htm-community / htm-2d-object-modeling Goto Github PK

Given an XY location in the environment, this layer should represent this space in many Grid Cell Modules, which work together to resolve the object.

Object Pool

Proximal input: sensory layer
Distal input: neighboring object pool layers
Represents: current object

Sensory Layer

Proximal input: sensory feature
Distal input: location from Location Layer / temporal context
Represents: feature at a location in object space

This layer needs to receive distal input in two ways: from itself as in the typical SP/TM operation defined in HTM School, and from the Location Layer. These distal inputs must be mixed.

Location Layer

Proximal input: displacement representing movement
Represents: location in 2D envrionment

When new proximal input representing movement is applied, GCM bumps move from current location to predicted location.

Create a layer that represents an XY location in the 2D grid with Grid Cell Modules. Every location must have a unique representation.

Trying to run the Python sample that fails with

No module named 'htm.bindings

Using Python38 and also tried with 3.6.

Have already imported htm and all requirements as described.

Thanks in advance

Given an XY location in the environment, this layer should represent this space in many Grid Cell Modules, which work together to resolve the object. This is not an encoder (like the HTM School videos imply). This layer must run GCMs that will respond to displacement vectors appropriately.

The only property these SDRs in this layer must have is consistent displacements between them, which will enable movements through the space to have meaning no matter where they originate.

To test this, let's assume an XY space that is 10 units square.
The SDR for location 2,2 is A, and the SDR for location 2,3 is B. The displacement A-B is d1.
The SDR for location 3,2 is C, and the SDR for location 3,3 is D. The displacement C-D is d2.

assert d1 == d2

Something like that. I need to review the Columns+ paper to provide more details.

Hi,

Trying to get the example made by @Zbysekz working but juts cant start the main.py:

(ri_venv) rdaneel (master *) python $ python main.py
Traceback (most recent call last):
File "main.py", line 14, in
from htm.bindings.algorithms import SpatialPooler
ImportError: No module named htm.bindings.algorithms

I tried to install and compile the nupic.cpp but ran into an issue: htm-community/htm.core#702

Any ideas? Suggestions?

Shouldn't be hard to hookup dataset from MNIST since now our 2D space is in yaml transparent format.

I think we should start simple with a small space like MNIST. So imagine something like 28*28 and a total of 784 input units, but topological.

Each location in this space:

• has a unique location representation
• may have a feature

We can put off specifying the feature type etc. until processing time.

If we do it this way, we can also use MNIST as a test set for object classification when we are ready, and we'll have the flexibility to decide what "features" are within each experiment.