Question on visual recognition example using layers L2/3, L4, L5tt, L6a and L6b


#1

Screenshot from this presentation:

Labeling signaling:

Some points:

  • There’s no distinction between features, sub-objects and objects. All are treated as objects. “structure on structure” (source).

  • The location-space of each object is practically infinite. So, features, sub-objects and objects get equally large location spaces for their compositional structure.

  • “When you have a sense for something there’s cells that are representing that” (source). Inverted, “When cells represent something you have a sense for that”.

Example (trying to recognize “cup”):

Senses:

L4: Sense of the union of “lines”, “edges”, “curves”, etc. at the union of their precise locations in the union of all location spaces of “logo”, “lid” and “container”.

L6a: Sense of precise location of “lines”, “edges” and “curves” in space AND inside the space of “logo”, “lid” and “container”.

L2/3: Sense of the union of “logo”, '“lid” and “container” objects.

L5tt: Sense of the union of “logo”, “lid” and “container” at the union of their precise locations in the location space of the “cup”.

L6b: Sense of precise location of “logo”, “lid” and “container” in space AND inside the space of the “cup”.

Representations:

SDR_A: Union of “lines”, “edges”, “curves”, etc.

SDR_B: Associated precise locations of “lines”, “edges” and “curves” in the location spaces of “logo”, “lid” and “container”.

SDR_C: Pooling of SDR_A and SDR_B into “logo”, “lid” and “container” objects.

SDR_D: Copy of the state of cells in L2/3.

SDR_E: Associated precise locations of “logo”, “lid” and “container” in the location space of the “cup”.

Questions:

Sense of the union of “logo”, '“lid” and “container” objects.

Where is the sense of the “cup”? Is the union of its sub-objects the sense of it? In L6b there is a sense of space for the “cup”. The L6b eventually forms compositions of parent objects such as “cup”, “table”, "“chair”. Do I have to change my attention to the “room” in order to be able to have a singular sense of the “cup”?

When trying to recognize the “room”:

L2/3: Sense of the union of “cup”, '“table” and “chair”.


#2

I’m not sure you are drawing the same conclusion I am. My primary take-away from these diagrams is that there are two similar two-layer circuits. They work together and give each other feedback. The lower layers of these circuits (L6a/b) contain internal unique representations of space. The SDRs here are transient and move quickly as sensors move across space. The upper layers (L4 / L5) contain stable pooled representations.

It is useful to try to break out all the separate SDRs at play as you’ve done above, but don’t take the exercise too far. As you think about all the ways data could be routed between layers, think also about how a connection can “fan out” from one layer into multiple layers. Disparate connections can also “fan in” from different sources into one layer of cortex. So keep in mind how complex these connections can be. Sometimes labelling them can restrict your thinking.


#3

From the framework paper on the hierarchy section:

For example, there is no single model of a coffee cup that includes what a cup feels like and looks like. Instead there are 100s of models of a cup. Each model is based on a unique subset of sensory input within different sensory modalities.

If I understand correctly, the singular “cup” sensation (feeling like it’s 1 cup and not 3 different parts put together) is the activity of all neurons across all the layers that it’s represented, including grid-cells, displacement cells, etc. This includes connections that “fan in/out”.