things that are like the cerebellum:
- cerebellum
- electrosensory lobe (ELL) of electric fish
- dorsal cochelar nucleus
other things that also seem to do sparsification:
- entorhinal cortex -> dentate gyrus
- antenna-lobe -> mushroom bodies
- olfactory bulb -> olfactory cortex
- possibly, the L1 synapses of in L1 cortex from the interlaminar thalamic nuclei; these synapses are too weak to cause the cortical cells to fire on their own, but seem to amplify the effect of synaptic input in cortical L4
"Marr motif"/"pattern separation"/"sparsification" theory of cerebellar function:
- (dense) inputs come from the precerebellar nuclei ("dense" meaning that most of the cells are firing most of the time)
- many many small granule cells receive these inputs. Each granule cell receives multiple inputs (but only a few; about 4 in the cerebellum)
- granule cells project to purkinje cells; these axons are called parallel fibers.
- purkinje cells are also contacted by climbing cell fibers which are axons from the (inferior?) olivary nucleus. There are many fewer purkinje cells than granule cells (unpublished scaling law)
- golgi cells, of which there are much fewer, also receive inputs from precerebellar nucleus, and also from granule cells. They output (inhibitory?) to granule cells.
- the purkinje cells project inhibitory axons to the precerebellar nuclei
supposed functions of cerebellum-like things:
- cerebellum: eyeblink conditioning, temporal difference learning
- cerebellum: vestibular-ocular reflex (note: can be consciously turned off; when this happens, the change is visible in the purkinje cell output)
- cerebellum: smooth, fine motor movement (note: being drunk is like cerebellar lesions)
- cerebellum: not acting lik a child (note: being drunk is like cerebellar lesions)
- ELL: subtract distortion in electric field caused by the fish itself moving
- dorsal cochelar nucleus: subtract effect of head position from sound (or ear position, for animals with movable ears)
- in general, create an inhibitory signal that will "subtract" some self-generated confound from some other signal
supposed functions of other supposedly sparsifying things:
- entorhinal cortex: inputs eg grid cells, dentate gyrus: place fields (then outputs to CA3, CA2, CA1, which presumably do autocompletion "pattern completion")
- antenna-lobe, olfactory bulb (where the glomeruli are): inputs (how much various odor receptors are activated). olfactory cortex: sparsely firing cells; approximately fixed number of cells firing at any one time (small fraction of total). The inhibitory cells in the olfactory bulb essentially subtract a constant amount of activation from the glomeruli but the glomeruli firing rate can't go below zero, so the effect is to silence the vast majority of the glomeruli
supposed goals of sparsification:
- start with a dense, typically topographic coding
- losslessly (approximately) transform to sparse, distributed aspatial coding
- the granule cells fire when the conjunction of their afferents are firing
- the purkinje cells at rest fire at a constant basal rate (because they are inhibitory; if this is "zero" then this allows them to transmit negative as well as positive)
- the golgi cells inhibit to granule cells to do something like mean-subtraction
- the climbing fibers are a supervised learning signal (error signal; tells the cerebellum what it is supposed to be learning to subtract)
- do the GRANULE cells have a basal firing rate, or do they fire sparsely?ws