I was thinking about how connectivity clustering is implemented in the cortex. The most likely substrate is lateral axonal branches of L2-L3 pyramids. Their target synapses would be reinforced when the nodes fire, although they initially fire due to vertical input from L4.
The question is how the cluster gets to be represented by a single higher-area neuron, to compress representations into compositional hierarchy.
I think this compression / selection is like done by WTA in thalamus, related discussion: Burst as a local learning rule in apical but not basal dendrites - #15 by bkaz, Corticothalamic Pathways From Layer 5: Emerging Roles in Computation and Pathology - PMC.
Thalamus mediates generalization: cortico-cortical loops from lower to higher areas, where the higher neuron should represent a cluster of related lower neurons. The centroid of the cluster will get the most reinforcement by lateral connections, so it will be the winner. The clusters here are AKA ensembles, macro-columns, or patches?
I could be terribly wrong but I think that the brain does not do “grandma cells” but instead, distributed representation.
While you consider the pros and cons, remember that cells die. A single point failure of some key memory would be catastrophic in the single neuron model, but of lessor importance in a distributed model.
I am simplifying here, it obviously doesn’t have to be single cell, the cells in in the same layer of minicolumn are mostly redundant. And the clusters are fuzzy, the same lower node is represented in several clusters. But there must be upward compression on the average, generalization is reduction, otherwise you get combinatorial explosion.
Hi, Mark 
That, and you get all the grandma-ish pattern matchers for essentially free !
Interesting question brought up by the OP. But replying to Bitking’s comment on grandma’s cells, it seems that the temporal pole is the most likely candidate where grandma cells (or rather ensemble or group of cells) are located.
Link #1: It's Not a Grandmother Cell, But Maybe It's a Grandma Cluster | Cell And Molecular Biology
I think it’s obvious that we can recognize and disambiguate individual objects and their systems, which are technically clusters. That means each cluster is represented by one cell. There can be many of such cells representing the same cluster, but each of them represents all or most of it’s elements.
The problem with the scheme I outlined is that active neurons in the same area can represent multiple clusters. I don’t know how WTA in thalamus can be localized within cluster, vs. within the whole source area.
Meh, I think that’s debatable. A repeating stream of patterns of several cells more likely. “Grandma” is not a single thing despite we conceptualise it this way. There is the word, the memories, the consecutive details which all call back the word, and her name, and more memories. Never a static thing. It seems static since the recurrent stream repeats again and again the same patterns but it isn’t. No pattern sticks forever.
It is a single thing, referred to in multiple ways.