This project trains and analyses a convolutional neural network (CNN) on the CIFAR-10 image classification dataset. The focus is not only on model accuracy, but also on interpretability through rich visualisations and error analysis.
The core model is a compact CNN built with TensorFlow / Keras using:
- stacked convolutional blocks (Conv2D → BatchNorm → ReLU → MaxPooling → Dropout)
- a dense classification head with softmax over 10 classes
The repository is organised as follows:
Computer-Vision-CIFAR10/
├── docs/ # Interactive Plotly HTML exports (for GitHub Pages)
│ ├── cifar10_acc.html
│ ├── cifar10_avg_confidence_per_true_class.html
│ ├── cifar10_confidence_hist.html
│ ├── cifar10_confusion_matrix.html
│ ├── cifar10_hard_prediction_grid.html
│ ├── cifar10_loss.html
│ ├── cifar10_misclassification_grid.html
│ ├── cifar10_per_class_accuracy.html
│ ├── cifar10_top_1_percent_correct_predictions.html
│ ├── cifar10_top_1_percent_wrong_predictions.html
│ ├── class_distribution.html
│ └── examples_per_class.html
│
├── models/ # Saved Keras models
│ └── cifar10_main.keras
│
├── notebooks/ # Jupyter notebooks (EDA and training)
│ ├── CNN-CIFAR10-NOTEBOOK.ipynb
│ └── EDA.ipynb
│
├── plots/ # Static PNG previews for README and index.html
│ ├── cifar10_acc.png
│ ├── cifar10_avg_confidence_per_true_class.png
│ ├── cifar10_confidence_hist.png
│ ├── cifar10_confusion_matrix.png
│ ├── cifar10_hard_prediction_grid.png
│ ├── cifar10_loss.png
│ ├── cifar10_misclassification_grid.png
│ ├── cifar10_per_class_accuracy.png
│ ├── cifar10_top_1_percent_correct_predictions.png
│ ├── cifar10_top_1_percent_wrong_predictions.png
│ ├── class_distribution.png
│ └── examples_per_class.png
│
├── results/ # Serialized training history and metrics
│ └── history_cifar10_main.json
│
├── src/ # Reusable Python modules
│ └── utils.py
│
├── tests/ # Unit tests
│ └── test_utils.py
│
├── index.html # Landing page for GitHub Pages (interactive gallery)
└── README.md # Project documentation (this file)
All interactive Plotly figures are available on GitHub Pages:
👉 https://42kiko.github.io/Computer-Vision-CIFAR10/
The gallery includes:
- dataset overview (class distribution, examples per class)
- training dynamics (accuracy and loss)
- confusion matrix and per-class accuracy
- confidence analysis and detailed grids of correct / wrong predictions
This plot shows how many training examples exist for each CIFAR-10 class.
The dataset is roughly balanced, which is helpful for training the classifier.
Click the image to open the interactive Plotly version in your browser:
This grid visualises several example images for each class.
It helps to build an intuitive understanding of what the model actually sees
and how visually similar some categories are.
Click the image to open the interactive Plotly version in your browser:
In the notebooks and interactive plots you can explore:
-
Training curves
- Accuracy and loss over epochs for train and validation sets
- Used to check convergence and detect overfitting / underfitting
-
Confusion matrix & per-class accuracy
- Shows which classes are recognised reliably
- Highlights pairs of classes that are frequently confused
-
Confidence & error analysis
- Confidence distribution for correct vs wrong predictions
- Top 1 % most confident correct and wrong predictions
- “Hard but correct” edge cases and misclassification grids
For full details, see the notebooks in notebooks/ and the interactive
visualisations on the GitHub Pages dashboard linked above.
Below are some of the most interesting plots with direct links to their interactive counterparts.
How fast and how far does the model learn?
This plot shows training and validation accuracy over time.
Complement to accuracy: the loss curve helps to see whether the model is still improving or has started to overfit.
The confusion matrix summarises which classes the model gets right and where it tends to make mistakes. Bright off-diagonal cells indicate frequent confusion.
Some classes are harder than others.
This bar chart shows the accuracy for each CIFAR-10 class individually.
How calibrated is the model?
This histogram compares predicted probabilities for the predicted class on:
- correctly classified samples
- misclassified samples
Here you can see, for each true class, how confident the model is on average about its predictions. This helps to spot overconfident or underconfident classes.
These are the “easiest” samples for the model – images where it is extremely confident and correct. Useful to see what the model has really internalised.
Even more interesting: images where the model is very confident, but still wrong. These often reveal dataset issues, label noise or systematic weaknesses in the model.
This grid focuses specifically on wrong predictions.
Each tile shows a misclassified image together with:
- the true class (what it really is)
- the predicted class (what the model thought it was)
- the model confidence for its (wrong) prediction
By scanning this grid you can quickly spot:
- which visual patterns tend to confuse the model
- recurring error types (e.g. 🐱 cats vs. 🐶 dogs,
✈️ airplanes vs. 🛳️ ships) - whether the model is confidently wrong or mostly unsure on difficult samples
Click the image to open the interactive Plotly version in your browser:
