Giters

numenta / htmresearch

Experimental algorithms. Unsupported.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Update the Python ETM to use the new "walk the columns" strategy

mrcslws opened this issue · comments

commented

Here's the current status of our temporal memory implementations:

C++ Python
TemporalMemory New New
ExtendedTemporalMemory New Old (phases)

The change here is to update the Python ExtendedTemporalMemory to use the new approach. At the end of this, the ETM algorithm tests should pass, and hopefully the Python and C++ ETM will behave identically.

The new approach walks each column once, doing all cell activation and learning for that column, then moving on to the next. It uses sorted lists of columns, cells, and segments, and it walks them all simultaneously, which avoids doing set lookups. It avoids building a pool of "learning segments", making it much easier to understand exactly where synapses are grown / reinforced / punished. This new approach is also easy to parallelize, should we ever want to.

Benefits of this change:

  • Faster code
  • Cleaner code (according to Marcus)
  • Fixes #582
  • Uses the updated Python Connections class which supports synapse/segment cleanup.
  • Unifies our two ETM implementations so that we can test them by verifying that they get identical results.
  • Removes a consumer of temporal_memory_phases and connections_phases, which we want to remove.

Here's how someone might do this:

  1. Copy-paste temporal_memory.py into extended_temporal_memory.py, and rename it.
  2. Change the code the same way that we changed the C++ TM to the C++ ETM.
  3. Copy-paste temporal_memory_test.py into etm_unit_test.py and change it to use the ETM instead of TM.

Notes:

  • I changed the C++ Connections to have a "SegmentExcitationTally" that could be used by the C++ ETM instead of computeActivity. I don't think this change will be necessary in the Python -- we can just concatenate lists together via itertools.chain (example) and pass it into computeActivity.
  • Definitely use our Python groupby2 function which is already used in temporal_memory.py. The current implementation is ready for the ETM.
commented

Before starting on this we should fix the tests broken by the recent maxNewSynapseCount change here. That is, running the tests in tests/extended_temporal_memory/ results in multiple failures. @mrcslws has confirmed this. @subutai do you have any suggestions or should I go ahead and fix these tests?

commented

We split out the two synapse change issues for temporal_memory:

numenta/nupic-legacy#3268
numenta/nupic-legacy#3292

For temporal_memory_phases the change is still in there - I'm working through it slowly as I might need that change for ColumnPooler.