/MNIST-Digit-Recognizer-CNN-Keras-99.66

Used the Dataset "MNIST Digit Recognizer" on Kaggle. Trained Convolutional Neural Networks on 42000 Training Images and predicted labels on 28000 Test Images with an Validation Accuracy of 99.52% and 99.66% on Kaggle Leaderboard.

Primary LanguageJupyter NotebookMIT LicenseMIT

MNIST-Digit-Recognizer-CNN-Keras-99.66

MNIST DIGIT RECOGNIZER COMPETITION on Kaggle.

  1. The Training Dataset has 42000 Images each of 784 pixels representing Digits from 0-9.
  2. The Test Dataset has 28000 Images.
  3. I have reshaped the Training and the Test Samples into 28 x 28 pixels for further processing.
  4. I have used Convolutional Neural Networks to Train the Data.
  5. After every layer, I have used Batch Normalization to transform the data points to have Zero Mean and Variance One.
  6. I have used Dropout layers to avoid Overfitting.
  7. I have performed Data Augmentation and generated random images with ImageDatagenerator to avoid Overfitting.
  8. I have used LearningRateScheduler after every Epoch to Reduce the Learning Rate in order to converge to Local Optimum.
  9. Then, I plot the Confusion Matrix to Evaluate the Performance of the model on the data.
  10. Lastly, I have generated csv file and made submission on Kaggle and acheived 99.66% Accuracy on its Public Leaderboard.