One interesting problem in systems neuroscience is how the nervous system’s motor output interacts with its sensory systems. Sensory inputs that result from motor commands must be either filtered out or used to guide future motor actions. In other words, the organism must distinguish between sensory inputs that are self-generated and those from the outside world. In the juvenile songbird, for example, motor commands for song generation must be sent to some internal critic (likely basal ganglia) so the bird can compare the actual song output to some internal tutor model and improve subsequent renditions.
The weakly electric fish produces regular electric pulses - electric organ discharge or EOD – to actively locate objects within its environment, not unlike a dog sniffing or a blind person feeling objects with his hands. The EOD produces an electric field around the fish that activates electroreceptors on the fish’s body, which the electroreceptors must cancel out (otherwise each EOD pulse would confuse the fish into thinking that a foreign object is nearby – imagine if every time you say something, you can’t tell if it was you speaking or someone else).
The electroreceptors or associated neurons must somehow predict what the EOD input will look like and create a signal that cancels the EOD (a negative image), while still responding to inputs that originate from the environment. But what is the mechanism by which those electroreceptors cancel out the EOD?
In a recent paper from Nate Sawtell's lab at Columbia, Kennedy et al recorded extracellular responses from mossy fibers in the eminentia granularis posterior (EGp) to pulses of the EOD and found a several classes of responses - some cells fired with a short latency after the pulse command (“early”cells), others fired at an intermediate (“medium”), or late latency; the last class was called “pause”cells, which produced prolonged responses after a brief silence. Importantly, these classes of responses to the EOD span the entire timescale of the sensory image that the EOD produces in Medium Ganglion (MG) cells, which is the cell type that must filter out the EOD. The authors demonstrate through a Random Mixture model that a random set of weights of the inputs to granule cells from mossy fibers, Golgi cells and Unipolar brush cells (in the EGp) is sufficient to fit the subthreshold membrane potential responses of granule cells, which suggests that granule cells sample inputs randomly.
A caveat here could be that the dimensionality of the inputs to granule cells in response to the EOD command vs. the dimensionality of the sensory input to MG following the EOD command: the basis function for the sensory input (or its negative image) needs to have a certain number of dimensions; if there are more dimensions among the functions that form the basis than there are in the basis itself (degeneracy), then those functions will definitely form the basis for the negative image. If true, this makes the observation that granule cells combine their inputs in a random linear fashion trivial and not necessarily true (i.e. given enough dimensions, any set of input shapes could be weighted appropriately to produce a desirable output shape). This is somewhat like asking someone to outline the shape of a square using a set of five or more sticks and being surprised that it’s possible.
The authors next modeled the convergent input from 20,000 granule cells to Medium Ganglion cells, which operates under an anti-Hebbian spike-timing dependent plasticity rule where granule cell spikes that precede MG spikes cause synaptic depression. Combining EOD command inputs (i.e. input from mixtures of early, medium, late and pause cells) via granule cells with sensory images from EOD to MG cells caused the MG cells to cancel the EOD-generated sensory input in under 1000 trials, which is as long as experimental data shows that fish take to form negative images.
One interesting data point that remains unaddressed in the paper is that in addition to cells that fire at early, medium and late phases of the EOD (and thus span the timescale of the EOD), there is a fourth class of cells, the “pause” cells, which fire after a brief delay and do so in a rhythmic pattern. It is not clear why the fish would possibly need the “pause”type responses onto granule cells. It makes sense that in order to cover the entire time of the EOD (~200 ms), granule cells would need efference inputs to span the 200 ms range. At first glance, it seems that the early, medium and late responses in mossy fibers would be sufficient to generate negative images. The experimental data in this paper points to unipolar brush cells, which hyperpolarize in response to EOD commands and fire rebound spikes after, as the source of “pause”inputs to granule cells. While the “pause”commands span more time (and later time points) than “late”cells, which is necessary to account for the entire length of the negative image, it is not clear why the pause responses need to fire rhythmically.
Overall, this paper demonstrates clearly how granule cells in the fish EGp could combine a temporally diverse set of EOD command-driven inputs and how Medium Ganglion cells use anti-Hebbian STDP plasticity to generate negative images of a somewhat arbitrary EOD-caused sensory input. Future experiments ought to tease out why the MG is restricted to learning only some, non-arbitrary, temporal dynamics of the sensory input, and if the fish’s natural environment’s statistics play a role in setting constraints on the MG cell’s processing or plasticity.
The electric organ model is relevant in general because of its similarity to mammalian cerebellar circuits, which use sensory feedback to modify future commands. It would be interesting to see if the songbird model of efference copies of motor commands, which works via cortico-basal ganglia loops, bears any similarity to the fish cerebellar-like circuit.