PMLE Scaling Prototypes into ML Models Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Freeze the first few layers and train the rest.

  Why this is correct

  Freezing early convolutional layers preserves the generic features ResNet learned on ImageNet, so only the later, task-specific layers train. Restricting updates to fewer parameters on a small dataset reduces overfitting while still adapting the model.

• ✗

  Add more convolutional layers to the model.

  Why it's wrong here

  Adding convolutional layers increases trainable parameters, giving the model greater capacity to memorise a small dataset and worsening overfitting. Fine-tuning should instead freeze the pre-trained base and train a lightweight classifier head. Extra layers are warranted when underfitting a large dataset, not when regularising a small one.

• ✗

  Use a larger learning rate to speed up training.

  Why it's wrong here

  A larger learning rate accelerates weight updates, which on a small dataset drives the network to memorise training examples rather than generalise. Overfitting is countered by freezing early layers and training few parameters. Large learning rates suit short runs on abundant data, not fine-tuning a pre-trained model on limited samples.

• ✗

  Train the entire model from scratch on the custom dataset.

  Why it's wrong here

  Training from scratch discards the pre-trained weights and requires far more labelled data than a small custom dataset provides, so the model overfits. Transfer learning exists precisely to reuse those learned features. Scratch training is the right approach only when the domain differs radically and ample labelled data is available.

Go deeper

Related to this question