This document outlines my submissions in the Rayan AI-Trustworthy Contest (ai.rayan.global) during the selection phase. I ranked 11th in this phase. The full problem statements and explorations are available in the corresponding directory.
We were provided with a dataset of 1,600 images of similar birds, which were noisy, and only 10% of them were labeled.
- Data Augmentation: Applied proper data augmentation techniques to enhance the dataset.
- Gradual Learning: Utilized a method called gradual learning, where the model is initially trained with limited data. Subsequently, data samples that the model is confident about are added to the training set.
- Model Architecture: Employed EfficientNet as the neural network architecture.
- Accuracy: Achieved 78.5% accuracy.
- Ranking: Placed 5th among all submissions.
The task was to perform segmentation of tooth caries, which was challenged by poisoned data. A significant portion of the data appeared as random noise.
- Noise Reduction: Removed noisy data using an image entropy technique.
- Model Architecture: Used U-Net as the segmentation model.
- Dice Score: Achieved a 0.5110 Dice score.
- Ranking: Placed 10th among all submissions.
This was a hierarchical clustering problem that required developing a classification system for the second layer of the hierarchy. The error could be computed at any arbitrary layer, with the minimum error being selected.
- Data Augmentation: Implemented smart augmentation techniques to handle imbalanced data due to time constraints.
- Model Strategy: Focused on optimizing classification for hierarchical layers.
- Average Accuracy: Achieved a 44.34% average accuracy.
- Ranking: Placed 15th among all submissions.