Neural networks, deep learning, NLP · insight-1410
Look at a labelled grid of augmented samples before you train on them
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- ML engineer, Data scientist
- Industries
- manufacturing
Summary: situation, action, reason, boundary
Situation
You add random crops, flips, jitter or noise to a training pipeline and plan to judge the result only by validation accuracy.
Action
Dump 16 to 32 augmented samples with their labels, exactly as the tensors enter the network, and have someone who knows the domain reject any that no longer carry their label.
Reason
The tensors fed to the model are the most reliable view of what it sees, so many preprocessing and augmentation bugs are visible there although they never raise an error.
Boundary
Every new augmentation or preprocessing step, and again after dependency upgrades.
Symptom or risk
An augmentation that changes the label, crops away the evidence or leaks a fixed artefact passes silently, and the model learns a shortcut or loses the class you needed.
Exceptions
Non-visual data still benefits; print a few rows of the transformed features instead.
Trade-offs
Ten minutes of human attention per pipeline change.
Pitfalls
Viewing the pre-transform data or the validation path only, which hides the training-path bug.
Lessons and concepts
Claims and their sources
Karpathy's recipe advises inspecting the exact data right before it enters the network, since that shows what the model actually sees.