MLS-C01 · Question #381
A machine learning (ML) specialist is developing a model for a company. The model will classify and predict sequences of objects that are displayed in a video. The ML specialist decides to use a hybri
The correct answer is D. Reinitialize the weights of the last fully connected layer of the CNN. Retrain the CNN on the E. Reinitialize the weights of the last layer of the RNN. Retrain the entire model on the prediction task. To adapt a pre-trained hybrid CNN-RNN model to classify a new set of objects with minimal effort, the specialist should leverage transfer learning. This involves reinitializing only the output layers of both the CNN and RNN parts.
Question
A machine learning (ML) specialist is developing a model for a company. The model will classify and predict sequences of objects that are displayed in a video. The ML specialist decides to use a hybrid architecture that consists of a convolutional neural network (CNN) followed by a classifier three-layer recurrent neural network (RNN). The company developed a similar model previously but trained the model to classify a different set of objects. The ML specialist wants to save time by using the previously trained model and adapting the model for the current use case and set of objects. Which combination of steps will accomplish this goal with the LEAST amount of effort? (Choose two.)
Options
- AReinitialize the weights of the entire CNN. Retrain the CNN on the classification task by using the
- BReinitialize the weights of the entire network. Retrain the entire network on the prediction task by
- CReinitialize the weights of the entire RNN. Retrain the entire model on the prediction task by using
- DReinitialize the weights of the last fully connected layer of the CNN. Retrain the CNN on the
- EReinitialize the weights of the last layer of the RNN. Retrain the entire model on the prediction task
How the community answered
(31 responses)- A3% (1)
- B3% (1)
- C10% (3)
- D84% (26)
Why each option
To adapt a pre-trained hybrid CNN-RNN model to classify a new set of objects with minimal effort, the specialist should leverage transfer learning. This involves reinitializing only the output layers of both the CNN and RNN parts.
Reinitializing the entire CNN would discard potentially useful pre-trained feature extraction knowledge, increasing training effort unnecessarily.
Reinitializing the entire network defeats the purpose of transfer learning, making it equivalent to training from scratch and requiring significantly more effort.
Reinitializing the entire RNN would discard valuable pre-trained knowledge about temporal dependencies, which could still be relevant for the new prediction task, thus increasing effort.
Reinitializing only the last fully connected layer of the CNN allows retaining the learned lower-level feature representations from the previous task, while adapting the output for the new set of objects.
Reinitializing the last layer of the RNN allows the model to learn the new mapping for sequence prediction, while preserving the learned temporal dependencies in the earlier RNN layers. The entire model is then retrained to fine-tune the updated layers.
Concept tested: Transfer learning with hybrid CNN-RNN models
Source: https://www.tensorflow.org/tutorials/images/transfer_learning
Topics
Community Discussion
No community discussion yet for this question.