Why transfer learning saves time and data in TensorFlow - Challenge Your Understanding

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why does transfer learning reduce training time?

Imagine you want to teach a model to recognize cats in photos. You start with a model already trained to recognize many objects. Why does this help reduce the time needed to train your cat-recognizing model?

A Because the model already knows useful features like edges and shapes, so it needs less time to learn new tasks.

B Because the model ignores previous knowledge and starts fresh, making training faster.

C Because transfer learning uses a smaller dataset which automatically speeds up training.

D Because transfer learning removes layers from the model, making it simpler and faster.

💡 Hint

Think about how learning to recognize letters helps you read new words faster.

Transfer learning uses a model that already learned basic features like edges and textures. This means it doesn't have to learn these from scratch, saving time.

🧠 Conceptual

intermediate

How does transfer learning reduce the amount of data needed?

Why can transfer learning work well even if you have only a small dataset for your new task?

A Because transfer learning creates synthetic data to increase dataset size automatically.

B Because the model already learned general features from a large dataset, it needs fewer new examples to adapt.

C Because transfer learning removes the need for data by using random weights.

D Because transfer learning trains only the output layer, ignoring all other layers.

💡 Hint

Think about how knowing the alphabet helps you read new words without seeing many examples.

The model has already learned useful features from a large dataset, so it can adapt to new tasks with fewer examples.

❓ Predict Output

advanced

Output of transfer learning model training metrics

Given this TensorFlow code snippet that uses transfer learning, what will be the shape of the output predictions?

TensorFlow

import tensorflow as tf
base_model = tf.keras.applications.MobileNetV2(input_shape=(96,96,3), include_top=False, weights='imagenet')
base_model.trainable = False
model = tf.keras.Sequential([
    base_model,
    tf.keras.layers.GlobalAveragePooling2D(),
    tf.keras.layers.Dense(5, activation='softmax')
])
import numpy as np
sample_input = np.random.rand(1,96,96,3).astype('float32')
predictions = model(sample_input)
print(predictions.shape)

A (1, 5)

B (96, 96, 3)

C (1, 1000)

D (5,)

💡 Hint

Look at the last Dense layer's output size and the batch size of the input.

The model outputs a batch of predictions with shape (batch_size, number_of_classes). Here batch size is 1 and classes are 5.

❓ Hyperparameter

advanced

Choosing layers to train in transfer learning

You have a pretrained model and want to use transfer learning. Which approach best balances training time and adapting to a new task?

A Freeze all layers and train only the input layer.

B Train all layers from scratch to fully adapt the model to the new data.

C Freeze most layers and train only the last few layers to adapt features to the new task.

D Freeze only the first layer and train all others.

💡 Hint

Think about keeping learned features but allowing some adaptation.

Freezing most layers keeps learned features and training only last layers adapts the model efficiently without long training.

❓ Metrics

expert

Interpreting transfer learning training curves

You train a transfer learning model and see training accuracy quickly reaches 95%, but validation accuracy stays around 70%. What is the most likely explanation?

A The learning rate is too low, preventing the model from learning.

B The model is underfitting and needs more training epochs.

C The pretrained weights are not loaded correctly, causing poor validation.

D The model is overfitting the small new dataset and not generalizing well.

💡 Hint

Think about what it means when training is good but validation is poor.

High training accuracy but low validation accuracy usually means the model memorizes training data but fails to generalize, a sign of overfitting.