A couple of references relevant to this discussion: CLA boosting is very similar to BCM learning, a form of Hebbian learning ( http://www.scholarpedia.org/article/BCM_theory). In 'vanilla' Hebbian learning, the change in weight of a synapse is proportional to the product of the presynaptic neuron's firing rate and the postsynaptic neuron's firing rate. This is always positive, so it only specifies a way to increase synaptic weight. With BCM learning, the change in weight of a synapse is instead proportional to the product of the presynaptic neuron's firing rate and a function of the postsynaptic neuron's firing rate that depends on a floating threshold that is based on the long term previous activity of the post-synaptic neuron. The more active the post-synaptic neuron has been in the long term, the higher that threshold will be. If the post-synaptic neuron's firing rate exceeds that threshold, the synaptic weight will increase, otherwise it will decrease. The result is that the more active a post-synaptic cell has been in the long term, the more likely its input synapse weights will decrease (except for those input patterns that result in the strongest activity for the post-synaptic cell, representing the very best fitting input patterns). Conversely the less active a post-synaptic cell has been in the long term, the more likely its input synapse weights will increase (resulting in the cell learning new pattern(s) to activate in response to).
BCM learning has had a lot of empirical support, but the biological mechanisms underlying it are not yet clear. A recent advance is that a detailed model of spike timing dependent plasticity (Urakubo et al, 2008: http://www.jneurosci.org/content/28/13/3310.abstract) was simulated by O'Reilly et al and they found that it resulted in BCM learning. An overview of this can be found in this paper: http://psych.colorado.edu/~oreilly/papers/OReillyHazyHerdIP.pdf on pp.10-12. That paper is also a very interesting read for many aspects of modeling the brain and cognition. They also found that using the same learning rule, but basing the floating threshold on a shorter term (hundreds of milliseconds) running average of the post-synaptic neuron's activity rather than a long term average, would result in the error-driven learning of predictions. Interestingly, this sounds a lot like the kind of learning that would need to take place at distal synapses for the CLA's temporal pooler. -Mike _____________ Michael Ferrier Department of Cognitive, Linguistic and Psychological Sciences, Brown University [email protected] On Wed, Dec 11, 2013 at 1:25 PM, Dennis Stark <[email protected]> wrote: > Jeff, wouldn't boosting essentially be similar to neurotransmitter > recovery in synapses? In short term a pre-senaptic axon would be less > likely to excite a post-synaptic dendrite due to the lack of > neurotransmitter. So it's sort of a boosting mechanism but is inverted. > There is also boosting in some of the mechanisms of how the short term > memory works, but I guess it does the opposite, as it's job is to remember > rather then to achieve sparsity. > > On Dec 11, 2013, at 3:55 AM, Jeff Hawkins <[email protected]> wrote: > > Great question. > > Boosting or something like it is essential. Without boosting it is > possible for some columns to never win (become active) and others to win > too much. We started without boosting but quickly saw that the spatial > pooler would have this problem of columns that never won and were > essentially wasted resources. So we added boosting to solve the problem. > > I am not aware of anything in the biological literature that relates > directly to our method of boosting, but I haven’t looked either. However, > a general observation is that most excitatory neurons have a low background > firing rate, maybe once a second or slower. Although this has been > observed and noted by many neuroscientists, I am not aware of anyone > studying the mechanisms that might cause it. It is possible that low > background firing rates could achieve the same result as boosting. All > cells will fire sometime and therefore be given a chance to learn. > > Jeff. > > *From:* nupic > [mailto:[email protected]<[email protected]> > ] *On Behalf Of *Marek Otahal > *Sent:* Wednesday, December 11, 2013 12:11 AM > *To:* NuPIC general mailing list. > *Subject:* [nupic-discuss] Boosting: biological support? > > > Hello, > > I understand why boosting is needed and how is it implemented > (algorithmically), my problem is: does it have an analogy in real brains? > > I'm comparing it with the inhibition (where boosting is like a > counterpart) which is known (local inh) from the real brains and we just > implement it. Or is boosting a new, artificial concept added for an > improved performance? > Thanks, Mark > > -- > Marek Otahal :o) > _______________________________________________ > 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
