TaskMet: Task-Driven Metric Learning for Model Learning
Fuente:
arXiv
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2023
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912046427144192 |
|---|---|
| author | Bansal, Dishank Chen, Ricky T. Q. Mukadam, Mustafa Amos, Brandon |
| author_facet | Bansal, Dishank Chen, Ricky T. Q. Mukadam, Mustafa Amos, Brandon |
| contents | Deep learning models are often deployed in downstream tasks that the training procedure may not be aware of. For example, models solely trained to achieve accurate predictions may struggle to perform well on downstream tasks because seemingly small prediction errors may incur drastic task errors. The standard end-to-end learning approach is to make the task loss differentiable or to introduce a differentiable surrogate that the model can be trained on. In these settings, the task loss needs to be carefully balanced with the prediction loss because they may have conflicting objectives. We propose take the task loss signal one level deeper than the parameters of the model and use it to learn the parameters of the loss function the model is trained on, which can be done by learning a metric in the prediction space. This approach does not alter the optimal prediction model itself, but rather changes the model learning to emphasize the information important for the downstream task. This enables us to achieve the best of both worlds: a prediction model trained in the original prediction space while also being valuable for the desired downstream task. We validate our approach through experiments conducted in two main settings: 1) decision-focused model learning scenarios involving portfolio optimization and budget allocation, and 2) reinforcement learning in noisy environments with distracting states. The source code to reproduce our experiments is available at https://github.com/facebookresearch/taskmet |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2312_05250 |
| institution | arXiv |
| publishDate | 2023 |
| record_format | arxiv |
| spellingShingle | TaskMet: Task-Driven Metric Learning for Model Learning Bansal, Dishank Chen, Ricky T. Q. Mukadam, Mustafa Amos, Brandon Machine Learning Artificial Intelligence Optimization and Control Deep learning models are often deployed in downstream tasks that the training procedure may not be aware of. For example, models solely trained to achieve accurate predictions may struggle to perform well on downstream tasks because seemingly small prediction errors may incur drastic task errors. The standard end-to-end learning approach is to make the task loss differentiable or to introduce a differentiable surrogate that the model can be trained on. In these settings, the task loss needs to be carefully balanced with the prediction loss because they may have conflicting objectives. We propose take the task loss signal one level deeper than the parameters of the model and use it to learn the parameters of the loss function the model is trained on, which can be done by learning a metric in the prediction space. This approach does not alter the optimal prediction model itself, but rather changes the model learning to emphasize the information important for the downstream task. This enables us to achieve the best of both worlds: a prediction model trained in the original prediction space while also being valuable for the desired downstream task. We validate our approach through experiments conducted in two main settings: 1) decision-focused model learning scenarios involving portfolio optimization and budget allocation, and 2) reinforcement learning in noisy environments with distracting states. The source code to reproduce our experiments is available at https://github.com/facebookresearch/taskmet |
| title | TaskMet: Task-Driven Metric Learning for Model Learning |
| topic | Machine Learning Artificial Intelligence Optimization and Control |
| url | https://arxiv.org/abs/2312.05250 |