Computational Perception — Jupyter Notebooks

Practical lab notebooks for the Computational Perception course, part of the Computer Science degree at the Universitat Politècnica de València (UPV).

---

Overview

This repository collects hands-on Jupyter notebooks covering the main topics of the Computational Perception course: classical machine learning, dimensionality reduction, deep learning with Keras, hyperparameter optimisation, and fine-grained image classification. Each notebook (TL prefix stands for Taller, i.e. workshop) builds on the previous one, progressively introducing more complex techniques.

---

Notebooks

#	Notebook	Topics
02	TL02 — PCA	Principal Component Analysis, dimensionality reduction, scikit-learn
03	TL03 — Generative Linear Models	Naive Bayes, joint distribution modelling, digits dataset
04	TL04 — Discriminative Linear Models	Logistic Regression, feature engineering, non-linear separability, MNIST
06	TL06 — Keras	Dense neural networks, ReLU / Softmax, early stopping, dropout
07	TL07 — Keras Tuner	Hyperparameter tuning, Random Search, optimal architecture for MNIST
08	TL08 — KerasHub	Pretrained models (ResNet, EfficientNet, MobileNet), transfer learning
09	TL09 — FGIC	Fine-Grained Image Classification, Oxford 102 Flowers, fine-tuning, data augmentation

---

Notebook Descriptions

TL02 — Principal Component Analysis

Applies PCA to reduce the dimensionality of datasets in machine learning tasks. Uses scikit-learn to explore how projecting data onto principal components affects model performance and visualisation.

TL03 — Generative Linear Models

Introduces generative models that learn the joint distribution P(x, y) of features and class labels. Implements the Naive Bayes classifier using the Bayes theorem and evaluates it on the scikit-learn digits dataset.

TL04 — Discriminative Linear Models

Covers discriminative approaches that model P(y|x) directly to learn decision boundaries. Topics include:

• Logistic Regression with sigmoid activation
• Feature engineering and polynomial transformation to tackle non-linearly separable problems
• Application to the MNIST handwritten digit dataset with reduced sample sizes

TL06 — Introduction to Keras

First contact with the Keras deep learning library. Builds fully-connected (dense) neural networks and covers:

• Network architecture: input layer, hidden layers (ReLU), output layer (Softmax)
• Training parameters: epochs, batch size, validation_split
• Regularisation techniques: Early Stopping and Dropout

TL07 — Keras Tuner

Automates the search for optimal neural network hyperparameters using Keras Tuner with Random Search. Finds the best dense architecture for classifying MNIST digits by exploring the number of layers, units per layer, and learning rate.

TL08 — KerasHub & Transfer Learning

Explores the KerasHub official library to load and use pretrained models (ResNet, EfficientNet, MobileNet). Demonstrates how to:

• Load a backbone (resnet_vd_50_ssld_v2_imagenet) pretrained on ImageNet
• Use ImageClassifier for end-to-end image classification with pretrained weights

TL09 — Fine-Grained Image Classification (FGIC)

Tackles the challenge of classifying visually similar categories using the Oxford 102 Flower dataset (102 classes). Covers:

• Dataset loading with tensorflow_datasets
• Transfer learning with ResNet as feature extractor
• Fine-tuning: unfreezing upper layers and using a lower learning rate
• Data augmentation to improve generalisation
• Training metrics visualisation and final evaluation on the test set

---

Tech Stack

Library	Purpose
scikit-learn	Classical ML models, PCA, datasets
numpy / matplotlib	Numerical computing and visualisation
keras / tensorflow	Deep learning framework
keras_tuner	Hyperparameter optimisation
keras_hub	Pretrained model hub
tensorflow_datasets	Standard dataset loading (Oxford Flowers 102)

---

Requirements

pip install -r requirements.txt

Python 3.9+ is recommended. GPU support (CUDA) is strongly advised for the Keras notebooks (TL06–TL09).

---

Course Information

• Course: Percepción Computacional (Computational Perception)
• Degree: Ingeniería Informática (Computer Science)
• University: Universitat Politècnica de València (UPV)
• Academic Year: 2024–2025

---

License

This repository is for academic and educational purposes. No license is explicitly stated; please contact the author before reusing any material.