Train and Test Accuracy Match: Success or Underfitting?
Problem statement
How would you improve an image classifier when training and test accuracy both fall well short of human performance? The data split is sound, but the model also misses the product’s accuracy goal. A separate binary classifier has similarly high training and validation losses. Validate the performance comparison and explain how you would diagnose and improve the fit.
Show hint
Separate what the gap tells you from what the absolute level tells you.
Relevant roles: Deep Learning Engineer, ML Engineer, Applied Scientist and Data Scientist.