The US-sponsored DARPA SyNAPSE project and the EU-sponsored FACETS (now Brainscales) project are examples of major initiatives in neuromorphic computing. As for many other fields, the time has arrived for emerging (and soon to be dominating?) economies to start heavily funding similar projects. The "China brain project" is one of such early example.
The project is led by Prof. Dr. Hugo de Garis, the Director of the Artificial Brain Lab in the School of Information Science and Technology,
Xiamen University, Xiamen, CHINA. This presentation and this link provide some additional information on the project, started in 2008 and to be completed this year.
The goal is, as I was mentioning, an early attempt to build China’s first artificial brain. In the presentation, we learn that Hugo de Garis adopts an “evolutionary engineering” approach, by evolving thousands of neural net modules and then connecting in an "artificial brain". The main network architecture is called Parcone (Partially Connected Neural Evolutionary). I could not get any particular insight on what this network is or does. In any case, the networks are evolved on nVidia Teslas, with the ultimate goal of building systems containing thousands of pattern recognizer modules to control the mobile robot.
Prof. de Garis is additionally pushing the Beijing government to invest heavily in the creation of a CABA, or a "Chinese Artificial Brain Administration" to administer and promote the country’s artificial brain industry, especially for the home robot industry.
Overall, the talk is very confusing, and I could not get anything novel or the impression that the project is in the right track. While evolving networks is neither new (I did it when I was a PhD students just for fun...) nor the solution to all problems faced by the dauting task of evolving artificial brains, the idea of evolving a new module when something new is encountered in the environment that the system cannot recognize is questionable from many point of view, ranging from chip design, to scalability, to ... simply the wrong thing to do!
Overall, the project also falls short in the idea of co-evolving hardware and software. These themes were picked up in the introduction of the presentation, but nothing in the talk mentions how they are tackling this issue, if at all.
Despite this project does not look like much more than a conventional (but well funded) attempt to simulate something "cool" with some completely obscure networks, with the occasional surprise occurring when some good results comes up, it is equally certain that more interesting projects will follow from China pretty soon.
Hi, Max,
Just to shed some light, I am slightly familiar with the nautre of Dr Hug De Garis’ approach. he was in charge of the failed (or cut-short) robo-koneko project in Japan.
He has also commented on the future in more speculative social terms, like the war between the terrans and cosmists (people for and against sentient machines).
His basic approach to co-evolving hardware and software was based on Xilinx FPGAs back in the day. The neural Net was to exist in a 3D cellular-automata type space, with dendrite trees and axons evolving from central nodes by turning cells of the ceullular automata to dendrite or axon etc. It’s more complex than it sounds. Basically the 3D CA was like a 3d voxel-space, in which the decrite and axon and synapse trees could “grow”. So a line of many of these cells designated as “axon” type might have constitited an axon.
As far as I remember…there was not a very in-depth approach to modeling biological neurons. Forget Hodgkin Huxley, it seemed more a like an LIP. Though the website was not very detailed.
Spikes where simulated simply as 1s in a stream of bits. So the interval between spikes was not a “real number” but rather a multiple of a limited clock frequency.
The big emphasis was on evolution. He was more focused on evolution than learning, though he hoped to move onto that at some point in the future. He also has a hardware CAM machine, which had some success in evolving some neural nets with approximated some wave patterns.
the robo-koneko robot was to have a simple “central” intelligence that was at best fuzzy logic. The neural nets were mostly dedicated to motor control and input parsing I guess. The approach was not very biology inspired at all. Sometimes Hugo sounded more like Ray Kurzweil than a scientist… talking about his writing on quasi-sinularity type subjects.
He was definitely ahead of his time in that he implemented evolvable massive neural nets in hardware… but the neurons and the networks were not very bio-mimetic, and it was all evolved insitinct, not learning as far as I had read (I could be wrong).
Also, for some reason, the robo-koneko project failed, and Dr Hugo De Garis lost or left his job in Japan.
Care to comment Max?
Just to clarify:
CAM = cellular automata machine
Also, his artificial neural nets were definitely digital, not analog.
As far as I remember, the robo-koneko robot was to have millions of neurons, sub-divided into modules that could be removed or added.
I do not remember Dr Hugo giving much info on any architecture for the neural nets, so it seemed like a soup of those neural-modules, which were soups of neurons. I could be wrong.
I am sort of disappointed by this latest announcement because it seems that his views and approach have not changed much…but this article is a very vague write up so …I could be wrong.
Maybe he has some secret sauce he is not telling us about.
Maybe he has pattern generators as well as pattern detectors. But the all evolution modle seems totally opaque. A brain built up of evolved pattern detector and generator modules will probably not be very intelligent. I may be wrong, but the human brain seems to be much more than that…and there seems to be more intimate coupling between the functions of disparate parts of the brain.
Also, what about imagination and internal modelling? We do not know how that works… so these pattern detectors and generators will just be hooked up as input parsers and motor-output producers. With no room for a “soul” in between.
again…maybe Dr Hugo has some architectural ideas he is not revealing to us…so my assumption of just input parsers and output generators is flawed.
Also… is he doing learning this time or not??
Hmmm… ok… reading the linked site… it does seem that this time it’s going to be more sophisitcated… with lots of “decision” modules….but … you still need a clever architecture and global variables and …I just don’t see how combining lots of special purpose modules is going to yield much…especially if you go for some simple “society of agents” arhcitecture.
Hmmmm…ok… his work might be more bio-mimetic this time. He is acknowledges Markram. Though… is he over-stating how revolutionary Markram’s work is? Blue Brain seems to have stalled, looking from the outside… not a lot of new announcements..and IBM seems to have scaled down it’s involvement too. I mean…obviously Markram is brilliant, he gave us STDP… but on the engineering and simulation side… on the AI side… how much is still happening in Blue Brain??
Thanks for the detailed account of Hugo De Garis’ work. I am not that familiar with his approach. If you would like to write a post about it, please let me know, I would be happy to host it. Failing in an ambitious project is certainly likely…. I got the same Kurzweilian impression in listening to some statement of his talk. However, I would agree with him on the exciting opportunities that robotics will offer in the next decade.
Yes, it is vague. I have not had the time to dig much more in terms of details, just got some high-level info from his talk.
The combination of modules is not what puzzles me, it is the proposal to learn a new module each time you encounter something that the agent does not recognize. I do not see how you escape major issues in connecting these modules, let alone the combinatorial explosion of # of modules, even assuming that they have some strong ability to generalize.
You are correct about the combinatorial explosion, and it’s the first time I have heard of using evolution for online-learning!
He has someone on his team, referred to as Ben, who will be in charge of “conciousness” so it is highly likely they are working on something more sophisticated that just banks of pattern recognizers and generators with a small decision-making module in the middle.
But he did not explicitly talk about any kind of “running internal simulation” or something…so maybe the initial machine will still be stuck in an input-output “flat” loop.
however… I don’t see any reason per-se why they could not, if clever enough, implement a more globally coherent network… with the equivalents of “global variables” global learning signals or timing or control signals.
But the initial robot seems to be pretty primitive…they are not going for a very thalamo-cortical approach maybe. They…want to try something very different from the brain (it might still work)
Ok I’ll send you something tomorrow evening (asia time). I’d be honoured if you would host it. If not.. just take the info and make a post of your own. Thanks.
Thanks Ibad. Out of curiosity: from this website it is indicated that Hugo has retired, which seems to indicate that he left the project before its end. If you know more about it, please let us know.
Thanks
I suspect an evolutionary approach is the correct approach, however, the selection process is non-trivial and probably cannot be rushed without undesirable consequences – I doubt anyone really knows what is needed, and perhaps even some early evolutionary steps are necessary. Our brains are hierarchical, the modern brain centers add capability and can inhibit our primitive reptilian-like brain. Such hierarchies probably exist down to low level decisions, starting with early organisms in the sea that merely attempted to move into the light.
I noted this:
“The project is led by Prof. Dr. Hugo de Garis, the Director of the Artificial Brain Lab in the School of Information Science and Technology,
Xiamen University, Xiamen, CHINA. This presentation and this link provide some additional information on the project, started in 2008 and to be completed this year.”
Was this completed? I noticed that Prof. de Garis has retired. What is the status of this project?
I am unaware of the status of the project. I suspect it has been terminated.