Please enable JavaScript.
Coggle requires JavaScript to display documents.
Structuring Machine Learning Projects (8 basic steps) (Choose your single…
Structuring Machine Learning Projects
(8 basic steps)
Choose your single number evaluation metrics
1a. Should not use both Precision & Recall
1b. But F1 score = "Average of P & R"
1c. Also balance with the running time
Mismatched train and dev set data
Carry out manual error analysis to try to understand the difference between the training set and the dev/test sets
Make training data more similar
Setup your goal
2a. Dev Set + Metrics
2b. Tips: Choose a dev set and test to reflect your data you expect to get in the future and consider important to do well on.
Build your first
QUICK, DIRTY
system, and then iterate
Compare to human level performance
4a. Avoidable bias error = Human level error - Training error
4b. Variance = Training error - Dev error
Improve your model performance
5a. Avoidable bias
Train longer/better optimization algorithms (Momentum/RMSProp/Adam)
NN architect/hyperparameters search (RNN/CNN)
Train bigger model
5b. Variance
More data
Regulization (L2, dropout, data augmentation)
NN architect/hyperparameters search (RNN/CNN)
Error Analysis
Manually find mislabeled data only in
DEV
set
Now we got 100 mislabeled data with (60% of category A (blurry image), 20 % of category B (...), 15% of Category C (...), 10% of Category D (...))
Now we know we should improve from Category A first.
Others
When to use Transfer Learning?
When to use multi-task Learning
When to use end-to-endend deep learning