Part2: Who/Where is the descision maker?

Part 1 I was wondering how to implement Q-value ??

As I was thinking and watching one of Jeff videos, when he mentioned Displacement cells, something clicked…
the solution was starring me in the face…

How do you implement a mechanism that allows algorithms like Gradient descent or RL i.e. how do you do REAL NUMBERS.
/there is no numbers in the Brain, just SDRs. But all our data and models and math is quantitative, so we have to find a way to bring those back into the brain/

I’ve played for a long time with SDR and Kanerva-hypervectors and believe you me there is no way of using overlap or hamming distance to measure SIMILARITY. After you add/union 3-5 items, good luck comparing them.
Comparing is different than checking for EXISTENCE (which unions handle well)

But now with Grid & Displacement cells you can put the SDR-points in Elastic-RefFrame and manipulate the GD-cells or the position of the SDR-points to nudge them towards a solution and use the DISTANCE as a measure .
A measure of which Action to choose. (RL)
A measure of what is the shortest path to a Goal. (Plan)
A measure of how similar is Concept A to Concept B. (asymmetric similarity, basic-level phenomena)

SDR-points are Location+Feature/Landmark probably

/I said Elastic-Frame to implore that changing GD-cell properties we can meta change the whole Concept space, just a thought./

These are just my initial thoughts I’m still unclear on exact implementation but frame with distances is exactly what the doctor ordered.

Another thing I’m not fully aware of is how does the process starts and develops. We do not have the grids fine tuned ready to go ?

I mentioned Displacement cells you can read about them here : A Framework for Intelligence and Cortical Function Based on Grid Cells in the Neocortx

Also Numenta tried to “hide” this little gem, but I “caught” them :wink:
/I knew modulo math was involved somehow/

Supplementary Material

Hey I really like your thinking, but I do not understand what you propose here.

How do you measure the distance between two SDRs, when you cannot use typical similarity metrics (hamming distance)?

Just speculation :

it has to be 4 SDRs : Loc1+Feauture1 and another Loc2+Feauture2 somehow via displacement+grid cells in another frame generate 5th SDR for distance

Or 2 SDR for Location, but keep the Features SDR active and connected.

how exactly is a tough question…

btw I’m thinking of this as algorithm of SDRs, in reality it is probably another neuron-loop.
It is hard to think in neurons, layers and synapses unless first I have grasp of the algorithm.