I like seeing HTM run from multiple angles.

Here’s a new angle: segment lifetimes.


Quick explanation:

  • The gray rectangles show dendrite segment addition and removal.
  • The colors show segment activity.
  • It’s interactive, and there’s a learning curve.
    • Use your mouse wheel on the top chart to zoom in on a timestep.
    • Use your mouse wheel on the bottom chart to zoom in on a segment.
    • Do both! Try zooming in on different colorful areas.

Other details:

  • This is a random sample of a large number of segments. The sorted list of segments is split into a small number of equal-sized buckets, and one segment is chosen randomly from each. So it’s not totally random.
  • If you zoom way in and then out, the random sample is different every time.

In lots of time series visualizations, the y dimension shows counts. This is a useful aggregate, but it obscures some of what’s going on. When you flatten a description of an HTM into counts, you can’t track an individual column or cell or segment across time.

Another approach is to use the y dimension to show identity. Sanity’s HTM region inspector does this with HTM columns, so you can follow a single column by moving left and right. It tells stories about individual HTM columns.

This visualization tells stories about individual HTM segments.

An example: Is white noise a neuralyzer?

Let’s run an experiment:

  1. Feed an HTM some data with sequencial patterns (hotgym).
  2. Feed it random garbage.
  3. Feed it the first data again.

What happens?

We should run this experiment twice. The Python temporal memory doesn’t currently implement segment cleanup – it will add dendrite segments forever, never clearing out old ones. The C++ temporal memory implements segment cleanup, so it’s worth testing how this impacts results.

The base algorithm: temporal_memory.py

Using the top chart, you can see where each of the following happened:

  • It received 1000 steps of orderly data.
  • Then it received 1000 steps of totally unpredictable data.
  • Finally it received 1000 steps of the same orderly data.

Using the bottom chart, you can see:

  • It grew 4500 segments which tended to be useful.
  • Then it grew 73000 segments which tended to never be used.
  • Finally it grew 500 useful segments, but mostly relied on the original 4500.
    • (Zoom and pan to the bottom of the chart to see these 500.)

The white noise filled the temporal memory with junk, but its early knowledge was preserved.

Now add segment cleanup: TemporalMemory.cpp

This Temporal Memory uses the following segment cleanup rules:

  • When a segment’s last synapse is removed, the segment is removed.
  • When a cell exceeds its maximum allowed number of segments, it removes its least recently active segment. The segment’s birth timestep is used as its initial “last active” timestep.

It’s good to see that this Temporal Memory cleans up its junk segments, but there’s clearly a huge flaw here. The entire set of useful segments is destroyed to make room for new segments. On timestep 2000, it has to relearn everything. With these cleanup rules, white noise will make an HTM forget everything it knows.

More opportunities

This graphic uses a fixed sort order, showing a semi-random sample of segments. This is really useful.

I’m excited to see what happens when we add creative sorting. You might enjoy my sorting demo. Imagine selecting a timestep and saying “sort by oldest segment”. It wouldn’t have to do a random sample, it could just show the top ~50. You could jump around, performing various sortings at various timesteps.

Using a random sample gives you a picture of typical tendencies within the population. But you can’t look at it and know for sure that, for example, no segments survived the white noise. You’d have to zoom and look at every segment to be confident. You can’t really solve this problem by showing aggregates, either. If, say, 10 segments survived, would you be able to see them in a chart that’s showing 80000 segments? This is a job for sorting.

Also remember

HTM dendrite segments don’t necessarily represent the same thing forever. If the segments’ connections tend to change dramatically, is this graphic still useful? Is there a way to see whether / how much this is happening? It’d be fun to explore this area.

Appendix: How I got this data

Here’s my Jupyter notebook:

This notebook is really cool. In it, I generated these graphics inline.

I used temporal_memory.py from NuPIC at 23c6686. For TemporalMemory.cpp, I used a private version that includes a few changes which make it faster, more visualizable, and which fix a few bugs (without changing the existing segment cleanup algorithm).