Deep Learning Practice – ANN & CNN

This repository contains my hands-on practice and mini-projects on Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Transfer Learning using TensorFlow & Keras. These notebooks/scripts were created as part of my learning.

---

📌 Topics Covered

1️⃣ Binary Classification using ANN (Breast Cancer Dataset)

• Dataset: sklearn.datasets.load_breast_cancer
• Problem Type: Binary Classification (Malignant vs Benign)
• Model Type: Fully Connected ANN
• Loss Function: binary_crossentropy
• Activation Functions: ReLU, Sigmoid
• Optimizer: Adam

Key Concepts Practiced:

• Data loading using sklearn
• Feature–target separation
• Train-test split
• ANN architecture design
• Model compilation & training
• Accuracy & loss visualization
• Threshold-based prediction
• Confusion matrix & accuracy score

---

2️⃣ Multi-Class Classification using ANN (Iris Dataset)

• Dataset: sklearn.datasets.load_iris
• Problem Type: Multi-class classification (3 classes)
• Output Activation: Softmax
• Loss Function: categorical_crossentropy

Additional Concepts:

• One-hot encoding using to_categorical
• Multi-class probability prediction
• argmax for class selection

---

3️⃣ Binary CNN (Day vs Night Image Classification)

• Data Source: Folder-based image dataset
• Model Type: Custom CNN
• Input Size: (256 × 256 × 3)

CNN Architecture:

• Conv2D + MaxPooling layers
• Flatten layer
• Dense ANN layers
• Sigmoid output layer

Image Processing Concepts:

• ImageDataGenerator
• Data augmentation (zoom, flip, shear)
• Rescaling pixel values
• Folder-based dataset handling

---

4️⃣ Multi-Class CNN (Butterfly Species Classification)

• Problem Type: Multi-class image classification
• Number of Classes: 74
• Output Activation: Softmax

Advanced Concepts Used:

• Training, validation & testing splits

• Callbacks:

  • ModelCheckpoint
  • EarlyStopping
  • ReduceLROnPlateau

• Class index mapping

• Prediction on external images

---

5️⃣ Transfer Learning – VGG16

• Pretrained Model: VGG16 (ImageNet weights)
• Task: Binary Image Classification

Transfer Learning Steps:

• Load pretrained model without top layers
• Freeze base layers
• Add custom classifier
• Train on custom dataset

---

📊 Model Evaluation

• Accuracy Score
• Confusion Matrix
• Training vs Validation plots
• Manual image testing using OpenCV

---

🛠 Libraries Used

Python
NumPy
Pandas
Matplotlib
Seaborn
Scikit-learn
TensorFlow / Keras
OpenCV

---

🎯 Learning Outcomes

• Clear understanding of ANN vs CNN
• Binary vs Multi-class classification
• Custom CNN design
• Image preprocessing & augmentation
• Transfer learning fundamentals
• Model evaluation techniques

---

• Use Sigmoid + Binary Crossentropy for binary classification
• Use Softmax + Categorical Crossentropy for multi-class problems
• ANN works well for tabular data
• CNN is best suited for image data
• Transfer learning reduces training time and improves accuracy

---

✍️ Author

Mounika Reddy Aspiring DataScientist

---

⭐ This repository is created for learning, revision, and interview preparation.