Very nice description Fergal! It's worth emphasizing your last point. The term HTM is more meaningful and should be used outside the NuPIC community. We're trying to stop using the term "CLA" altogether.
--Subutai On Fri, Jan 9, 2015 at 5:27 AM, Fergal Byrne <[email protected]> wrote: > Hi Dinesh, > > HTM refers to the general theory developed by Jeff Hawkins and Numenta > over the past 1-15 years. You can think of HTM as the general "big idea" of > how we believe the neocortex works. The key aspects of HTM are Jeff's six > principles, which refer to hierarchy, sparse distributed representations, > online learning from streaming data, a uniform algorithm, combination of > sensory and motor function everywhere, and attention. While the theory will > accumulate detail (for example what the roles of the layers inside a region > might be doing), it grows outwards stably from this kernel. > > Officially, CLA refers to the particular detailed algorithmic design for a > single layer of neurons, which is outlined in the 2011 White Paper and > (partially) implemented by NuPIC. Jeff Hawkins and Numenta have indicated > that they wish to "freeze" this meaning of CLA and use a different name for > new versions of their detailed algorithmic designs. > > The rest of us have become accustomed to using "CLA" to refer to an > algorithmic design which is close to Numenta's, but might differ in some > minor or major aspects. The key features of CLA, which generalise across > most of our models, are: > > - Neurons arranged in columns ("mini-columns" in neocortex) which share > feedforward inputs and have similar feedforward responses. > - Sparsity imposed by a columnar inhibition algorithm. > - Feedforward inputs appear on proximal dendrites (to a column in official > CLA, also to cells in some models). > - Neurons in a layer have axons connected to distal dendrites in the same > layer, allowing for prediction. > - Proximal dendrites perform some version of linear summing. > - Distal dendrite segments act independently as coincidence detectors. > - Layers can learn first-order transitions between feedforward patterns, > and also higher-order sequences using choices of active cells in an active > column. > - Columns which correctly predict their activity have one cell active, > otherwise several cells activate (burst). > > HTM is quite general, allowing for many more detailed theories and designs > to be claimed to correspond to HTM, but It's much easier to quantify how > well a design matches up with CLA proper. > > We tend to use CLA when referring to processes in some detail (at the > layer, column, neuron, dendrite, synapse levels), and HTM when talking > about how things work at the layer, region and brain levels. We'll also be > seen using "HTM" when we propose ideas which supercede or contradict > assumptions underlying Numenta's "official" CLA design. > > The other thing to bear in mind is that CLA is an internal name (within > the community) which has no general currency in either neuroscience or > AI/ML, while HTM is well-known (at least by name) to researchers in both > fields. > > Regards, > > Fergal Byrne > > > On Fri, Jan 9, 2015 at 12:48 PM, David Ragazzi <[email protected]> > wrote: > >> Dear Dinesh, >> >> > 1.What is the difference between CLA and HTM? 2.Is CLA generalization >> of HTM as the CLA(the agorithms based on cortex) name suggests so? >> Explain if wrong. >> >> CLA => Cortical Learning **ALGORITHMS** >> HTM => Hierarchical Temporal **THEORY** >> >> As the names say, CLA tries simulate what the HTM states about how cortex >> could work. Something we use wrongly HTM acronym to refer to CLA. But the >> names are clear, one is the theory, the other is the algorithmic model of >> it. Just remember neither all features addressed on HTM are implemented on >> CLA (yet). >> >> David >> >> On 9 January 2015 at 10:08, Dinesh Deshmukh <[email protected]> wrote: >> >>> Hi >>> >>> 1.What is the difference between CLA and HTM? >>> 2.Is CLA generalization of HTM as the CLA(the agorithms based on cortex) >>> name >>> suggests so?Explain if wrong. >>> >>> Thank you. >>> >>> >> >> >> -- >> David Ragazzi >> MSc in Sofware Engineer (University of Liverpool) >> OS Community Commiter at Numenta.org >> -- >> "I think James Connolly, the Irish revolutionary, is right when he says that >> the only prophets are those who make their future. So we're not >> anticipating, we're working for it." >> > > > > -- > > Fergal Byrne, Brenter IT > > http://inbits.com - Better Living through Thoughtful Technology > http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne > > Founder of Clortex: HTM in Clojure - > https://github.com/nupic-community/clortex > > Author, Real Machine Intelligence with Clortex and NuPIC > Read for free or buy the book at https://leanpub.com/realsmartmachines > > Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014: > http://euroclojure.com/2014/ > and at LambdaJam Chicago, July 2014: http://www.lambdajam.com > > e:[email protected] t:+353 83 4214179 > Join the quest for Machine Intelligence at http://numenta.org > Formerly of Adnet [email protected] http://www.adnet.ie >
