Regarding Transfer Learning

[amankharb]

i want to ask that we can apply transfer learning only on that dataset which is similar to dataset of previously pre-trained model.If there is very less similarity between the datasets then transfer learning fails??

For eg using Resnet50 trained on imagenet and using it as a pretrained model for self driving car dataset for predicting steering angles

But it can also be seen that using transfer learning in this case the convolution layers will extract good features from the image like we can extract the other cars,edges etc. and this can be applied to any image and will give good results.

and one more qs imagenet challenge was a classification problem and can we apply resnet50 on a regression problem too if dataset is similar??

[prashant_ml (Prashant Arora)]

hey @amankharb ,

@amankharb

viously pre-trained model.If there is very less similarity between the datasets then transfer learning fails??

Its not that it fails completely , but yeah the performance that can be achieved ,say 100 , then it will be only be able to get you too around 40 . Yeah which is a alot degraded.

@amankharb

But it can also be seen that using transfer learning in this case the convolution layers will extract good features from the image like we can extract the other cars,edges etc. and this can be applied to any image and will give good results.

yeah it should give good results. But the imagenet dataset is very large and diverse , having flowers ,roads , cycles , etc… And using a model based on such dataset to predict over the cars , i hope it wont work.

@amankharb

one more qs imagenet challenge was a classification problem and can we apply resnet50 on a regression problem too if dataset is similar??

Yes You can use it. Basically the task which matters is how you process the information you gain from the pre trained models.
The more better you process ,more better you performance will be. You can use it for any task either classification or regression.

I hope this helped you.