helins.htm.space — io.helins/htm 0.0.0-alpha0

Spatial pooling.

Based on :

Byrne, F. (2015). Encoding reality: Prediction-assisted cortical learning algorithm in hierarchical temporal memory. arXiv preprint arXiv:1509.08255.

Cui, Y., Ahmad, S., & Hawkins, J. (2017). The HTM spatial pooler—a neocortical algorithm for online sparse distributed coding. Frontiers in computational neuroscience, 11, 111.

Hawkins, J. et al. 2016. Biological and Machine Intelligence. Release 0.4. Accessed at http://numenta.com/biological-and-machine-intelligence/.

Leake, M., Xia, L., Rocki, K., & Imaino, W. (2015, January). Effect of Spatial Pooler Initialization on Column Activity in Hierarchical Temporal Memory. In AAAI (pp. 4176-4177).

Mnatzaganian, J., Fokoué, E., & Kudithipudi, D. (2017). A mathematical formalization of hierarchical temporal memory’s spatial pooler. Frontiers in Robotics and AI, 3, 81.

Pietroń, M., Wielgosz, M., & Wiatr, K. (2016, October). Formal analysis of HTM spatial pooler performance under predefined operation conditions. In International Joint Conference on Rough Sets (pp. 396-405). Springer, Cham.

Glossary :

active-inputs Sequence of inputs active as a result of some stimulus.

active-minicols Sequence of minicols active as a result of active-inputs.

center-input When relevant, a minicol might have a natural center in the input space.

cnx, connection A connection is made when perm >= `cnx-threshold``.

cnxs Vector where indices are minicols and item are vectors of inputs with established connection.

cnx-threshold, connection-threshold Treshold for deciding if a perm is strong enough for establishing a cnx.

flat-index Cf. flat-index in helins.htm.grid namespace.

grid Cf. grid in helins.htm.grid glossary.

grid-inputs grid describing the topology of the input space.

grid-minicols grid describing the topology ofminicols`.

inhibition-radius Radius around a minicol within grid-minicols where inhibition happens during local inhibition.

input Input bit represented by its flat-index within the grid-inputs.

input-pool-mapping Vector where indices represent inputs and items are 2-tuples where the first element is a minicol and the second one is the offset of the corresponding input in that minicol's pool.

minicol Modelisation of a neocortical mini-column represented by its flat-index within the grid-minicols.

n-pool Size of a pool.

overlap-score Number of currently active-inputs a minicol is currently connected to.

overlap-scores Vector of overlap-scores where each index represent minicol;

perm, permanence Value between 0 and 1 (inclusive) for deciding, in conjunction with a cnx-threshold, if a cnx is established.

perms Vector of perms layed out in the same order as their corresponding pool.

perm-table Vector of perms where each index represents a minicol. perms are layed out in the same order as corresponding pools.

pool, potential-pool Set of inputs a minicol can potentially connect to.

pools, potential-pools Vector of pools where each index represents a minicol.

receptive-field Span of inputs within grid-inputs a minicol is currently connected to.

stimulus-threshold During inference, in order for a mini-column to even be considered potentially active, it must have an overlap-score of at least that much. During initialization, it is important to garantee that at least stimulus-threshold connections are randomly established for each mini-column otherwise they will never have the chance to compete. This parameter is a measure against noise and should be low. It could even be 0.

active-duty-cycles clj

(active-duty-cycles n-minicols)

adapt-perms clj

(adapt-perms perm-
             perm+
             n-inputs
             pools
             active-inputs
             perm-table
             active-minicols)

Spatial learning, adapts permanences of active-minicols.

avg-inputs-minicols-ratio clj

(avg-inputs-minicols-ratio grid-inputs grid-minicols)

Computes, on average, the number of mini-columns existing for every input bit.

Needed for computing the inhibition-radius.

boost-factors clj

(boost-factors n-minicols)

cnxs clj

(cnxs cnx-threshold pools perm-table)

Returns a vector where indices are minicols and items are vectors of inputs with established connections.

global-inhibition clj

(global-inhibition overlap-scores n-active)

Using overlap-scores of mini-columns, performs global inhibition by selecting n-active i-minicols with the best overlap.

global-pool clj

(global-pool grid-inputs n-pool)

(global-pool grid-inputs n-pool rng)

Creates a pool by sampling inputs from the whole grid-inputs.

inhibition-radius clj

(inhibition-radius grid-inputs avg-inputs-minicols-ratio cnxs)

Computes the inhibition-radius for local inhibition.

input-pool-mapping clj

(input-pool-mapping n-inputs pools)

Given pools, returns perm-mapping.

local-inhibition clj

(local-inhibition grid-minicols n-active inhibition-radius overlap-scores)

Using overlap-scores of mini-columns, performs a local inhibition.

local-pool clj

(local-pool grid-inputs potential-radius n-pool grid-minicols minicol)

(local-pool grid-inputs potential-radius n-pool grid-minicols minicol rng)

Creates a pool for the given minicol by sampling inputs from a hypercube around the minicol's corresponding center-input in grid-inputs.

overlap-scores clj

(overlap-scores cnx-threshold perm-table input-pool-mapping active-inputs)

Given active-inputs, returns overlap-scores.

perms clj

(perms n-pool n-cnxs cnx-threshold cnx-delta)

(perms n-pool n-cnxs cnx-threshold cnx-delta rng)

Returns a sequence of perms size n-pool where n-cnxs values are above cnx-threshold. All values are within cnx-threshold +/- cnx-delta.

Keep in mind n-cnxs should be >= stimulus-threshold later used otherwise the minicol associated with those intiial perms will not have a chance of ever be considered active.

receptive-field clj

(receptive-field grid-inputs minicol-cnxs)

Given a list of inputs with established connections, computes the receptive-field of a minicol.

update-active-duty-cycles clj

(update-active-duty-cycles period active-minicols active-duty-cycles)

update-duty-cycle clj

(update-duty-cycle period duty-cycles updates)