spwnn - reading list

spwnn - the "shallow pair-wise neural network"

My articles about it in chronological order:

spwnn - 40 years in the making

spwnn - main algorithm at github

spwnn - the fundamental question

spwnn - fun with letter pairs

spwnn - fun with golang

spwnn - iterating quickly to finish this thing

spwnn - big 370K word dictionary

spwnn - awkward word anana

spwnn - fun with aws lambda

spwnn - embarrassingly parallel

spwnn - delightfully parallel

spwnn - aws sweet spot

spwnn - more cpu core scaling tests

Programming Retreat

I've mentioned I did some work on spwnn during a programming retreat in 2018.

The place I stayed was in the middle of this satellite view:

The internet was provided by an old-school (not Starlink) satellite.  It was just good enough that I could look up articles about Go and AWS but not fast enough to be distracted by anything else.

Sometimes I would have visitors in the evening.

Each day I would drive to Clifton, Az, for lunch and coffee.

You can rent it if you'd like (click picture to go to AirBnB posting).

spwnn - scaling

The dictionary size has a big impact on the speed of spwnn (spwnning?).

c7g.16xlarge
dict size	time (seconds)	words / sec
23,115	1.47	15,768
49,167	6.13	8,024
370,104	388.47	953

This could be for many reasons.  One is that the number of words traversed per letter pair is much larger.  Another could be exceeding the cache size which slows down memory access.

Every time I think I'm done measuring things I think up something else.

All right, all right.  Here are the increases in times to validate the dictionary vs. the max size of the neurons:

time vs. 23k	time vs len
1.0	1
4.2	3.257148909
265.0	16.02753654

That doesn't directly correlate, so most likely the big dictionary exceeds the cache size.