Skip to content

Latest commit

 

History

History
31 lines (21 loc) · 943 Bytes

README.md

File metadata and controls

31 lines (21 loc) · 943 Bytes
Raw

Alphabet Recognition - Neural Network

Experiment to create an artificial neural network in Typescript to recognize capital letters.

Inspired by a video by Sebastian Lague: How to Create a Neural Network (and Train it to Identify Doodles)

Installation

  1. Clone Repo
  2. Install Dependencies (npm install)
  3. Create new folder : data
  4. Put dataset files with this scheme :
    • data/train/[letter]\_[index].png
    • data/test/[letter]\_[index].png
  5. Run CreateDatase() to create dataset.json

Training Test

The first training session used Huggins Face's dataset : pittawat/letter_recognition

Params :

  • Layers : [784, 100, 26]
  • LearnRate : 0.24
  • Activation Function : ReLU
  • Perturbations : 4%

Results (816 iterations) :

  • TrainDataset : 93%
  • TestDataset 91%