Hierarchical Temporal Memory of a smoker

Modifications

Temporal memory

Bursting

• Modification of bursting unpredicted column. I activate only winner cell
• Fortunately, the model haven't lost its noise resistance.
• For more statistic confidence of trainability, I plan to add some random in bursting column.
• Besides, it can make sense to burst all the column while learn = false.

Module

• Being in the same folder with HTM.so you can import it with

import HTM

or

from HTM import temporal_memory

Temporal Memory API:

• init:

tm = temporal_memory(
    (1, 2), # column_dimensions
    1,      # cells_per_column, default = 32
    1,      # min_threshold, default = 1
    1,      # activation_threshold, default = 2
    0.4,    # initial_permanence, default = 0.21
    0.5,    # connected_permanence, default = 0.5
    0.1,    # permanence_increment, default = 0.1
    0.1,    # permanence_decrement, default = 0.1
    0.05,   # predicted_segment_decrement, default = 0.0
    2,      # max_segments_per_cell, default = 255
    2       # max_synapses_per_segment, default = 255
)

• computation/learning:

tm.compute([[0, 1]], True) # computation of 2-d SDR. The second parameter is learning

• evaluating model:

tm.get_anomaly() 
# the percentage of unpredicted active cells and wrong predicted inactive cells 

• synapses list:

tm.print_connections() # == tm.print_connections(False), 
# prints only connected synapses
tm.print_connections(True)
# prints all potential synapses

# The bool argument is all_synapses

• Warning! kwargs don't work. Sorry:( You can NOT to write parameters this way:

tm = temporal_memoty(
    column_dimensions=(1, 2),
    cells_per_column=4,
    # ...
)