https://towardsdatascience.com/e2e-the-every-purpose-ml-method-5d4f20dafee4

University of Michigan - Ann Arbor

End-to-end training is a core methodology in deep learning where a complete system is built and tuned jointly to map raw inputs to final outputs. This approach contrasts with traditional machine learning pipelines that assemble individually optimized components, such as a separate, manually engineered feature extractor followed by a shallow classifier. By simultaneously training all layers of computation, the model automatically discovers the optimal representations needed for the task.

End-to-End Training

End-to-end learning, the (almost) every purpose ML method

- **Lets the data speak:** For a large enough data set, the neural network can figure out the appropriate function mapping from X (input) to Y (output). Moreover, by relying on a pure machine learning approach, the data and its underlying statistics are the main factors for classifying/predicting the output, and it does not have to reflect on human preconceptions.
- **Less hand-designing of components:** This simplifies the design workflow and saves a lot of time designing individual features of the pipeline. This may contribute to a more ubiquitous adoption of this learning method.

Advantages of End-to-End Deep Learning

- **Large amount of data is necessary:** incorporating some prior knowledge into the training improves performance and is considered a key element. For end-to-end learning more training examples must be provided because we don't incorporate this prior knowledge.
- **Difficult to improve or modify the system:** if some structural change must be applied, the old model is of no use and the whole model has to be replaced and trained all over again.
- **Potentially useful hand-designed modules cannot be used:** many hand-designed modules hold a lot of human knowledge and are efficient in solving some tasks, especially when the available data is limited.
- **Difficult to validate:** due to the complex architecture, the potential number of input/output pairs can be big enough to make high level validation impossible.

Limitations of End-to-End Deep Learning

- Speech Recognition
- Autonomous Driving

Learn Before

Related