iakremnev / multitask-classification

Analysis of hard parameter sharing scenario in multitask classification problem.

multitask-classification

We train 2 DNN classifiers on 2 different datasets and a single multitask classifier and examine their performance.

• Base model: Resnet20.
• Dataset 1: Fashion MNIST
• Dataset 2: Imagewoof
• Metrics: top-1 accuracy, Confusion Matrix

Metrics

Top-1 accuracy, %

Model	Fashion MNIST	Imagewoof
Single task	87.96	65.74
Multitask	81.67	62.28

To do

1. Models aren't trained to their best since time shortage.
2. Training is slow, especially for multitask model. Can we take smaller model?
3. Use lr schedulers for multitask training to escape plateau.