leotiez/htmresearch

Experimental algorithms. Unsupported.

Incorporating A Continuous Bayesian Learning Rule Into A Discrete Hierarchical Temporal Model

This repository contains the Python code that was used to extend Numenta's HTM approach to a continuous stochastical network that embeds a Hebbian-like Bayesian learning rule to adapt to changing input statistics. The project was conducted within the research class DD2465 at KTH, and the report can be found here.

The goal was to compare the approaches regarding their respective convergence. We hypothesised that our network achieves similar results, while increasing interpretability and reducing the number of model parameters that are needed to be tuned.

Unfortunately, we could not verify our claim. The algorithm comes with the drawback of large computational costs for both memory and time. Moreover, we could not obtain a similar performance as the original model (measured in number of sensation-location pairs needed to infer an object correctly).

Nevertheless, we are convinced that the work opens a new way to investigate thoroughly the dynamical behaviour of the network. Moreover, the output activity of the neurons of the object representation can be seen as a measure of confidence. Thus, we believe that with more time it is possible to tune the performance of the Bayesian version to be similar to the original approach, whilst increasing interpretability.

The original README that was published by Numenta can be found here

Installation

To install the project, change to the project directory and run

python2 -m pip install .

Execution

The first file for a large parameter search can be executed via

python experiments.py

To create the plots, run the command

python tests/bayesian/bayesian_conergence_activity_plot.py

Contribution

Please feel free to contribute to the work of Numenta and to our own adaption. Clone, fork and try. For hints and recommendations, don't hesitate to contact us via email: [email protected] and [email protected].

htmresearch's Issues

Segment count without duplications

At the moment we add a segment each time a column bursts.
However, we only need to add a segment if the cell has less than the current max segments used.
We could keep track of the number of segments used per cell and then compare if an increment is needed.

Experiments/Cleanup/Todos 17/11 Discussion

Issue list from discussion 17/11

[E] = Test in Experiments
[C] = Cleanup
[T] = Todos

• [T] Compute method in apical temporal memory
• [T] Getter/Setter redefined to new variables
• [C] Initial permanence -> can be removed?
• [E] Basal/Apical weights -> one segment per cell randomly initialized -> better results? (e.g. NN convergence) or hinderly?
• [E] Use apical input: (1) as tiebreak, (2) not at all, (3) same weights as basal (bias only needed once)
• [E] Normalization: Threshold is (1) before (independent thresholding), (2) after (competition based on threshold), (3) in-between (2-times normalize)

Include apical information in TM