This project addresses two learning problems defined in the course assignment
Train convolutional neural networks to classify blood cell images from the BloodMNIST dataset.
- 8 classes, images resized to 3×28×28.
- Models are trained with Adam and CrossEntropy loss for 200 epochs.
- Compare behaviour with and without Softmax on the output layer.
- Add max-pooling layers and measure the impact on accuracy and runtime.
- Report training loss and validation/test accuracy over epochs.
The goal is to study architectural choices in small CNNs and their effect on performance and efficiency.
Predict continuous binding affinity between RNA sequences and the RNA-binding protein RBFOX1.
- Input: one-hot encoded RNA sequences from the RNAcompete dataset.
- Output: normalized fluorescence intensity.
- Task is regression with masked MSE loss.
- Evaluation metric is Spearman rank correlation.
Work includes:
- Implementing and comparing two different neural architectures.
- Extending one architecture with attention and analysing its effect.
- Designing a conceptual extension to handle multiple RNA-binding proteins.
The task focuses on learning data-driven approximations of RNA-protein binding preferences rather than explicit biophysical or combinatorial modelling.