AIGS538 Deep Learning Project : Analyzing the Impact of Extracting Representative Data on Deep Learning Model Performance

• Our motivation was to reduce the number of data so that we only use key representative samples during training. We used 5 algorithms to reduce the datasets ; interval mean, k-medoid, k-mean, min-max, and random reduction.
• If you want to run each reduction algorithm, you can run ipynb files below including the name.

Code List

• 딥러닝 interval mean 정리.ipynb
• 딥러닝 k_medoid.ipynb
• 딥러닝 k mean 정리.ipynb
• 딥러닝 max_min.ipynb
• 딥러닝 random reduction.ipynb

Description 1

• K-means, Interval mean code description
• You can change the capacity in 4-3) fixed capacity results, and the fixed capacity used in this research is 2 layers.
• You can change the number of reduced datasets in 4-2) dataloader with reduced dataset, and the range of number of datasets is from 3~200.

1. Setting
   • Import Required Libraries
   • Hyperparameters
2. Data
   • Generate data
   • Dataset reduction
     • (a) reduction algorithm
       • : interval mean or kmean
     • (b) dataloader with reduced dataset
     • (c) plot reduced dataset
3. Training with reduced samples
   • Define model
   • Training
     • (a) Learning curve
     • (b) True Function vs Model Prediction
   • Model Selection
     • (a) Validation curve
4. Results
   • Example
   • Whole result (with optimal capacity)
     • (a) train, evaluate, test whole reduced dataset
     • (b) plot whole reduced dataset
   • Whole result (with fixed capacity = 2)

Description 2

• K-medoid, max-min, random reduction description
• You can change the capacity in 3-2) Training, # of hidden layer varaible 'i', and the fixed capacity used in this research is 2 layers.
• You can change the number of reduced datasets in 3-2) Training, argument of training reduction as variable 'num', and the range of number of datasets is from 3~200.
  • In case of random reduction, you have to change # of reduced datasets after 2**-2) Dataset Reduction**, using 'num' variable

1. Settings
   • Import required libraries
   • Hyperparameters
2. Data
   • Generate data
   • Dataset reduction
     • : K-medoid, min-max, random reduction
3. Training with reduced samples
   • Define model
   • Training
   • Learning curve
   • True function vs Model Prediction
   • Model selection
   • Results