (NuPIC) on the Numenta web site:

Why are we making NuPIC available now?

We have been contacted by dozens of researchers and scientists who are excited about HTM and by our work at Numenta. These people are anxious to work on HTM, are willing to be pioneers, and are willing to accept the uncertainty associated with a new technology. We are making our tools available so that these sophisticated developers can start building a community around HTM technology. NuPIC has been under development for 18 months, is pretty solid, and is well documented – including several examples to make it easy to get started – so we’re ready to open up to more developers, even while knowing that we do not yet have benchmarking data, and we cannot make guarantees about applicability to specific problems.

Here’s why Hawkins thinks that HTMs are new.

We have been covering Hawkins’ work for a while now. See these previous posts for more background info.

Neurodudes is actively soliciting code reviews of the newly released software. Is NuPIC the next big thing, or are you left feeling cold? Post your thoughts yourself using the instructions on the right-hand column, or let us know at contactus -AT- neurodudes.com!

Recently, a company named AnyBots appears to have been able to build a humanoid walking robot whose gait is much more humanlike that Honda’s Asimo (If you aren’t familiar with Asimo, check it out)

The robot is named Dexter, and here’s a link to a blog explaining why its better than Asimo. From the blog:

There are of course biped robots that walk. The Honda Asimo is the best known. But the Asimo doesn’t balance dynamically. Its walk is preprogrammed; if you had it walk twice across the same space, it would put its feet down in exactly the same place the second time. And of course the floor has to be hard and flat.

Dynamically balancing—the way we walk—is much harder. It looks fairly smooth when we do it, but it’s really a controlled fall. At any given moment you have to think (or at least, your body does) about which direction you’re falling, and put your foot down in exactly the right place to push you in the direction you want to go. Practice makes it seem easy to us, but it’s a very hard problem to solve. Something as tall as a human becomes irretrievably off balance very rapidly.

This is important because robots that dynamically balance can handle arbitrary terrains and can withstand attempts to knock them over, much like the Big Dog robot we reported on earlier, whereas pre-programmed robots like Asimo cannot. If Honda were smart, they’d buy AnyBot ASAP.

Here’s a link to the Slashdot article on this.