the_htm_spatial_pooler_a_neocortical_algorithm_for_online_sparse_distributed_coding

Spatial Pooler Research

This repository contains research scripts for the following paper:

The HTM Spatial Pooler—A Neocortical Algorithm for Online Sparse Distributed Coding https://www.frontiersin.org/articles/10.3389/fncom.2017.00111

The goal is to demonstrate various computational properties of SP with artificial and real dataset.

SP metrics

The SP performance is quantified using a set of metrics

• entropy, which measures efficient use of columns
• stability, the same inputs should map to the same outputs even with continuous learning
• input-output overlap, which measures the noise robustness of SP output
• classification accuracy, which measures whether the SP outputs are classifiable

Datasets

A variety of input datasets were considered, such as

• random SDRs ('randomSDR')
• random SDRs with varying sparsity ('randomSDRVaryingSparsity')
• correlated SDRs from two input fields ('correlatedSDRPairs')
• random bars (horizontal or vertical) ('randomBarPairs')
• random crosses ('randomCross')

Usage

Spatial Pooler Without Topology

• random SDRs, fixed sparsity

python train_sp.py -b 1 -d randomSDR --spatialImp py --trackOverlapCurve 1 -e 50

• random SDRs, varying sparsity

python train_sp.py -b 1 -d randomSDRVaryingSparsity --spatialImp cpp --trackOverlapCurve 1 -e 100 --name randomSDRVaryingSparsityNoTopology

• Continuous learning experiment

  • Train SP on random SDR dataset until converge
  • Then switch to a different dataset
  • The SP should adapt to the new dataset

python train_sp.py -b 1 -d randomSDRVaryingSparsity --trackOverlapCurve 1 --name continuous_learning_without_topology --spatialImp cpp --changeDataSetAt 50 -e 120

• Fault tolerance experiment (no topology)

  • Train SP on random SDR dataset until converge
  • kill a fraction of the SP columns

python train_sp.py -b 1 --name trauma_boosting_without_topology --spatialImp faulty_sp --killCellsAt 50 --killCellPrct 0.5 --trackOverlapCurve 1

• Random Bar Pairs Vs. Random Cross (No Topology)

python train_sp.py -d randomBarPairs -e 200 -b 1 --showExampleRFs 1 python train_sp.py -d randomCross --spatialImp monitored_sp -e 200 -b 1 --showExampleRFs 1

• Two input fields with correlated SDR pairs

python train_sp.py -b 1 –d correlatedSDRPairs

Spatial Pooler With Topology

• Random SDRs with varying sparsity

python train_sp.py -t 1 -b 1 -d randomSDRVaryingSparsity --spatialImp cpp --trackOverlapCurve 1 -e 100 --name randomSDRVaryingSparsity

• Continuous learning experiment

python train_sp.py -t 1 -b 1 -d randomSDRVaryingSparsity --spatialImp cpp --trackOverlapCurve 1 -e 120 --changeDataSetAt 50 --name randomSDRVaryingSparsityContinuousLearning

Fig. 2A, 2C will be generated automatically Fig. 2B requires python plot_noise_robustness.py Fig. 2D, Fig. 3E requires python plot_noise_robustness.py

Fig. 2B, 2E, Fig. 3 * Run runRepeatedExperiment.sh first * Run plot_traces_with_error_bars.py

• Random Bar Pairs Vs. Random Cross (With Topology)

Fig. 4A, 4B python train_sp.py -t 1 -d randomCross --spatialImp py -e 200 -b 1 --showExampleRFs 1 python train_sp.py -t 1 -d randomBarPairs --spatialImp py -e 200 -b 1 --showExampleRFs 1

• Random bar sets (more than two random bars per input)

python train_sp.py -t 1 -d randomBarSets -b 1 --name random_bars_with_topology \ --spatialImp monitored_sp --changeDataSetContinuously 1 --boosting 1

MNIST experiment

Fig. 4C python train_sp.py -t 1 -d mnist --spatialImp py -e 2 -b 1 --showExampleRFs 1

Example receptive fields will be saved in figures/exampleRFs/dataType_mnist_boosting_1_seed_XX To download the MNIST data, please follow the instructions in the nupic.vision repo: https://github.com/numenta/nupic.vision/tree/master/src/nupic/vision/mnist/data

• install nupic.vision
• run ./build.sh in nupic.vision/src/nupic/vision/mnist/data/
• run python extract.py
• Move data/training/ and data/testing/ should be moved to sp_paper/data/mnist/

Fault tolerance experiment (with topology)

• Fault tolerance to SP column death (Results for Fig. 5A-C, upper panel) Train faulty_SP on random bar set dataset

python train_sp.py -t 1 -d randomBarSets -b 1 --name trauma_boosting_with_topology \ --changeDataSetContinuously 1 --spatialImp faulty_sp --killCellsAt 180 \ --trackOverlapCurve 0 -e 600 --checkTestInput 1 --checkRFCenters 1 \ --saveBoostFactors 1 --checkInputSpaceCoverage 1

• Fault tolerance to input afferents death (Results for Fig. 5A-C, lower panel)

python train_sp.py -t 1 -d randomBarSets -b 1 --name trauma_inputs_with_topology \ --changeDataSetContinuously 1 --spatialImp faulty_sp --killInputsAfter 180 \ --trackOverlapCurve 0 -e 600 --checkTestInput 1 --checkRFCenters 1 \ --saveBoostFactors 1 --checkInputSpaceCoverage 1

• Analyze fault tolerance experiment results (Fig. 5)

mkdir figures/traumaMovie/ python analyze_trauma_experiment.py -e [expname] cd figures/traumaMovie/ ffmpeg -start_number 100 -i [expname]_frame_%03d.png traumaMovie_[expname].mp4

expname should be trauma_inputs_with_topology_seed_41 or trauma_boosting_with_topology_seed_41

NYC Taxi experiment

• Run with random SP (no learning, no boosting)

  python run_sp_tm_model.py --trainSP 0

• Run with learning SP, but without boosting

  python run_sp_tm_model.py --trainSP 1 --boostStrength 0

• Run with learning SP, and boosting

  python run_sp_tm_model.py --trainSP 1 --boostStrength 20

• plot results (Fig. 6) python plot_nyc_taxi_performance.py