Scribd
Upload
9 views2 pages

Deep Learning - As2

The document outlines key components of deep learning, including convolutional, pooling, flatten, and fully connected layers, each serving a specific role in feature extraction and classification. It discusses the importance of pooling methods, the impact of filter sizes, and the number of convolutional layers based on task complexity. Additionally, it explains the initial challenges in training deep neural networks, such as weight initialization and gradient stabilization.

Uploaded by

dev.madusha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
9 views2 pages

Deep Learning - As2

The document outlines key components of deep learning, including convolutional, pooling, flatten, and fully connected layers, each serving a specific role in feature extraction and classification. It discusses the importance of pooling methods, the impact of filter sizes, and the number of convolutional layers based on task complexity. Additionally, it explains the initial challenges in training deep neural networks, such as weight initialization and gradient stabilization.

Uploaded by

dev.madusha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Deep Learning - AINT 44032

Assignment 2
CS/2019/056

1.​
a.​ Convolutional layer: Applies filters to the input to extract features
like edges or textures in an image, creating feature maps for further
processing.
b.​ Pooling layer: Reduces the size of feature maps by summarizing
regions, making the model faster and less sensitive to small shifts in
the input.
c.​ Flatten layer: Converts the 2D feature maps into a 1D vector so they
can be fed into fully connected layers for classification.
d.​ Fully connected layer: Connects all neurons to compute the final
output, such as class probabilities, by combining features learned
earlier.

2.​ Pooling in CNN reduces the size of feature maps while keeping important
information. It works by taking a small region, like a 2x2 area, and
summarizing it into one value, which helps reduce computation and makes
the model less sensitive to small changes. There are two main types:

a.​ Max pooling: Takes the largest value in the region, focusing on the
most prominent features, like strong edges.
b.​ Average pooling: Takes the average value in the region, which
smooths out the features but may lose some detail.

3.​ Deep neural networks have many layers, allowing them to learn hierarchical
features, such as edges, shapes, and objects in images. This makes them
better for complex tasks like image recognition or natural language
processing. Shallow neural networks, with only one or two layers, can’t
capture these complex patterns as effectively, so they often perform worse on
challenging problems.

4.​ The filter size in a CNN depends on the type of features you want to capture.
Smaller filters, like 3x3, are good for detecting fine details, such as edges or
textures in an image. Larger filters, like 5x5 or 7x7, capture broader patterns
but require more computation. The choice is often made through
experimentation, testing different sizes to see which works best for the
specific task.

5.​ The number of convolutional layers depends on the complexity of the


problem. Simple tasks, like classifying basic shapes, may need only a few
layers, while complex tasks, like identifying objects in detailed images,
require more. Start with a small number of layers, test the model’s
performance on validation data, and add more layers if needed. You also need
to consider computational resources, as more layers increase training time.

6.​
a.​ The weights are randomly initialized at the start, so the model’s
predictions are far from correct, leading to high error.
b.​ In the early epochs, the model hasn’t seen enough data to learn
meaningful patterns, as it needs multiple passes to adjust weights.
c.​ The initial gradients are often large because the model is far from the
optimal solution, causing the error to be high until the gradients
stabilize and the model begins to converge.

You might also like