In a system where size in bytes of an SDR is of concern (such as synchronizing between multiple threads, or transmitting over a network), the thought occurred to me that indices to the 1 bits could be relative to each other. If the space is too large to fit into the data type used, noise could be a…

A slightly different take on the same issue … [image] Path based SDRs (Dendrite Trees) Numenta Theory In an earlier post, I mentioned that I have been trying to work out how to form the cortical.io retina using HTMs. This effort has led me to work through what operations …

The offset is limited by the number of input bits. It is easy to store the on indices as Int16 or Int32, and don’t impact the memory usage practically speaking, since not every SDR will be stored, just the currently produced one. Using offset adds to processing and is a limited optimisation.

Let’s try some numbers and see if this is important or not. Start with an image processing application, using a modest 1K x 1K map. Each cell may have 8 dendrites for proximal and 16 for distal, Each dendrite may have a large number of potential synapse locations, say 1024. The data structure for…

[image] abshej: The offset is limited by the number of input bits. It is easy to store the on indices as Int16 or Int32, and don’t impact the memory usage practically speaking, since not every SDR will be stored, just the currently produced one. Using offset adds to processing and is a limited o…

I see. I am not sure. It would still make sense to store them in larger data types instead of processing them to obtain individual large ones.

Another way to think about it, is imagine yourself a node located in Mumbai, part of a large HTM network spanning the globe. You are responsible for doing temporal memory, and you are relying on a node located in Redwood City to tell you what were all the active cells at timestep T in the layer yo…

[image] abshej: How is the path stored? What is used to define it? I thought this seemed obvious - my bad. The paths are stored in an array just like any other data structure in the program. I am using random walks with direction bias to generate the paths now. I am thinking of adding a br…

BTW, the numbers @Bitking broke down could be directly applied to the classic Spatial Pooling algorithm with very little effort. A single index to one of a list of pre-defined paths (or a seed to define it procedurally) and list of permanence values would have a notably smaller memory footprint com…

[image] Paul_Lamb: [2,14,238,13,242,4,239,11,257,3 …] In this case, lets say we want to encode the indices as uint8, which has a rang of 0 - 255. One of the values (257) is too large. However, if we insert a single bit of noise, then we could use uint8 (thus halving the footprint compared to …

Trading byte size for noise

Engineering

abshej February 25, 2018, 9:01am 5

How is the path stored? What is used to define it?

Path based SDRs (Dendrite Trees)

Topic		Replies	Views
Memory layout strategies for a Cell Engineering	11	1566	March 23, 2021
Developing a better NuPIC. Need suggestions Implementations htm-implementations , community , projects	17	1594	July 4, 2018
Most frequent SDR overpowering all other bits Engineering encoders , question , community	21	713	April 15, 2024
The Principle of Temporal Memory Numenta Theory	21	852	April 25, 2021
Triadic Memory — A Fundamental Algorithm for Cognitive Computing Related Papers sparse-distributed-m	282	7200	October 18, 2023

Trading byte size for noise

Related topics