Yesterday I listened to an ITConversations podcast by Jeff Hawkins at ETech last year titled Why Can't a Computer Be More Like a Brain? I have been following Jeff, his HTM technology, and his new company Numenta since they released their development environment early last year. Anyway, listening to Jeff's marketing spiel again got me thinking about a broader application of abstracted brain models to the web.
In his book titled On Intelligence, Hawkins outlined his theory for human intelligence called the Memory Prediction Framework. The approach is focused a functional abstraction of the neocortex (part of the brain that does lots of useful stuff). The premise is that the hierarchical architecture for this region of the brain relates to the interesting information processing that it does. The model focuses on the architecture, the spatiotemporal abstraction (feature selection) up the hierarchy, and the memory-based prediction afforded by such acquired abstractions.
Numenta's Hierarchical Temporal Memory is a machine learning realisation of the theory. It's kind of like an Artificial Neural Network (neurons, layers, and propagation) although uses a varied inspiration, nomenclature, and training algorithm (Bayesian belief networks). They provide a non-commercial platform called NuPIC to play around and try out examples. Alternatively, there is an open source realisation based on the Hawkins book called Neocortex. As a tech, HTM is an interesting and popularised variation on the theme (ANN broadly conceived), although I suspect hardly a breakthrough in either machine learning or neuroscience.
The specific train of thoughts I had involved relating the decentralised information processing of the brain to the web. For example:
- Content Perspective: Web content are the neurons connected implicitly (pointers) and implcitly (contextually or semantically) to other content. Users provide the active information processing and learning algorithms for the otherwise passive and unintelligent system.
- Users Perspective: Users are the neurons connected to each other through their local interfaces such as popular sites and web applications. Each has wide variations to their connections and therefore reach to other users. Content manifests artifacts of communication (links) and therefore the information (memory).
A final thought is in relating the hierarchical nature of information processing to the transformations in user interaction space (geography, application, social, interests, profile clusters, etc). Each involves feature selection as a means for defining and raising the level of abstraction. I believe there is a pattern to those successful/intelligent web applications that exploit easy (localised) connections across multiple user interaction spaces simultaneously (for example social and profile cluster or geography and application). The interesting question is can you emerge (learn) such levels of user interaction, or do they have to be engineered (system) or contrived (user)?
2 comments:
A far more interesting analogy is modelling a users online interests (neurons) and inter-interest relationships (connections) as a brain. Managing novelty is a big problem, but assessing old web artifacts in this context is relatively straightforward.
Hawkins posits HTM as a repetitive feedback loop where essentially the only divergent activity is due to the appearance of novel inputs, or, “Is it new?” What’s missing is the concept of importance, or, “Is it important?” Why importance? The answer is definitional: unlike a software program, which exists at the whim of its developer, a living creature is an independent complex system that exists separately from its environment, an environment with which it must interact in order to survive. Novelty alone is not enough in this case; the very existence of the organism is dependent upon its ability to categorize and associate inputs and input patterns in accordance with their relative importance to its survival.
Like HTM itself this idea is an excellent fit with the biology of the brain. The hypothalamus, common to all vertebrate brains, is the likely main repository of this determination, acting in coordination with the cortex via complex pathways that flows mainly through the thalamus. HTM is at work here, I believe, and most likely originated in this area, but the precise method whereby the activities of the hypothalamus and the cortex are combined and coordinated involve other mechanisms as well, including the sleep cycle common to all advanced brains....
From rmars.blogspot.com....
Post a Comment