Recently I came across this paper: The Cortical Column: A Structure Without a Function
No doubt, if cortical columns are not functionally relevant structures (which there appears to be evidence for) that would be a very damaging fact to HTM theory. I am curious what some responses might be from HTM theorists. I think HTM theory at the macro scale would need to change quite a bit in that case.
It’s really hard to disprove that columns exist. For example, would three cells scattered around randomly but controlled as if in a minicolumn count as a minicolumn? I think so.
Minicolumns don’t seem necessary unless there’s a reason they’re specifically needed. You could probably use one cell per minicolumns in the temporal memory, where each activates only if it is predicted or not enough nearby cells are predicted, and it would still work as far as I know. It just requires cells with direct spike-driving inputs, biasing signals sent a little bit of time beforehand, and probably some sort of competition (which is necessary anyway for spatial pooling). Minicolumns are efficient though, since they let it pick far less spatial patterns than sequence contexts, which allocates cells better since there are way more sequence contexts.
Cortical columns are probably more complicated than has been established in HTM theory so far because there’s also the where/egocentric/how pathway. There are also neurons in the spaces between cortical columns which operate on larger scales sensory patch size-wise (and also e.g. binocularity). This space is kind of weird. For example, at least in primary visual cortex and barrel cortex, it is mostly parallel lines rather than at every border of adjacent columns. It may or may not function as a separate region.
In primary cortex, those sections between columns might be part of the where pathway because that’s where they project, at least one region does that. Also, some layers and sublayers also might not be arranged into cortical columns.
I see cortical columns as an efficiency thing like minicolumns. The fact that they exist usually but maybe not always is still evidence for treating the brain as if it looks at the world through a straw (not that everything it does is like that). The ideas that produces are applicable without cortical columns.
Unaccountably, ocular dominance columns are present in some species, but not others.
Rodents lack ocular dominance columns whereas primates and cats have them. Rats have very poor vision compared to those and their eyes point more to the sides, so they might not need them, at least not most of their visual cortex since most of it maps parts of the visual field which only one eye can see. They could still have cortical columns for places in the visual field.
Although the column is an attractive concept, it has failed as a unifying principle for understanding cortical function.
I think a dozen neuroscientists could in a couple decades summarize research on the brain sufficiently for us to know exactly how intelligence works. We have an amount of agreement proportional to how much people summarize things and draw conclusions which aren’t just educated guesses. We have many regions in many species to figure out what is essential for the cortex to work, what is essential for the cortex to do things which most of the cortex does (e.g. there’s no primary thalamus for smell), and what is just a specialization (specializations are numerous and a problem but they do exist in the context of fundamental things so they’re not useless).
Hubel and Wiesel’s classic model showing orientation columns as discrete slabs is misleading because the columns are borderless in real life.
It’s very hard to prove whether columns are quantal like they discuss. They aren’t physically exactly quantal (or they are and that’s just noisy results) but they might be functionally, and it doesn’t really matter if they aren’t exactly quantal because two cells with nearly identical orientation preferences will either both fire or not both fire, just at slightly different average rates.
First, nobody seems willing to venture how many minicolumns constitute a column; the number is arbitrary. Second, there is no evidence that short-range connections bind minicolumns into discrete, larger structural entities . . . These short-range connections do not end abruptly along fixed borders in the cortex, as they should if they actually defined the edges of a structurally distinct column.
Cells in barrel cortex (in macrocolumnar layers) have a main whisker to which they respond, which corresponds to the cortical column, but they usually also respond to a few other whiskers somewhat.
It’s not a problem for the idea of columns if it isn’t neatly organized like a man-made machine. HTM theory doesn’t require cortical columns to be so precise. It seems more about simplifying the sensory input by just considering a piece of it and figuring stuff out about objects and figuring out locations, then combining that information in other layers.
No one has demonstrated a repeating, canonical cellular circuit within the cerebral cortex that has a one-to-one relationship with the minicolumn.
If it’s in a lot of different systems, it’s probably part of the fundamental circuitry, or at least it indicates it quite well because otherwise it wouldn’t be in a lot of systems. For example, if minicolumns aren’t fundamental, they still point somewhat strongly at temporal memory, so that’s evidence for temporal memory even if it isn’t always implemented with strict minicolumns.
It remains true that nothing is known about the physiological correlates (if, indeed, any exist) of the minicolumn.
We do not know anything for certain about the fundamental cortical circuitry’s mechanisms (maybe a few things but basically nothing), and there’s no getting around that so the best we can do is show that a bunch of different things point towards a function and then show that it is powerful.
An alternative explanation is that species lacking orientation columns have relatively poor visual acuity or a small visual cortex. Perhaps orientation columns develop only in species that depend on form vision and those that have a cortical surface area sufficiently large to require subdivision into columns. This argument is contradicted, however, by a recent study of orientation tuning in the squirrel. This animal has a cone-rich retina and it surpasses the tree shrew and mink in visual acuity and visual cortical area. Although the tree shrew and mink both have beautifully organized orientation columns, the squirrel has none
That means rodents probably don’t have orientation columns. The visual stimulus is not just bars though so the minicolumns might be different, like a plain spatial pooler.
Mini-column? Any feature I can make out with a microscope is likely to have some purpose.
Macro-columns? If you have read my posts you know I subscribe to a different model of a sparsity enforcement mechanism. I do have some support from papers that describe lateral connection that have the properties needed to form hex-grids. The fact that these are actually observed in living tissue seems to be very strong support.
Please note that this is compatible with the HTM system - it just changes the shape of the macro-column. Spatial pooling needs some sort of voting and a winner-takes-all for a local vote. The essential feature is a local selection of the strongest response to effect sparsification. Picking the local hub of a hex-grid serves this function very well.
Do you mind sharing these with me?
@bkutt The grid-forming method I favor is lateral reverberant connections combined with lateral inhibition interneurons. While L2/3 happens to be the one boss structure with extensive cortico-cortical efferent bundles, the other layers would be boss’ing it too.
Synaptic reverberation underlying mnemonic persistent activity, Xiao-Jing Wang
http://www.cns.nyu.edu/wanglab/publications/pdf/wang2001a.pdf
He proposes four mechanisms to support reverberant activity for short-term memory.
has this information in it:
Given that the stripes revealed by a given tracer injection in the PFC appear to be reciprocally connected (Pucak et al., 1996), it is reasonable to hypothesize that other pyramidal neurons in the superficial layers are the principal synaptic targets of these connections (Melchitzky et al., 1998).
The paper is a hard read (lots on clamp-patch recording) and is heavy on lab technique.
They do confirm that the target is “300–500 μm away” The inhibitory circuit responses are evident in the results but it does not look like they choose to interpret them this way.
" initial experiments showed that the PSCs typically had mixed excitatory and inhibitory components (EPSCs and IPSCs, respectively …"
In fact - they considered these inhibitory signal a nuance and tried to separate them out.
They did not consider the possibility that it may take a trio of cells to start a strong response.
Conclusion offered: The Majority of Layer 3 Neurons Are Postsynaptic Targets of Long-distance Horizontal Monosynaptic Connections
and
A large proportion of the excitatory input to cortical pyramidal neurons is generally assumed to be provided by short-distance, local axon collaterals of neighboring pyramidal cells (Douglas et al. , 1995; Markram, 1997). However, in many of our experiments, monosynaptic EPSCs elicited from distal stimulation sites had a similar or larger amplitude than EPSCs elicited from more proximal sites (snip) Therefore, this finding suggests that long-distance, horizontal, intrinsic projections are a relatively strong source of excitatory input to layer 3 pyramidal neurons.
Finally:
What Proportion of Layer 3 Pyramidal Cells Receive Long-distance, Excitatory, Monosynaptic Inputs?
Our findings also suggest that most pyramidal neurons in layer 3 are targets of long-distance, horizontal projections. Specifically, low-intensity stimulation at long distances from the recorded layer 3 pyramidal cell evoked monosynaptic EPSCs in the majority (77%) of these neurons. However, this proportion is likely to be an underestimate since some long-distance axon collaterals were probably severed by slicing of the tissue blocks.**
" Interestingly, the range of horizontal distances (135–810 μm) between the stimulation sites that evoked peaks in EPSC amplitude was very similar to the range of distances (200–1200 μm) between centers of stripe-like clusters in anatomical studies of monkey PFC (Levittet al.,1993). In fact, previous studies in other cortical regions suggest that clustered connections are the basis of a peaked distribution in EPSC amplitudes. For example, using whole-cell recordings and multisite electrical stimulation in visual cortex slices, Weliky and colleagues (Weliky and Katz, 1994; Weliky et al.,1995) observed EPSC peaks when stimulation was applied at the locations of clusters of neurons that shared orientation selectivity with the recorded cell. These iso-orientation clusters of visual cortex neurons are known to be linked by intrinsic horizontal axons (Gilbert and Wiesel, 1989)"
So - griddy cells are possible useful when they are not forming grids.
I think that the ratio between inhibitory and excitatory connections varies from area to area, causing different behaviors.
A slight rebuttal to one of the points raised in the paper you linked:
