Machine Learning Specialist is training a model to identify the make and model of vehicles in images. The Specialist wants to use transfer learning and an existing model trained on images of general…

Question

Options

• AInitialize the model with random weights in all layers including the last fully connected layer.
• BInitialize the model with pre-trained weights in all layers and replace the last fully connected layer.
• CInitialize the model with random weights in all layers and replace the last fully connected layer.
• DInitialize the model with pre-trained weights in all layers including the last fully connected layer.

• A
  7% (2)
• B
  89% (24)
• C
  4% (1)

Explanation

Standard transfer learning practice is to retain the pre-trained weights for all convolutional/feature-extraction layers (which have learned rich, general visual features from ImageNet) and replace only the final fully connected classification layer with a new one sized to the number of vehicle make/model classes. The new classification head is then trained (and optionally the earlier layers fine-tuned) on the custom dataset. Initializing with random weights everywhere (A, C) discards the pre-trained knowledge and defeats the purpose of transfer learning. Keeping the original last layer with pre-trained weights (D) is incorrect because the output size corresponds to ImageNet classes (e.g., 1,000), not vehicle classes.

No community discussion yet for this question.