Slashdot Mirror
Jeff Hawkins' Cortex Sim Platform Available
UnreasonableMan writes "Jeff Hawkins is best known for founding Palm Computing and Handspring, but for the last eighteen months he's been working on his third company, Numenta. In his 2005 book, On Intelligence, Hawkins laid out a theoretical framework describing how the neocortex processes sensory inputs and provides outputs back to the body. Numenta's goal is to build a software model of the human brain capable of face recognition, object identification, driving, and other tasks currently best undertaken by humans. For an overview see Hawkins' 2005 presentation at UC Berkeley. It includes a demonstration of an early version of the software that can recognize handwritten letters and distinguish between stick figure dogs and cats. White papers are available at Numenta's website. Numenta wisely decided to build a community of developers rather than trying to make everything proprietary. Yesterday they released the first version of their free development platform and the source code for their algorithms to anyone who wants to download it."
http://en.wikipedia.org/wiki/Baum-Welch_algorithm http://en.wikipedia.org/wiki/Viterbi_algorithm
The first is an alogorithm which utilizes forward and back-tracking "to find the unknown parameters of a hidden Markov model." The second is a similar algorithm used for learning 'known' causes (for reference).
I work in computational linguistics and the time an algorithm takes to run and the amount of memory it requires are serious limitations. That's why ad-hoc systems are so common.
Hawkins is a rich guy, and no-one feels like telling him that his stuff is crap. He had a few smart people working for him at some point, but when they told him his ideas were half baked and not new, he just fired their asses.
Here is what many people in machine learning and computer vision think about Hawkins stuff:
- it's way, way behind what other people in vision and machine learning are doing. Several teams have biologically-inspired vision systems that can ACTUALLY LEARN TO RECOGNIZE 3D OBJECTS. Hawkins merely has a small hack that can recognize stick figures on 8x8 pixel binary images. Neural net people were doing much more impressive stuff 15 years ago.
- Hawkins's ideas on how the brain learns are not new at all. Many scientists in machine learning, computer vision, and computational neuroscience have had general ideas similar to the ones described in Hawkins's book for a very long time. But scientists never talk about philosophical ideas without actual scientific evidence to support them. So instead of writing popular book with half-baked conceptual ideas, they actually build theories and algorithms, they build models, and they apply them to real data to see how they work. Then they write a scientific paper about the results, but they rarely talk about the philosophy behind the results.
It's not unusual for someone to come up with an idea they think is brand new and will revolutionize the world. Then they try to turn those conceptual ideas into real science and practical technologies, and quickly realize that it's very hard (the things they thought of as mere details often turn out to be huge conceptual obstacles). Then, they realize that many people had the same ideas before, but encountered the same problems when trying to reduce them to practice (which is why you didn't hear about their/your ideas before). These people eventually scaled back their ambitions and started working on ideas that were considerably less revolutionary, but considerably more likely to result in research grants, scientific publications, VC funding, or revenues.
Most people go through that "naive" phase (thinking they will revolutionize science) while they are grad students. A few of them become successful scientists. A tiny number of them actually manage to revolutionize science or create new trends. Hawkins quit grad school and never had a chance to go through that phase. Now that he is rich and famous, the only way he will understand the limits of his idea is by wasting lots of money (since he obviously doesn't care about such things as "peer review"). In fact, many reputable AI scientists have made wild claims about the future success of their latest new idea (Newell/Simon with the "general theorem prover", Rosenblatt with the "Perceptron", Papert who thought in the 50's that vision would be solved over the summer, Minsky with is "Society of Minds", etc......).
No scientist will tell Hawkins all this, because it would serve no purpose (other than pissing him off). And there is a tiny (but non-zero) probability that his stuff will actually advance the field.
- Anonymous Scientist
I've been working for some time on technology with hierarchical NN architecture like Hawkin's HTM, but mine in part involves SIPO FIFOs with attached neural networks, and the output of the NNs go to the next layer's SIPO+NNs, and so on up the chain. It's intended to extract meaning from symbol flow over time. Like speech primitives into language. Hawkins embeds temporal symbol handling in each HTM layer in a different way. Both of us are trying to emulate some of the processing the neocortex does, but I am less concerned with matching closely the brain and more concerned with outperforming the limitations of the brain. I believe there are classes of problems his architecture will solve, but can't handle others. There's lots of room for people to explore what his technology can do, and I expect it will work well for some things.
Most "grand-scale theories of brain operation", in fact, fail to make claims that can be tested, at least not in the foreseeable future. They predict the large-scale algorithms by which the brain operates. They do not make any claims as to the behavior of any individual neurons, and this is the data we have to work with. Moreover, these theories generally fail to provide any explanation for existing data, such as the diversity in neuronal phenotypes, the connectivity architecture, functional segregation, the wealth of neurotransmitters, laminar structure and why the details of this structure varies across the neocortex, differences in histochemical labelling, and so on and so forth. In short, these theories tend to be computer science, and not neuroscience. They might represent major progress in the question, "how do we make a machine that can solve a difficult computational problem?" but they have very little significance in answering the question, "what are the principles that underlie neural performance?"
Not really. The ability to stop learning is a crucial element of learning. In the existing computational literature, this is related to the problem of overfitting: there comes a point where additional learning is dominated by attempts to explain noise in the data, and can actually lead to degraded performance. A classic example of "frozen learning" in living animals include the zebra finch male (who learns one song only in his life, which remains unchanged past adolescence). Anecedotally, you can also think of human accents in speech; most people never lose the accents they develop in childhood, no matter how hard they might try. Of course, both of the examples illustrate that it is not a matter of stopping learning; learning can in fact occur, but much slower and under much more extreme conditions.
From the perspective of neuroscience (and in fact, from the unsupervised learning perspective in general), given that there are lots of models that can learn and stop learning, the much more relevant question is: how can the system switch between these two states?
Is the problem a supernatural one? Of course not. It is a very tough problem. The issue is not, at this point, a lack of theory.