Branch Point Computing: Beyond Synaptic Plasticity

For years, plasticity has been taken as synonymous of synaptic plasticity. In other words, the neuronal correlate of learning is believed by most neuroscientists to be in the connections between neurons. Yet synaptic plasticity is a relatively modest form of learning: it ties an input to an output, creating pairwise associations. Complex functions of many inputs can be created by the convergence of multiple inputs onto a dendritic tree, but synaptic plasticity as we know it still operates on input-output pairs. A more powerful form of learning would lie in a mechanism that modified the pairing of two inputs to one output, yielding an input/output function that takes two inputs and converts them into one output. Neurons are full of such transformations: dendritic branchpoints. Dendritic trees are essentially sets of binary trees. At each branchpoint, two denditic branches converge onto one. When one or more electric signals reach a branchpoint, the branchpoint carries out a computation: the signal may carry on towards the cell body, amplified or unamplified, or it may die, in what neurobiologists call a \"branchpoint failure\". The result of this computation is no doubt modulated by which ion channels are expressed in its vicinity. Thus, modulation of the expression of these ion channels will change the computation. It appears to me improbable in the greatest degree that evolution has not exploited this mechanism. Thus, Hebb\'s simple 1 input-1 output law ought to have a 2 input-1 output analog, one that dictates the way in which the input-output function is modified with activity or experience. Compared to Hebbian learning, such a mechanism would allow for far more contextual information to be stored: as opposed to synapses, which remember simply that each of N inputs successfully predicted suprathreshold activity of the entire set, the branchpoints can remember exactly which pairs of inputs were active together. Similarly, branchpoint plasticity would allow for an unprecedented mechanism for generalization, as the activation of two inputs that converge at a branchpoint could strengthen the association between the corresponding branches in a way that generalizes to activation of \'\'any \'\'set of inputs that activates those branches. This could make each branch a representation of a particular concept whose associations can be manipulated independently. Given the small set of synapses between sensoria and neuronal correlates of perception, such independence of manipulation could be crucial. Whether such a branchpoint plasticity rule exists can be assayed these days by carrying out optical imaging of a branchpoint during stimulation of inputs to dendrites on either input (or both inputs) to the branchpoint. I cannot think of a more important task ahead for researchers of learning today. cti.itc.virginia.edu/~psyc220/golgi1.gif

Alex Bäcker Altadena, California, December 2006 (This note is an expanded version of the first version written in January 2004).

Addenda: # For recent evidence of experience-dependent changes in dendritic branchpoint morphology, see http://cercor.oxfordjournals.org/cgi/content/full/14/6/655 . It is only natural to search for changes in the branchpoint transfer (input-output) function now. # Note that branchpoint plasticity can affect information flow in both directions: from dendrites to soma for computation and from soma to dendrites for the function relating activity to plasticity or learning. With regard to the latter, Golding et al showed that backpropagation showed a digital, binary, distribution, exhibiting a bimodal attenuation distribution (Golding NL, Kath WL, Spruston N.: J Neurophysiol. 2001 Dec;86(6):2998-3010. Erratum in: J Neurophysiol 2002 Feb;87(2): Dichotomy of action-potential backpropagation in CA1 pyramidal neuron dendrites.)

Keywords: branch point plasticity, branch point computing