Experiment Documentation
-
Developed CNN from scratch , created 2 CNN layer and ran on 100 epoch but data was overfiting , introduced dropout, it didn't had drastic change and due to shortage of data it was not giving goog result.
-
And problem statement with less data are well solved using transfer learning, the current result is based on transfer learning, the results can be improved further using hyper tuning such as changing optimizer to adam.
-
The another rescue for less data and ambiguous is one shot learning refere to below paper https://arxiv.org/pdf/1605.06065.pdf