ANNDL Challenges 2023-2024

Artificial Neural Networks and Deep Learning Course - Politecnico di Milano

Challenges Overview

Challenge 1: Plant Classification

• Task: Classify plants based on health condition using a provided dataset.
• Dataset: Contained noisy data, including non-useful images. The nature of the noise was not known beforehand.
• Objective: Predict the correct class label (0 or 1).
• Type: Binary classification.
• Grade: 5/5

Challenge 2: Time Series Forecasting

• Task: Predict future samples from input time series data.
• Dataset: Composed of 48000 time series, padded to a length of 2776, belonging to six categories ('A', 'B', 'C', 'D', 'E', 'F').
• Objective: Develop models that can generalize effectively across different time domains.
• Requirement: Predict multiple uncorrelated time series, emphasizing robust generalization. Tested multiple different prediction windows to ensure model versatility.
• Grade: 5/5

For the first challenge, we performed extensive data cleaning, augmentation, and employed both custom CNNs and transfer learning approaches, leading to high accuracy. For the second one, we addressed dataset biases, applied advanced windowing techniques, and explored multiple model architectures including LSTM and ResNet.

Each folder contains the developed models and corresponding reports:

• Challenge 1 Report
• Challenge 2 Report

Authors

Name	Email	GitHub
Luca Lain	luca.lain@mail.polimi.it	@lucalain
Alessandro Mosconi	alessandro2.mosconi@mail.polimi.it	@Alessandro-Mosconi
Martino Piaggi	martino.piaggi@mail.polimi.it	@martinopiaggi