Update usage.rst

Author: MMathisLab

docs/source/usage.rst

@@ -1,7 +1,7 @@
 Using CEBRA
 ===========
 
-This page covers a standard CEBRA usage. We recommend checking out the :py:doc:`demos` for in-depth CEBRA usage examples as well. Here we present a quick overview on how to use CEBRA on various datasets. Note that we provide two ways to interact with the code:
+This page covers a standard CEBRA usage. We recommend checking out the :py:doc:`demos` for CEBRA usage examples as well. Here we present a quick overview on how to use CEBRA on various datasets. Note that we provide two ways to interact with the code:
 
 * For regular usage, we recommend leveraging the **high-level interface**, adhering to ``scikit-learn`` formatting.
 * Upon specific needs, advanced users might consider diving into the **low-level interface** that adheres to ``PyTorch`` formatting.
@@ -22,7 +22,7 @@ We recommend to start with running CEBRA-Time (unsupervised) and look both at th
 
 (1) Use CEBRA-Time for unsupervised data exploration.
 (2) Consider running a hyperparameter sweep on the inputs to the model, such as :py:attr:`cebra.CEBRA.model_architecture`, :py:attr:`cebra.CEBRA.time_offsets`, :py:attr:`cebra.CEBRA.output_dimension`, and set :py:attr:`cebra.CEBRA.batch_size` to be as high as your GPU allows. You want to see clear structure in the 3D plot (the first 3 latents are shown by default).
-(3) Use CEBRA-Behavior with many different labels and combinations, then look at the InfoNCE loss - the lower the loss value, the better the fit (see :py:doc:`cebra-figures/figures/ExtendedDataFigure5`), and visualize the embeddings. The goal is to understand which labels are contributing to the structure you see in CEBRA-Time, and improve this structure. Again, you should consider a hyperparameter sweep.
+(3) Use CEBRA-Behavior with many different labels and combinations, then look at the InfoNCE loss - the lower the loss value, the better the fit (see :py:doc:`cebra-figures/figures/ExtendedDataFigure5`), and visualize the embeddings. The goal is to understand which labels are contributing to the structure you see in CEBRA-Time, and improve this structure. Again, you should consider a hyperparameter sweep (and avoid overfitting by performing the proper train/validation split (see Step 3 in our quick start guide below).
 (4) Interpretability: now you can use these latents in downstream tasks, such as measuring consistency, decoding, and determining the dimensionality of your data with topological data analysis.
 
 All the steps to do this are described below. Enjoy using CEBRA! 🔥🦓