202502012212
tags: #machine-learning #deep-learning #architecture

End-to-end deep learning trains a single neural network to map directly from input to output, replacing traditional multi-step pipelines.

Traditional pipeline example (speech recognition):
Audio → Features → Phonemes → Words → Transcript

End-to-end approach:
Audio → Transcript (single neural network)

Advantages:

• Automatically learns intermediate representations
• Eliminates hand-designed feature engineering
• Can discover patterns humans might miss
• Simpler system architecture

Disadvantages:

• Requires massive amounts of data
• Less interpretable than pipeline components
• Harder to debug individual components
• May need more computational resources

When to use end-to-end:

• Large datasets available (typically millions of examples)
• Task is complex with many pipeline stages
• Traditional approaches are hard to optimize

When to use pipelines:

• Limited data available
• Need interpretability for individual components
• Existing pipeline components work well
• Want to optimize individual stages separately

The choice impacts your Error Analysis approach and debugging strategies.

Reference

Machine Learning Yearning by Andrew Ng