Martijho-PathNet-search-plots

From Robin

(Difference between revisions)
Jump to: navigation, search
(Ny side: {|style="margin: 0 auto;" | [[File:600binMNIST_module_reuse_histogram.png|300px|thumb|upright|alt=module reuse plot|Distribution of module reuse in s+s and p+s searches along side the distr…)
 
(32 intermediate revisions not shown)
Line 1: Line 1:
 +
= Population Diversity =
 +
== Diversity metrics ==
 +
Three different diversity calculations have been performed, which resulted in three similar diversity messures:
{|style="margin: 0 auto;"
{|style="margin: 0 auto;"
-
| [[File:600binMNIST_module_reuse_histogram.png|300px|thumb|upright|alt=module reuse plot|Distribution of module reuse in s+s and p+s searches along side the distribution for a random module selection]]
+
| [[File:Average_population_diversity_pairwise_hamming.png|300px|thumb|upright|alt=Hamming distance|Pairwise Hamming distance: Calculated for each layer and then averaged across each path. This plot took 22 minutes and 40 seconds to generate.]]
-
| [[File:600binMNIST_training_boxplot.png|thumb|300px|upright|alt=average training|Average training each module in a path undergoes for s+s and p+s searches plotted over the amount of module reuse]]
+
| [[File:Average_population_diversity_homemade.png|thumb|300px|upright|alt=Homemade distance|Homemade diversity metric: This diversity metric resulted in a scaled version of Pairwise Hamming, but took only 36 seconds to generate. See description below]]
-
| [[Fil:600binMNIST_reuse_by_layer.png|thumb|300px|alt=reuse by layer|Amount of module reuse for each layer in s+s and p+s searches alongside amount of reuse when randomly selecting modules]]
+
| [[File:Average_population_diversity_euclidean_centroid.png|thumb|300px|alt=Euclidean distance|Average euclidean distance to centroid: Calculated by averaging the euclidean distance between every genotype in a generation, and a "centroid genotype". The plot was generated in 40 seconds]]
-
| [[Fil:600binMNIST_evaluation_vs_training.png|thumb|300px|alt=evaluation vs training|Validation accuracy for each model as a function of average training each module in the path recieved.]]
+
|}
 +
 
 +
== Home-made for each layer ==
 +
{|style="margin: 0 auto;"
 +
| [[File:Population_diversity_recomb.png|300px|thumb|upright|alt=Population diversity in recomb search|Average population diversity for each layer in generation. Experiment: Recomb]]
 +
| [[File:Population_diversity_low.png|thumb|300px|upright|alt=Population diversity in low search|Average population diversity for each layer in generation. Experiment: Low]]
 +
| [[File:Population_diversity_high.png|thumb|300px|alt=Population diversity in high search|Average population diversity for each layer in generation. Experiment: High]]
 +
|}
 +
{|style="margin: 0 auto;"
 +
| [[File:Population_diversity_low2high.png|thumb|300px|alt=Population diversity in low2high search|Average population diversity for each layer in generation. Experiment: Low to High]]
 +
| [[File:Population_diversity_high2low.png|thumb|300px|alt=Population diversity in high2low search|Average population diversity for each layer in generation. Experiment: High to Low]]
 +
|}
 +
 
 +
= Path size =
 +
{|style="margin: 0 auto;"
 +
| [[File:Average_path_size.png|300px|thumb|upright|alt=Path size|The change in average path size for each generation. ]]
 +
| [[File:Average_size_by_layer.png|thumb|300px|upright|alt=Layer size|The average size of each layer in each optimal path found.]]
 +
|}
 +
 
 +
= Capacity and Reuse =
 +
{|style="margin: 0 auto;"
 +
| [[File:Capacity_pr_generation.png|300px|thumb|upright|alt=Capacity development during search|Average number of modules used during the different searches plotted consecutively. The gray dotted line is what number of modules would be used if modules were selected randomly for each task.]]
 +
| [[File:Capacity_pr_validation_accuracy.png|thumb|300px|upright|alt=Capacity vs accuracy|Used capacity in a PathNet plotted against that networks average classification accuracy for all tasks]]
 +
| [[File:Module_reuse_pr_generation.png|thumb|300px|upright|alt=Module reuse development during search|Average reuse for each generation plotted for each experiment type. The gray dotted line is what reuse would be reached by random module selection]]
 +
|}
 +
 
 +
= Training =
 +
{|style="margin: 0 auto;"
 +
| [[File:Training_accuracy.png|300px|thumb|upright|alt=Training accuracy|Average training accuracy for each of the five search algorithms during the search for each task. The dotted line is the validation accuracy reached after the search was completed.]]
 +
| [[File:Training_value.png|thumb|300px|upright|alt=Training value|Number of training iterations in total for all modules in the PathNet plotted against the cumulative validation accuracy reached for all tasks. The size of the circle corresponds to the amount of modules used.]]
 +
| [[File:Used_training_ration.png|thumb|300px|upright|alt=Used Training ratio|Plot of the ratio of unit training in locked and stored modules on the total number of unit training in all modules for each task]]
|}
|}

Current revision as of 16:12, 28 February 2018

Contents

Population Diversity

Diversity metrics

Three different diversity calculations have been performed, which resulted in three similar diversity messures:

Hamming distance
Pairwise Hamming distance: Calculated for each layer and then averaged across each path. This plot took 22 minutes and 40 seconds to generate.
Homemade distance
Homemade diversity metric: This diversity metric resulted in a scaled version of Pairwise Hamming, but took only 36 seconds to generate. See description below
Euclidean distance
Average euclidean distance to centroid: Calculated by averaging the euclidean distance between every genotype in a generation, and a "centroid genotype". The plot was generated in 40 seconds

Home-made for each layer

Population diversity in recomb search
Average population diversity for each layer in generation. Experiment: Recomb
Population diversity in low search
Average population diversity for each layer in generation. Experiment: Low
Population diversity in high search
Average population diversity for each layer in generation. Experiment: High
Population diversity in low2high search
Average population diversity for each layer in generation. Experiment: Low to High
Population diversity in high2low search
Average population diversity for each layer in generation. Experiment: High to Low

Path size

Path size
The change in average path size for each generation.
Layer size
The average size of each layer in each optimal path found.

Capacity and Reuse

Capacity development during search
Average number of modules used during the different searches plotted consecutively. The gray dotted line is what number of modules would be used if modules were selected randomly for each task.
Capacity vs accuracy
Used capacity in a PathNet plotted against that networks average classification accuracy for all tasks
Module reuse development during search
Average reuse for each generation plotted for each experiment type. The gray dotted line is what reuse would be reached by random module selection

Training

Training accuracy
Average training accuracy for each of the five search algorithms during the search for each task. The dotted line is the validation accuracy reached after the search was completed.
Training value
Number of training iterations in total for all modules in the PathNet plotted against the cumulative validation accuracy reached for all tasks. The size of the circle corresponds to the amount of modules used.
Used Training ratio
Plot of the ratio of unit training in locked and stored modules on the total number of unit training in all modules for each task
Personal tools
Front page