learning-to-defer

Overview

This repository contains code and resources for the project "Learning to Defer", which focuses on developing a concise package for implementing learning-to-defer.

Installation

1. Clone the repository:

   git clone https://github.com/csangelahu/learning-to-defer.git

2. Navigate to the project directory:

   cd learning-to-defer

3. Install the required dependencies:

   pip install -r requirements.txt

Project Structure

• l2d_functionality/: Implementation of loss functions, inference (l2d_inference.py), and evaluation (l2d_eval_combined.py) classes.
• demos/: Directory for demonstrations showcasing package usage. ALL_refactorized_demo.ipynb demonstrates usage for all L2D settings, using the inference and combined evaluation classes.

Parametrization types

• softmax: [https://arxiv.org/abs/2006.01862]
• asymmetric_sm: [https://openreview.net/pdf?id=TWb9y4PNSW]
• one_vs_all: [https://proceedings.mlr.press/v162/verma22c.html]
• realizable_sm: [https://proceedings.mlr.press/v206/mozannar23a/mozannar23a.pdf]
• softmax and one_vs_all for multiple experts: [https://arxiv.org/pdf/2210.16955]
• Learning from a population: [https://arxiv.org/pdf/2403.02683]