CIFAR-10 Image Classification with CNN and ResNet18

---

Project Overview

This project builds and compares two deep learning architectures for image classification on the CIFAR-10 dataset:

• A baseline Convolutional Neural Network (CNN)
• A ResNet-18 architecture

The goal is to explore how deeper residual architectures improve performance and training stability compared to standard CNNs.

The models were trained and evaluated using PyTorch with GPU acceleration.

---

Dataset

CIFAR-10

• 60,000 images
• 10 classes
• Image size: 32×32
• 50,000 training images
• 10,000 test images

Classes:

airplane
automobile
bird
cat
deer
dog
frog
horse
ship
truck

---

Model Architectures

Baseline CNN

Architecture:

Input Image
   ↓
Conv Layer
   ↓
ReLU
   ↓
Max Pool
   ↓
Conv Layer
   ↓
ReLU
   ↓
Max Pool
   ↓
Fully Connected Layer
   ↓
Softmax

This model serves as a baseline computer vision classifier.

---

ResNet-18

ResNet introduces residual connections, allowing deeper networks to train effectively.

Instead of learning:

H(x)

the network learns a residual:

F(x) = H(x) − x

which improves gradient flow during backpropagation.

Architecture overview:

Input Image
   ↓
Conv Layer
   ↓
Residual Block × 8
   ↓
Global Average Pooling
   ↓
Fully Connected Layer
   ↓
Softmax

---

Code

The full training and evaluation pipeline is provided in the notebook:

vision_cnn_classifier.ipynb

The notebook includes:

• dataset loading
• CNN architecture
• ResNet-18 architecture
• training loops
• evaluation metrics
• confusion matrix generation
• prediction visualization
• feature map visualization

---

Training Setup

Framework: PyTorch

GPU: NVIDIA Tesla T4

Loss Function:

Cross Entropy Loss

Optimizer:

Adam

Training Pipeline:

Dataset
 → DataLoader
 → Model
 → Loss Function
 → Backpropagation
 → Optimizer Step

---

Training Curves

CNN Training Loss

ResNet Training Loss

---

Confusion Matrix

CNN Confusion Matrix

ResNet Confusion Matrix

---

Feature Map Visualization

To better interpret what the network learned, I visualized first-layer feature maps from the trained ResNet-18 model.

These activations highlight low-level visual patterns such as:

• edges
• contrast boundaries
• texture-like structures
• simple shape detectors

---

Example Model Predictions

These predictions highlight both correct classifications and failure cases.

Typical mistakes occur between visually similar classes:

cat ↔ dog
deer ↔ horse
automobile ↔ truck

This is expected due to the small resolution (32×32).

---

Results

Model Comparison

Model	Test Accuracy	Notes
Baseline CNN	70.43%	2-layer custom CNN baseline
ResNet-18	87.50%	Deeper residual architecture with stronger feature extraction

Improvement from CNN to ResNet-18: +17.07 percentage points

ResNet-18 significantly outperformed the baseline CNN due to:

• deeper hierarchical feature learning
• residual connections improving gradient flow
• stronger representation capacity for complex visual patterns

Typical remaining errors occurred between visually similar classes such as:

• cat vs dog
• deer vs horse
• automobile vs truck

Repository Structure

cifar10-resnet18-image-classifier
│
├── cnn_training_curve.png
├── resnet_training_curve.png
├── cnn_confusion_matrix.png
├── resnet_confusion_matrix.png
├── resnet_predictions.png
│
├── cnn_baseline_model.pth
├── resnet18_model.pth
│
├── cnn_training_losses.json
├── resnet_training_losses.json
├── resnet_per_class_accuracy.json

ResNet-18 Per-Class Accuracy

Class	Accuracy
airplane	89.00%
automobile	94.80%
bird	87.00%
cat	81.30%
deer	88.80%
dog	84.30%
frog	77.60%
horse	86.90%
ship	93.00%
truck	92.30%

---

Live Demo

Try the model in your browser:

https://huggingface.co/spaces/dschechter27/cifar10-resnet-demo

Upload an image and the ResNet-18 classifier will predict the most likely CIFAR-10 class.

---

Key Takeaways

• building a full deep learning training pipeline
• training CNN and ResNet architectures
• evaluating models with confusion matrices
• visualizing training curves
• analyzing classification errors

---

Future Improvements/Possible Extensions

• data augmentation
• deeper ResNet variants
• transfer learning from ImageNet
• attention mechanisms
• feature map visualization

---

Author

David Schechter

Incoming MIT '30
Interested in AI, Machine Learning, and Computer Vision