Learning specific timing sequences via "mobile synapses"

I find the premise of neural timing mechanisms fascinating, and though in the recent HTM chat with Jeff exact timing is raised and there’s mention that it’s not a current priority for HTM, i’m optimistic that capturing it may at some point add robustness to sequence learning, among other things.

I recently finished Coursera’s Computational Neuroscience (hosted by UW), which mentions the way timing is used in the auditory system, specifically (in lecture 5.4), the way delay lines may enable sound localization by detecting a specific order of incoming action potentials from neurons receiving input from each ear (see cs528 lecture notes, slide 21). That’s intriguing in and of itself, but has made me think a lot more about the value of timing-based learning, and potential mechanisms.

As illustrated, it’s easy to imagine a setup where even a single dendrite segment is able to detect a coincidence of action potentials from sources with a specific relationship between firing time and (temporal) distance, and as such fire only when a specific sequence occurs. This is clearly useful, but how could such a relationship be learned?

What if synapses were semi-mobile, and could translate (either during initial path-finding/growth, or after connection with the post-synaptic cell)? I haven’t found any resources to confirm such a possibility, but have recently been reading about the incredibly complex multi-player system by which axon pathfinding occurs. These materials show us that branching and select synapse destruction can occur during navigation, so even if explicit ‘translation’ is not feasible, a selective dynamic acting upon a population of co-located synapses certainly is.

Clearly there’s a significant (and mostly unfounded) assumption here, but I followed it to see if any useful properties would result.

Further assume a learning rule is feasible in which synapses to a dendrite segment can be influenced by nearby action potential activity in that dendrite, to move or be attracted in a particular direction (maybe via some sort of local voltage/chemical gradient). Specifically, what if synapses could move, at the time of synaptic transmission, towards an active action potential on the same target segment, if any existed within some limited learning radius.

I found it useful to sketch out what a number of simple scenarios would look like under these (perhaps far-fetched) assumptions. The scenarios are JS, so are available here: synapse timing sketches.

This is very speculative, so I’d love to hear from anyone with more background in the neuroscience on how such a mechanism may or may not be biologically feasible.

Thanks for reading,




Hey Jeremy. Timing is also an interest of mine. I have a hunch that various timing characteristics will end up being quite important.

I’m skeptical of spatially mobile synapses. But what I do know about is activity-dependent plasticity of axon myelination, which suggests that neurons may modify conduction delays for computational purposes [1,2,3,4]. Jeff Krichmar’s group has used an algorithm based on this to plan paths in robotics [5].

[1] Fields, R. Douglas. “A new mechanism of nervous system plasticity: activity-dependent myelination.” Nature Reviews Neuroscience 16.12 (2015): 756-767.

[2] Fields, R. Douglas. “White matter in learning, cognition and psychiatric disorders.” Trends in neurosciences 31.7 (2008): 361-370.

[3] Wang, Runchun Mark, et al. “An FPGA implementation of a polychronous spiking neural network with delay adaptation.” Frontiers in neuroscience 7 (2013): 14.

[4] Wang, Runchun M., et al. “A mixed-signal implementation of a polychronous spiking neural network with delay adaptation.” (2014).

[5] Hwu, Tiffany, et al. “Adaptive Robot Path Planning Using a Spiking Neuron Algorithm with Axonal Delays.” IEEE Transactions on Cognitive and Developmental Systems (2017).


Many thanks for your response, Jake. Looking forward to reviewing these papers.