Abstract

This publication presents two deep learning projects focused on Natural Language Processing (NLP) and Computer Vision. The first project uses LSTM-based sequence modeling to summarize text, while the second applies Convolutional Neural Networks (CNN) for handwritten digit classification using the MNIST dataset. Both models are implemented using TensorFlow/Keras and demonstrate strong performance on their respective tasks.

Introduction

Artificial Intelligence (AI) has revolutionized the way we interact with data. Text summarization helps in reducing information overload, and digit recognition enhances automation in banking, postal services, and digital forms. This publication combines both techniques into a single repository:

Text Summarization using LSTM (Long Short-Term Memory networks)
Handwritten Digit Recognition using CNN (Convolutional Neural Networks)

Methodology

Text Summarization using LSTM
Model Type: Encoder-decoder LSTM sequence-to-sequence model
Preprocessing: Tokenization, padding, word embedding
Framework: TensorFlow/Keras
Objective: Generate a concise summary from a longer input text

Handwritten Digit Recognition using CNN
Dataset: MNIST (60,000 training images, 10,000 test images)
Model Architecture: Conv2D → MaxPooling → Flatten → Dense layers
Framework: TensorFlow/Keras
Objective: Predict the digit (0–9) from input grayscale image (28x28)

Experiments

Text Summarization
Trained on AI-related text samples
Model prints: Text summarization model training complete!
Output: Predicted summaries for new input sentences

Results

Handwritten Digit Recognition
Achieved accuracy: ~97.2%
Output: Predicted digit label with visual output
Sample result: Test accuracy: 97.2%
Predicted: 4

Conclusion

The combined project demonstrates the versatility of deep learning in solving diverse AI problems using LSTM for NLP and CNN for vision. It highlights how TensorFlow/Keras simplifies the development of end-to-end models with high accuracy and reproducibility. Future improvements could include using pre-trained embeddings for summarization and model optimization using dropout or batch normalization for CNNs.