🧠 CIFAR‑10 Convolutional Neural Network (TensorFlow / Keras)

---

📘 Project Overview

A fully implemented Convolutional Neural Network (CNN) for CIFAR‑10 image classification, written in TensorFlow 2 / Keras.

This project demonstrates:

• Modern CNN architecture design (Conv‑BN‑Pool‑Dropout)
• Data augmentation integrated directly into the model
• He initialization, L2 regularization, and Batch Normalization
• Adam optimizer with exponential‑decay learning rate schedule
• Real‑time training visualization and evaluation via confusion matrix + classification report

---

📂 Dataset: CIFAR‑10

CIFAR‑10 contains 60 000 color images (32 × 32 × 3) in 10 classes:

✈️ airplane • 🚗 automobile • 🐸 frog • 🐱 cat • 🐶 dog • 🐴 horse • 🐦 bird • 🚢 ship • 🦌 deer • 🚚 truck

• Images normalized to [0 – 1]
• Labels converted to one‑hot vectors using to_categorical()
• Usually split: 80 % train / 20 % validation + separate test set

---

🧠 Model Architecture

Figure 1. CIFAR‑10 CNN architecture (Graphviz horizontal view).
Conv blocks use HeUniform initialization + L2 regularization (1e‑5 dense) and progressive dropout (0.25).
Final classifier is Dense(10) activated by softmax and initialized with GlorotUniform.
Optimizer = Adam+ExponentialDecay (LR = 1e‑3, decay rate = 0.9, every 10 000 steps).

✅ Total Params: 2.23 M (8.52 MB)
✅ Trainable Params: 2.23 M
✅ Regularization: L2 → Dense (1e‑5)

---

🧾 Results Summary

Metric	Value
Best Validation Accuracy	≈ 89.7 %
Test Accuracy	≈ 88.8 %
Test Loss	~ 0.43
Total Parameters	2,233,546
Loss Function	categorical_crossentropy
Optimizer	Adam + ExponentialDecay (LR: 1e‑3 → decay 0.9 / 10 000 steps)
Regularization	L2 Dense = 1e‑5; Dropout = 0.25
Augmentation	Flip · Rotation · Zoom · Translation
Framework	TensorFlow 2.16 / Keras
Python Version	3.10

🎨 Visuals & Results Summary

📈 Training & Validation Curves

Figure 2. Loss and accuracy progress over 150 epochs (batch size 128).
The smooth convergence and small gap indicate well‑balanced regularization.
Best validation accuracy: ≈ 89.7 % on held‑out set.

---

🔍 Confusion Matrix & Classification Report

Figure 3. CIFAR‑10 test set confusion matrix with per‑class performance.
Diagonal dominance shows robust feature discrimination.
Remaining misclassifications occur mostly between visually similar classes
(cat ↔ dog, automobile ↔ truck), highlighting realistic domain overlap.

---