Deep Learning Learning Journey

This repository is a personal learning project where I explore and implement various aspects of Deep Learning — starting from basic machine learning models to advanced neural network architectures.
The aim is to build a solid foundation, experiment with different techniques, and understand their practical applications.

---

📂 Project Structure

1. Basic Machine Learning Models

Located in the OtherMLModels/ folder.

Currently implemented:

📄 Support Vector Machines

Path: OtherMLModels/SVMs/

• SVMs.pdf — A detailed write-up explaining the theory behind Support Vector Machines, including:
  • Mathematical formulation
  • Kernel functions
  • Margin maximization
  • Soft vs. hard margin SVMs
  • Real-world applications

📄 Naive Bayes Classifier

Path: OtherMLModels/Naive-bayes/

• NaiveBayes.pdf — Detailed explanation of the Naive Bayes algorithm, including:

  • Probabilistic model fundamentals
  • Conditional independence assumption
  • Formula derivation
  • Pros and cons
  • Real-world use cases

• NaiveBayes.ipynb — Jupyter Notebook containing code to implement and analyze Naive Bayes on the Breast Cancer Wisconsin dataset, including:

  • Data preprocessing
  • Model training
  • Evaluation metrics
  • Performance analysis and visualization

📄 Multi-Layer Perceptron (MLP)

Path: OtherMLModels/mlp/

• MLP.pdf — In-depth explanation of Multi-Layer Perceptrons, including:
  • Network architecture and activation functions
  • Backpropagation and optimization
  • Overfitting and regularization techniques
  • Experimental results on CIFAR-10 and CIFAR-100 datasets
  • Insights from accuracy/loss curves

2. Convolutional Neural Networks (CNNs)

Located in the CNNs/ folder.

📄 Image Captioning with CNN + RNN

Path: CNNs/image_captioning_with_CNN/

• Image_captioning_with_RNNs.pdf — Detailed study of an image captioning system that integrates:

  • CNN (InceptionV3) for high-level image feature extraction.
  • RNNs (LSTM and GRU) for sequential caption generation.
  • Use of GloVe word embeddings for semantic-rich input representations.
  • Preprocessing steps for image and text data (tokenization, vocabulary building, embedding).
  • Comparison of model architectures:
    • Baseline LSTM
    • Baseline GRU
    • Stacked GRU (3 layers + dropout)
    • Stacked LSTM (3 layers + dropout)
  • Experimental evaluation using BLEU scores and qualitative analysis of generated captions.
  • Discussion on optimizer choices (Adam, SGD, RMSprop) and overfitting mitigation techniques.

• image_caption.ipynb — Jupyter Notebook implementation for:

  • Data preprocessing (Flickr8k dataset)
  • Feature extraction with InceptionV3
  • Caption generation with different RNN architectures
  • Model training and evaluation with BLEU scores

📄 Comprehensive Comparison of CNN Architectures

Path: CNNs/Comprehensive_comparision_CNNs/

• Writing/ — Contains the detailed research study (Comprehensive_comparison_CNNs.pdf) covering:

  • Comparative analysis of five ResNet-18-based models:
    • Base CNN (vanilla ResNet-18)
    • Local Soft Attention CNN
    • Global Soft Attention CNN
    • Hard Attention CNN (MetaDOCK-based kernel selection)
    • Omni-Directional CNN (ODConv)
  • Task coverage:
    • Image Classification on Tiny ImageNet
    • Image Segmentation on Pascal VOC 2012
    • Time Series Analysis on UCR Adiac dataset
  • Findings:
    • Attention mechanisms and dynamic convolutions consistently outperform baseline CNNs in accuracy and adaptability.
    • ODConv achieved the highest classification accuracy (73.4%) and mIoU (73.09%) in segmentation.
    • Dynamic CNNs improved time-series classification (mean accuracy 0.653) over base CNN (mean 0.571).
  • Discussion of:
    • Efficiency vs. computational cost trade-offs (FLOPs analysis)
    • Task-specific advantages of different attention mechanisms
    • Future directions in dynamic CNN optimization.

• code/ — Implementation of all CNN variants:

  • Model definitions for Base CNN, Local & Global Soft Attention, Hard Attention, and ODConv.
  • Training pipelines for classification, segmentation, and time series tasks.
  • Evaluation scripts for mIoU, accuracy, and FLOPs computation.

---

🎯 Goals of the Project

This project is designed as a comprehensive learning journey through the different stages of machine learning and deep learning, with the following objectives:

1. Build Strong Foundations

   • Start with classic machine learning algorithms (e.g., SVM, Naive Bayes) to understand the fundamentals of supervised learning, probabilistic reasoning, and evaluation metrics.

2. Transition to Neural Networks

   • Implement and study basic deep learning models like Multi-Layer Perceptrons (MLPs) to bridge the gap between traditional ML and advanced neural architectures.

3. Explore Convolutional Neural Networks (CNNs)

   • Study CNN theory, implement standard architectures, and experiment with their applications in image classification, segmentation, and other domains.

4. Investigate Advanced CNN Variants

   • Conduct a detailed comparative study of dynamic CNNs, including attention-based models (local, global, hard) and Omni-Directional CNNs, to evaluate their performance across multiple tasks.

5. Integrate Multi-Modal Deep Learning

   • Develop an image captioning system combining CNNs for feature extraction and RNNs (LSTM, GRU) for natural language generation, learning how to merge vision and language models.

6. Hands-on Experimentation & Analysis

   • Implement models from scratch, train on real-world datasets, evaluate with relevant metrics (accuracy, mIoU, BLEU), and document findings for each experiment.

7. Understand Trade-offs in Model Design

   • Analyze the balance between performance, computational cost (FLOPs), and model complexity when choosing architectures for different tasks.

8. Encourage Reproducibility & Knowledge Sharing

   • Maintain well-documented code, structured project folders, and detailed theory write-ups to help others replicate experiments and learn from them.

---

🛠 Technologies Used

This project leverages a combination of programming languages, frameworks, and tools for implementing, training, and evaluating models across different machine learning and deep learning tasks.

Programming Languages

• Python 3.x — Primary language for all implementations and experiments.
• Jupyter Notebook — Interactive environment for code, visualizations, and documentation.

Core Libraries & Frameworks

• NumPy — Numerical computations and array operations.
• Pandas — Data loading, cleaning, and preprocessing.
• Matplotlib / Seaborn — Data visualization.
• scikit-learn — Classic ML algorithms (SVM, Naive Bayes, etc.) and preprocessing utilities.
• TensorFlow / Keras — Deep learning model building, training, and evaluation.
• PyTorch (optional/future) — Alternative deep learning framework for experimentation.

Model-Specific Tools & Techniques

• GloVe Embeddings — Pre-trained word embeddings for semantic-rich text representation in NLP tasks.
• InceptionV3 — Pre-trained CNN for image feature extraction in image captioning.
• ResNet-18 — Backbone architecture for CNN and dynamic CNN experiments.
• Attention Mechanisms — Local Soft Attention, Global Soft Attention, Hard Attention.
• Omni-Directional Convolution (ODConv) — Rotation-invariant CNN variant.
• Dynamic Convolutions — Adaptive kernel modulation for efficiency and performance.

Datasets Used

• Flickr8k — Image captioning dataset with 8k images and human-annotated captions.
• Breast Cancer Wisconsin Dataset — For Naive Bayes classification analysis.
• CIFAR-10 / CIFAR-100 — For MLP image classification experiments.
• Tiny ImageNet — Image classification benchmark.
• Pascal VOC 2012 — Semantic segmentation benchmark.
• UCR Adiac — Time series classification dataset.

Development & Workflow Tools

• Git — Version control.
• GitHub — Repository hosting and project collaboration.
• VS Code — Main code editor.
• Google Colab — Cloud-based GPU acceleration for model training.