Hi folks, I'm a computer scientist PhD student, and I'm working on implementing Multi-Task Learning architecture for a better generalization aims, it will be throughout a Deep Learning model. I have some questions concerning MTL algorithms and its feasibility, for those whom already worked on the same project, here are my questions:
1- Can we design an MTL architecture model based on different task's definition ? Example: task 01: is a classification, task 02: is a clustering (mixing between supervised and unsupervised tasks) is it possible, or we have to design a common and homogeneous architecture ?
2- Is it a mandatory to assign a specific dataset for each task ? Or, we can use a common and global dataset for both shared layers and tasks specific layers (example: an ecommerce historical purchase) ?
3- According to you, what are the best pretraining MTL architecture models that I could rely on ? Thanks in advance !
1. Combining Different Task Types
Yes, you can design an MTL architecture that combines different tasks, like classification (supervised) and clustering (unsupervised). Use shared layers to learn common features, followed by task-specific layers for each task type.
Example Architecture:
- Input Layer
- Shared Layers (e.g., convolutional or dense)
- Task-Specific Branches: (Classification: Dense layers → Softmax, Clustering: Dense layers → Embedding output)
2. Dataset Allocation
You can use a common dataset for both shared layers and task-specific layers, especially if the tasks are related.
Approach:
- Shared Dataset: Useful for tasks that can benefit from shared representations.
- Task-Specific Augmentation: Apply specific preprocessing if needed.
3. Recommended Pretrained MTL Models:
- BERT: Great for NLP tasks, with shared transformer layers and task-specific heads.
- MT-DNN: An extension of BERT, designed for multiple NLP tasks.
- Multi-Task CNNs: Shared convolutional layers with task-specific fully connected layers for image tasks.
- U-Net: Shared encoder with multiple decoder heads for tasks like segmentation and classification.
Pretraining Approach:
- Transfer Learning: Fine-tune a pretrained model on your specific tasks.
- Joint Training: Train on multiple tasks from the start to learn generalized features.
- These approaches will help you effectively implement a robust MTL architecture.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Control Theory
- Control Systems