Exam AWS Certified Machine Learning - Specialty topic 1 question 48 discussion

Ans B sounds correct

In transfer learning, a pre-trained model is used as a starting point to train a new model on a different task, typically using a smaller dataset. The pre-trained model contains weights that have been learned from a large amount of data on a related task, and these weights can be leveraged to train the new model more efficiently. To re-train the model with the custom data, the Specialist should initialize the model with pre-trained weights in all layers, as these weights can provide a good starting point for the new task. The Specialist should then replace the last fully connected layer, which is responsible for making the final predictions, as this layer will likely need to be modified to reflect the new task. By keeping the pre-trained weights in the other layers, the Specialist can take advantage of the knowledge learned from the previous task, and potentially speed up the training process.

Explanation: The Machine Learning Specialist wants to use transfer learning with an existing model trained on general object images and fine-tune it for vehicle make and model classification. The best approach is: Use pre-trained weights from the existing model for feature extraction. Replace the last fully connected (FC) layer to match the number of vehicle classes. Fine-tune the new model on the vehicle dataset. Why This Works? Lower training time: The model has already learned useful features from general objects (e.g., edges, shapes). Improves accuracy: Instead of training from scratch, transfer learning leverages knowledge from large datasets (e.g., ImageNet). Avoids catastrophic forgetting: Reusing pre-trained weights preserves learned low- and mid-level features while adapting the last layer for new classes.

Transfer learning helps accelerate the training and at this point, model has yet to learn from the new data. So, all layers including the fully-connected by replaced. Eventually, the training will update the fully-connected layer. The question is about initialization, so we should initialize the fully-connected layers too.

A. NO - random weights does not allow transfer learning B. YES - the last layer gives the final classes, we want to have new classes C. NO - random weights does not allow transfer learning D. NO - the last layer gives the final classes, we want to have new classes

Option B

For Transfer Learning, A and C are incorrect because we restart the model. The correct is letter B

B. The reason is, fine-tuning a model means to use the weights/biases trained before. also no matter which strategy you go for in transfer learning (fine-tuning or feature extraction) you always replace the last or last few layers.

The task is to " to re-train it with the custom data". That means, it is not transfer learning anymore. The "transfer learning" is just a title to make a question tricky. So, in this case we should randomize the weights and retrain whole model from scratch on custom user's images only. The correct answer is C.

The task is to " to re-train it with the custom data". That means, it is not transfer learning anymore. The "transfer learning" is just a title to make a question tricky. So, in this case we should randomize the weights and retrain whole model from scratch on custom user's images only. The correct answer is C.

The fully connected layer will need to be trained from scratch to incorporate the features of his domain problem (Car models)

12-sep exam

D is the best - here is why Question is not to design a final production with deep lense - it is to use it as a dev platform to comeup with a edge ML vs. dump load all to S3 -which is very wasteful! AWS did not mae deeplense as a toy for devs! it is meant to help companies experiment with edge ML And then copy and reuse the open hardware platfom

one of the method to implement transfer learning

I will go with B, we are mainly concerned with the output layer for us to get the desired results, hence we need to replace it.

B is correct

Actually, it should be NONE of IT!.... it should be like B with exception that 20-40% top layers should be retrained :) -- this is classic transfer learning setup, so B is the answer here.