Add links to the paper (#63)

* Add link to the paper in UPGrad's docstring
* Add link to the paper in README.md
* Add citation section at the end of README.md
* Add hint admonition with link to paper in iwrm.rst
* Add link to the paper in the main documentation page
* Update link to paper in CHANGELOG.md
* Add hint with link to the paper in aggregation index

CHANGELOG.md

@@ -38,7 +38,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   - `Sum` to sum the rows of the matrix.
   - `TrimmedMean` from [Byzantine-Robust Distributed Learning: Towards
       Optimal Statistical Rates](https://proceedings.mlr.press/v80/yin18a/yin18a.pdf).
-  - `UPGrad` from [Jacobian Descent for Multi-Objective Optimization](https://arxiv.org/search/?query=jacobian+descent+for+multi-objective+optimization&searchtype=all&source=header).
+  - `UPGrad` from [Jacobian Descent for Multi-Objective Optimization](https://arxiv.org/pdf/2406.16232).
 - `backward` function to perform a step of Jacobian descent.
 - Documentation of the public API and of some usage examples.
 - Tests:

README.md

@@ -1,6 +1,6 @@
 # ![image](docs/source/icons/favicon-32x32.png) TorchJD
 
-TorchJD is a library enabling Jacobian descent with PyTorch, for optimization of neural networks
+TorchJD is a library enabling [Jacobian descent](https://arxiv.org/pdf/2406.16232) with PyTorch, for optimization of neural networks
 with multiple objectives.
 
 > [!IMPORTANT]
@@ -25,3 +25,14 @@ TorchJD requires python 3.10, 3.11 or 3.12. It is only compatible with recent ve
 ## Contribution
 
 Please read the [Contribution page](CONTRIBUTING.md).
+
+## Citation
+If you use TorchJD for your research, please cite:
+```
+@article{jacobian_descent,
+  title={Jacobian Descent For Multi-Objective Optimization},
+  author={Quinton, Pierre and Rey, Valérian},
+  journal={arXiv preprint arXiv:2406.16232},
+  year={2024}
+}
+```

docs/source/docs/aggregation/index.rst

@@ -27,3 +27,7 @@ This package provides several aggregators from the literature:
     sum.rst
     trimmed_mean.rst
     upgrad.rst
+
+.. hint::
+    Most of these aggregators are analyzed theoretically in `Jacobian Descent For Multi-Objective
+    Optimization <https://arxiv.org/pdf/2406.16232>`_.

docs/source/examples/iwrm.rst

@@ -5,6 +5,11 @@ This example shows how to use TorchJD to minimize the vector of per-instance los
 paradigm, called IWRM, is multi-objective, as opposed to the usual empirical risk minimization
 (ERM), which seeks to minimize the average loss.
 
+.. hint::
+    A proper definition of IWRM and its empirical results on some deep learning tasks are
+    available in `Jacobian Descent For Multi-Objective Optimization
+    <https://arxiv.org/pdf/2406.16232>`_.
+
 For the sake of the example, we generate a fake dataset consisting of 8 batches of 16 random input
 vectors of dimension 10, and their corresponding scalar labels. We train a very simple regression
 model to retrieve the label from the corresponding input. To minimize the average loss, we use

docs/source/index.rst

@@ -17,7 +17,9 @@
 |
 
 TorchJD is a library enabling Jacobian descent with PyTorch, for optimization of neural networks
-with multiple objectives.
+with multiple objectives. It is based on the theory from `Jacobian Descent For Multi-Objective
+Optimization <https://arxiv.org/pdf/2406.16232>`_ and it contains algorithms from many other related
+papers.
 
 .. important::
     This library is currently in an early development stage. The API is subject to significant changes

src/torchjd/aggregation/upgrad.py

@@ -14,7 +14,8 @@
 class UPGrad(_WeightedAggregator):
     """
     :class:`~torchjd.aggregation.bases.Aggregator` that projects each row of the input matrix onto
-    the dual cone of all rows of this matrix, and that combines the result.
+    the dual cone of all rows of this matrix, and that combines the result, as proposed in
+    `Jacobian Descent For Multi-Objective Optimization <https://arxiv.org/pdf/2406.16232>`_.
 
     :param pref_vector: The preference vector used to combine the projected rows. If not provided,
         defaults to the simple averaging of the projected rows.