Assignment 8
Ahzam Zobairi
Deep Learning
September 6, 2024
Answer 1:
I would choose the following three design for my architecture
Convolutional Neural Network (CNN) Architecture :
• Description: Utilize a pre-trained CNN, such as VGG16 or ResNet50, which has been trained on a large dataset
like ImageNet.
Transfer Learning with Feature Extraction:
• Description: Use the pre-trained model as a fixed feature extractor by removing the final classification layer
and adding a custom classifier on top.
Fine-Tuning the Model:
• Description: After feature extraction, selectively unfreeze some of the deeper layers of the pre-trained model
to allow for fine-tuning on the specific dataset.
Answer 2:
The merits and demerits of the above 3 design architecture are as follows
Convolutional Neural Network (CNN) Architecture
• Merits:
– High accuracy in image classification tasks due to hierarchical feature learning.
– Ability to leverage transfer learning, reducing training time and resource requirements.
• Demerits:
– Requires substantial computational resources for training and fine-tuning.
– May overfit if the dataset is small or not diverse enough.
Transfer Learning with Feature Extraction
• Merits:
– Significantly reduces training time since the model uses pre-learned features.
– Effective when labeled data is scarce, as it utilizes knowledge from large datasets.
• Demerits:
– The model may not adapt well to the specific characteristics of the new dataset if the domain differs
significantly.
– Limited flexibility since the base model’s layers are not retrained.
Fine-Tuning the Model
• Merits:
– Allows the model to adapt better to the specific dataset by adjusting weights in deeper layers.
– Can improve performance on tasks where the pre-trained model’s features are not fully aligned with the
new data.
• Demerits:
– The model may not adapt well to the specific characteristics of the new dataset if the domain differs
significantly.
– Limited flexibility since the base model’s layers are not retrained.
Answer 3:
In the context of image classification, fine-tuning deeper layers is generally a better choice. This is because:
• Deeper Layers Capture High-Level Features: Deeper layers in a CNN are responsible for capturing more abstract
features, which are crucial for distinguishing between classes in complex datasets. Fine-tuning these layers allows
the model to learn these specific high-level features relevant to the new task.
• Shallow Layers Generalize Well: The initial layers of a CNN often capture low-level features (like edges and
textures) that are common across many datasets. These features are usually transferable and do not require
significant adjustment. Fine-tuning shallow layers may lead to unnecessary changes in these well-established
features, potentially degrading performance.
• Adaptation to New Data: Fine-tuning deeper layers allows the model to adapt to the nuances of the new dataset
while retaining the foundational knowledge learned from the original dataset. This balance is crucial for achieving
high accuracy in image classification tasks.
