What is the difference between column and mini-column?


The documentation says: Mini-column: See “Column”.

So I don’t know if mini-column is a synonym for column or is part of a column or something else.


You don’t tell us which document you’re referring to, but in general a column is a macrocolumn. When a paper talks about a minicolumn, it specifically uses the term minicolumn.

A few tips:

  • A minicolumn is a cross-section of the cortex of about 100-120 neurons.
  • A minicolumn is about 28 to 40 µm wide
  • All neurons in the sequence memory section of the minicolumn (L4) receive proximal input from the same source.
  • A (macro)column contains between 60 to 80 minicolumns.
  • Minicolumns are visible with a microscope; (macro)columns are not.
1 Like

In the most recent Numenta research meeting videos the concept of (macro) cortical columns seems fundamental to the theoretical model that is being discussed. A (macro) column is said to hold one or perhaps many complete object models.

However, actual modeling of the HTM spooler depends only on minicolumns (Frontiers | The HTM Spatial Pooler—A Neocortical Algorithm for Online Sparse Distributed Coding | Frontiers in Computational Neuroscience). Likewise, the last slide of the most recent research meeting says (albeit theoretical) movement vectors of grid cell integration map to minicolumns.

The wikipedia monograph for cortical columns discusses the concept as being somewhat arbitrary and obsolete, maybe being an artifact of the seminal paper on the topic. Cortical column - Wikipedia .

Is the Numenta definition of a (macro) cortical column a functional definition rather than a physical one, or should it be? That is to say, should there be a theoretical structure defined by the ability to hold a complete object model? Would it not also make sense to use the term ‘concept model’ to encompass the modeling of actions and compositional objects as well as discrete objects?

I like the idea that function follows form. In the absence of physical evidence for (macro) cortical columns shouldn’t the circuitry of the brain guide the theory? For instance, does anyone know the ratio of thalamic connections to minicolumns in the cortex ( or any other ratio of communicating fibers to minicolumns?

If there were such a thing as a physically discrete (macro)cortical columns then would we not expect forward projecting communication fibers entering the cortex to occur in a ratio that approaches a discrete integer, to the limits of our ability to count such things?


Thanks for the answers.

I’ll take a look at the article.


In the thousand brains theory, it’s not required that there are discrete cortical columns. It’s just easier to explain as if they are discrete. They also provide a useful lens for figuring out what the cortex does, e.g. voting and duplicate models.

In this paper, there’s ~18 neurons in a cortical column per neuron in the corresponding part of a thalamic nucleus. I dunno how many minicolumns are in a cortical column there.

1 Like

Thanks. You have an impressive command of the literature to be able to dredge up the ratio of whisker barrel to barreloid neurons on the fly.

it’s not required that there are discrete cortical columns. It’s just easier to explain as if they are discrete. They also provide a useful lens for figuring out what the cortex does, e.g. voting and duplicate models.

So I’m apparently on the right track thinking about columns as an abstraction (despite whisker barrels being discretely visible under a microscope).

I’m actually surprised to see that the number of input neurons per column is a tight integer even though my thought experiment predicted it. I’m even more surprised that it is a multiple of six. The ratio of thalamic to columnar neurons being 18 such as it is, and a single column representing a single whisker such as it does, I wonder if three neurons represent each vertex of a grid, or if there are three grids per whisker.

I came here after listening to Jeff Hawkins’ interviews with Lex Fridman which are actually pretty comprehensive overviews of the HTM model. So far I’ve viewed all the HTM School videos and Thousand Brain videos and I’m slowly working my way through the Research Meeting videos and the papers linked in the School videos. As an HTM theorist can you recommend any other fundamentals to study?

1 Like

I can provide a couple suggestions for additional resources:

First is a paper that Subutai co-authored with a Kjell Hole, Director of Simula UiB in Norway, called “A Thousand Brains: Toward Biologically Constrained AI” . It covers the components of HTM and the Thousand Brains Theory.

The second is actually a set of videos from recent research meetings. Jeff and Subutai have gone through each of our core papers that comprise the TBT. It’s a good overview of the theory to date, right up to and including Jeff’s current thoughts on what we now know, new ideas, and future research:

  • How Active Dendrites Enable Prediction and Context Integration in Neurons (Video)
  • Voting in the Thousand Brains Theory (Part One / Part Two)
  • Objecting Modeling in the Thousand Brains Theory (Part One / Part Two)
1 Like

Thanks very much.

I should probably post this question as separate topic but, since you are an employee you may know, is Numenta participating in the Intel Neuromorphic Research Community which apparently is the gateway for access to their Loihi chips?

1 Like

Thank you too. I just know what terms to search on google scholar.

Right. I think I remember Jeff or Subutai saying TBT doesn’t require discrete columns, but I could be remembering wrong or misrepresenting details. To be clear, functionally discrete columns are common in the cortex, they’re just not always a thing, probably.

That’s an interesting connection. I think it was 18 plus or minus 3, but biology is messy, so it could still be an integer ratio as far as theory is concerned. It’d be interesting it that’s a general rule. I couldn’t find any papers which give the ratio for other parts of the brain.

I can’t think of anything too important. There are a lot of talks on youtube, although some are outdated.

numenta.com has a page about cortical columns. The section " What is the difference between a cortical column and a mini-column?" includes the statements:

Cortical columns span from the top to bottom of the neocortex and are much larger. [than minicolumns]

The input layer of each cortical column are arranged in mini-columns. In our simulations, there are typically 150-250 mini-columns per cortical column, with 16 cells per mini-column.

Some (most?) Numenta software uses variable names incorporating “column” (eg cellsPerColumn) in which column refers to simulated mini-columns.