Artificial Neural Networks and Deep Learning Challenges - PoliMi

This repository contains the solutions and code for challenges completed as part of the Artificial Neural Networks and Deep Learning course at the Polytechnic University of Milan.

📚 Overview

The goal of these challenges, completed between September and December 2023, was to design and implement deep learning models to solve a variety of tasks, including binary classification of plant health based on image datasets and forecasting future samples in uncorrelated time series. The challenges emphasized building effective models, optimizing their performance, and analyzing errors to gain insights.

Challenge 1: Binary Classification of Plant Health

The solutions include:

• Preprocessing datasets: Removing outliers, duplicates, and handling imbalances.
• Model implementation: ResNet, ConvNeXtBase, EfficientNet, and other state-of-the-art architectures.
• Optimization techniques: Transfer learning, fine-tuning, and hyperparameter tuning.
• Performance evaluation: Ensemble methods and error analysis.

Challenge 2: Time Series Forecasting

In this challenge, the task was to predict future samples of uncorrelated time series based on past observations. The solutions include:

• Data preprocessing: Removing padding, handling variable time series lengths, and categorizing datasets.
• Model implementation: Bidirectional LSTM networks with masking layers for sequence-to-vector predictions.
• Feature engineering: Building and expanding sequences using custom algorithms.
• Performance evaluation: Validation through Mean Squared Error (MSE) and ensemble methods for improved generalization.

📂 Repository Structure

|-- challenge1/
    |-- notebooks/
        |-- ResNetFromZero.ipynb          # Implementation of ResNet from scratch
        |-- ConvNextBase.ipynb            # ConvNeXtBase model training and evaluation
        |-- EfficientNetV2S.ipynb         # EfficientNetV2S model training and evaluation
        |-- DeleteShrekAndTrololoDuplicatesFromDataset.ipynb  # Dataset cleaning script
    |-- scripts/
        |-- model.py                      # Python script for ensemble model loading and prediction
    |-- Report.pdf                        # Detailed report on the challenge and solutions
|-- challenge2/
    |-- notebooks/
        |-- one_dimensional_convolutiol_network.ipynb  # Implementation of 1D convolutional models
        |-- recurrent_network.ipynb                  # Implementation of recurrent models
        |-- official_data_preprocessing.ipynb        # Data preprocessing and sequence generation
    |-- Report.pdf                                    # Detailed report on the time series challenge
|-- README.md                                         # Repository documentation

🚀 Key Features

Challenge 1

• Transfer Learning: Leveraging pre-trained models (ImageNet) to improve performance on limited datasets.
• Data Augmentation: Applying techniques like rotation, flipping, and translation to enhance training robustness.
• Ensemble Models: Combining predictions from multiple models to improve accuracy and reduce variance.
• Error Analysis: Identifying and interpreting misclassified samples for insights.

Challenge 2

• Bidirectional LSTMs: Using sequence-to-vector networks for robust forecasting.
• Feature Engineering: Expanding time series data through sequence generation.
• Trend and Seasonality Analysis: Handling non-stationary time series by removing trends and periodic components.
• Ensemble Models: Combining models trained on different splits for better generalization.

🧑‍💻 Getting Started

1. Clone the repository:

   git clone https://github.com/yourusername/ANN-DL-Challenges.git

2. Explore the notebooks for detailed implementations and results.

📝 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Happy coding! 🎉