Update papers/pasteur_labs/main.md

component -> pipeline

Author: xalelax

papers/pasteur_labs/main.md

@@ -20,7 +20,7 @@ Building pipelines that propagate gradients effortlessly across components intro
 To the best of our knowledge, modern scientific computing tools do not provide native support for service-level AD, limiting all autodiff processing to a single component. However it is certainly possible to combine multiple tools to create AD-capable components that can communicate over network. Typically such solution would involve an AD-native framework (e.g., JAX, TensorFlow, PyTorch, Julia) and a service layer that exposes gradient-specific endpoints. Service layer can be implemented using general-purpose web application frameworks (e.g. Flask or Django) or specialised serving solutions (e.g. Ray Serve, AWS SageMaker, Goolge Vertex AI). Crucially, there is no standardized gradient APIs, no REST/gRPC protocols to follow, so all of these solutions for service-level AD require large effort to define AD-relevant endpoints, input and output schemas. Additionally, it is necessary to factor in orchestration of distributed gradient calculation jobs and workflows.
 
 We are not the first to recognise this gap, and there are several examples in academic literature of proposed solutions for distributed AD. @baker2021peering proposed a technique for dstributed training of deep neural networks that leverages
-the outer-product structure of the gradient of a network layer. @rush2024federated adopted AD to the context of federated learning of ML models. @tang2023auto considered the problem of differentiating computations expressed in relational databases. Notwithstanding these selected examples, even experimental support for component-level AD is nascent, and it is remarkable that recent surveys on the state of AD do not consider it being an important research direction (@van2018automatic, @baydin2018automatic).
+the outer-product structure of the gradient of a network layer. @rush2024federated adopted AD to the context of federated learning of ML models. @tang2023auto considered the problem of differentiating computations expressed in relational databases. Notwithstanding these selected examples, even experimental support for pipeline-level AD is nascent, and it is remarkable that recent surveys on the state of AD do not consider it being an important research direction (@van2018automatic, @baydin2018automatic).
 
 ## Tesseracts enable differentiable physics programming at scale