Hi Jeff, Thanks for the fast response (and attached paper). You definitely answered my questions, plus some questions I didn't know I had yet!
Regards Gideon On 20 February 2014 06:06, Jeff Hawkins <[email protected]> wrote: > Gideon, > > Not sure I completely understand your question but I will attempt an > answer anyway. > > > > The "general rule" I was referring to applies to cortex, brain stem, and > spinal cord. A rising afferent axon projects to some structure that > generates behavior (such as spinal cord, brain stem, superior colliculus, > etc.) That same axon splits at that point and continues to a higher brain > region. For example retinal axons project to the superior colliculus > (controls eye movement) but also split and send a branches to the thalamus > (LGN which projects to V1). The same basic thing occurs between regions of > the cortical hierarchy. L5 cells project through the thalamus to the next > higher cortical region (thus are afferents to next region) but these same > axons split and project to some subcortical structure involved in > generating behavior. Guillery and Sherman have written extensively about > this idea (as well as others). I have attached one of their papers that > talks about this. ("The thalamus as a monitor of motor outputs" Phil. > Transs. R. Soc Lond. B 2002). > > > > Now to the new ideas I wrote about recently. The CLA is a high order > sequence memory and as currently implemented has nothing to directly do > with motor behavior. It just learns high order sequences in the afferent > stream and predicts the next input. I now can show that the same CLA > mechanism can learn to predict the next input when the changes in in the > afferent stream are not repeatable high order sequences but are caused by > movements of the body (such as saccades). We can do this if we provide a > copy of the motor command that generated the behavior to the CLA. I worked > this out on paper under the assumption that an SDR coding of the behavior > is provided to the distal dendrites of the CLA. I propose that L4 cells > are doing this, building a predictive model of changes due to our own > behavior. > > > > We have subsequently worked out (on a white board) how L4 and L3 work > together so that L3 forms stable representations of predictable changes in > L4, but any changes L4 can't predict are passed on to L3. I find the idea > compelling. > > > > The exact biology of how the copies of motor commands get to L4 is a bit > of a mystery. We know that copies of motor commands are sent to the > thalamus and thus will arrive in the cortex, almost certainly in L4. > However, I have not found reliable data (or conflicting data) on the > following questions. > > - Do these motor commands project to the same thalamic neurons as the > sensory data? I have found evidence that they do but also evidence that > they may project to separate thalamic relay areas. > > - If the motor commands are kept separate in the thalamus do they project > to the distal dendrites of L4? > > - We know that 60% of the distal inputs to L4 come from cells in L6. > These L6 cells also send a branch of their axons to the thalamus are > believed to be an attention command. Attention is considered a type of > behavior. Thus these L6 cells are another example of a motor command > splitting and being passed up. A lot is unknown about these layer 6 cells > but this may be the copy of the motor command we need. > > > > All of this is confusing but I am confident the L4 predictive model will > work and we apply it to many cool problems. I am working on trying to > understand the exact biology but I don't feel we have to wait to get all > these details. > > > > To answer your second specific question, the splitting occurs > sub-cortically. It also occurs between regions of the cortex via the L5 > cells that project both to the thalamus (up) and sub-cortically (down). > And it may be occurring in the L6 cells that project both down to the > thalamus and up to L4. > > > > Let me know if I mis-understood your question! > > > > Jeff > > > > > > *From:* nupic [mailto:[email protected]] *On Behalf Of *Gideon > Kowadlo > *Sent:* Wednesday, February 19, 2014 1:53 AM > > *To:* NuPIC general mailing list. > *Subject:* Re: [nupic-discuss] New Temporal Pooler Proposal > > > > Hi Jeff, > > > > I've been following your work for a few years now and am very excited to > read your latest thoughts on the TP mechanism. I was hoping (a little > belatedly) to clarify some aspects. Regarding the paragraph: > > > > *"As a general rule, sub-cortical neurons that generate behavior have > split axons. One branch generates the behavior and the other is sent to the > cortex. If the cortex didn't get copies of motor commands you couldn't > function. Every time you moved your eyes or turned your head it would > appear as if the world was spinning and shifting."* > > > > In previous descriptions of the biology in the white paper and in the > literature (such as Sherman 2007), a signal is passed a) to motor regions > and b) to higher cortical regions via a thalamic relay. How does your new > approach fit into this biological understanding? > > > > More specifically: > > · Is this the copy of the motor command that is input to Layer 4? > If not, is it additional or an alternative understanding? > > · Does the connectivity of the splitting axons as described, > occur between layers as you move up the hierarchy, or just at the > sub-cortical to cortical level? > > > > Regards > > Gideon > > > > > > On 30 January 2014 12:10, Dean Horak <[email protected]> wrote: > > Blender is a pretty powerful open source 3D modelling and animation tool. > > > > The learning curve is a bit steep if you're not familiar with 3D modelling > and animation, but for relatively simple modelling it's not too bad. It > supports a python scripting interface and supports many import and export > formats (3D Studio [3ds], AutoDesk animator [fbx] and Wavefron [obj]). > > > > http://www.blender.org/ > > Another nice open source tool is POVRay (Persistence of Vision Raytracer). > > > > This uses a text based modelling language to define scenes and supports > animations. It also has a large community of tools for visual modelling, > format conversions and add-ons. > > > > http://www.povray.org/ > > > > Both of these are capable of producing studio quality animations (in fact, > I believe Pixar uses - or at least used to use Blender for their > animations) and have a very active community for support and plenty of > examples to learn from. > > > > While these may be overkill for simple animation sequences, it's always > nice to have the advanced features available if you need them. > > > > Dean > > > > On Wed, Jan 29, 2014 at 6:18 PM, Jeff Hawkins <[email protected]> > wrote: > > I was thinking more of a markup language. I would favor power and > quality of the output over ease of use. > > > > *From:* nupic [mailto:[email protected]] *On Behalf Of *Jeff > Fohl > *Sent:* Wednesday, January 29, 2014 9:54 AM > > > *To:* NuPIC general mailing list. > *Subject:* Re: [nupic-discuss] New Temporal Pooler Proposal > > > > I don't know of any such tool that would provide what you are looking for > right out of the box, but it should be possible to create something. Making > it Web-based would be useful, to make it as portable as possible. Graphics > and animation on the Web are quite mature now. Not sure what the effort > might be. What sort of interface are you thinking of? A GUI, or something > like a markup language? > > > > On Wed, Jan 29, 2014 at 9:19 AM, Jeff Hawkins <[email protected]> > wrote: > > I would love to have an animation tool available to teach HTM/CLA > principles. The output should look good and be useful for conveying > concepts, something that could be used on a website or in a presentation. > Ideally there would be some kind of language that generates animations. > People like myself could modify this as needed for new ideas, > presentations, etc. Does anyone have a suggestion on such a language > and/or want to work on creating animations like this? > > Jeff > > > > *From:* nupic [mailto:[email protected]] *On Behalf Of *Sergey > Cleftsow > *Sent:* Wednesday, January 29, 2014 6:03 AM > > > *To:* NuPIC general mailing list. > > *Subject:* Re: [nupic-discuss] New Temporal Pooler Proposal > > > > Consider another kind of the HTM (API etc) visualisation like a mind map. > I'm sure many of us do the same like this: > http://www.mindmeister.com/ru/352874116/nupic > But why make it separately while it is possible to work on it together? > It may be accessable for every community member. And every of them (i.e. - > us) can easily contribute when he/she has time to do that. But may be > somebody should be a *corretor*. > > > > 2014-01-28 Brev Patterson <[email protected]> > > Hi Subutai, > > > > In my personal time, I've just finished a serious study of On > Intelligence, and will be starting the Whitepaper shortly. > > > > As part of teaching myself this, I'm going to start organizing my > notes/illustrations/etc into documentation for others. > > I should be right in line (both topic and timing) to help on the > documentation/illustration you mentioned. > > > > Thanks, > > Brev > > > > > > *From: *Subutai Ahmad <[email protected]> > *Reply-To: *"NuPIC general mailing list." <[email protected]> > *Date: *Tuesday, January 28, 2014 8:46 AM > *To: *"NuPIC general mailing list." <[email protected]> > > > *Subject: *Re: [nupic-discuss] New Temporal Pooler Proposal > > > > One thing we'll want to do with the new proposal is to create detailed > step by step illustrations of every aspect. This was sorely missing I > think in the whitepaper. It would be awesome if anyone wants to contribute > with this. > > > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > > > > > -- > > SY Sergey V. Cleftsow > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > > > > > -- > Gideon Kowadlo, Ph.D. > http://gideon.kowadlo.net > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > -- Gideon Kowadlo, Ph.D. http://gideon.kowadlo.net
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
