Scribd
Upload
13 views

Lecture 4 - Visualizing What Convnet Learn

Uploaded by

Joel Lim
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
13 views

Lecture 4 - Visualizing What Convnet Learn

Uploaded by

Joel Lim
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 26

Official (Closed) - Non Sensitive

Deep Learning in Image


Recognition
Lecture 4:
Visualizing What Convnets
Learn

• Specialist Diploma in Applied


Generative AI

• Academic Year 2024/25


Official (Closed) - Non Sensitive

Topics

1. Introduction to Visualizing
Convnets
2. Activation Visualization
3. Visualizing Convnets Filters
4. Visualizing Heatmaps of Class
Activation
Official (Closed) - Non Sensitive

1. Introduction to Visualizing Convnets


Official (Closed) - Non Sensitive

Deep Learning Models as Black


Boxes
 Deep-learning models as black boxes
• Learning representations that are hard to
extract
• Not true for convnets
 Convnets and visual concepts
• Representations are highly amenable to
visualization
• Representations of visual concepts
 Techniques for visualization and
interpretation
• Wide array of techniques since 2013
• Three accessible and useful techniques
Official (Closed) - Non Sensitive

Techniques for Visualizing


Convnets
1. Visualizing intermediate
convnet outputs
• Helps understand how
successive convnet layers
transform their input
• Provides insight into the
meaning of individual convnet
filters
Official (Closed) - Non Sensitive

Techniques for Visualizing


Convnets
2. Visualizing convnet filters
• Helps understand what visual
pattern or concept each filter is
receptive to
3. Visualizing heatmaps of class
activation in an image
• Helps understand which parts of
an image were identified as
belonging to a given class
• Allows localization of objects in
images
Official (Closed) - Non Sensitive

2. Activation Visualization
Official (Closed) - Non Sensitive

What is intermediate activation


 Intermediate activations are the outputs
generated by each layer in a Convnet when
an image is passed through it
Official (Closed) - Non Sensitive

Why Visualize intermediate


activation
 Ability to understand what the model
“sees” or “focuses on” at each layer
 Useful for debugging and interpreting
model behavior
 Able to provide explainability to users
Official (Closed) - Non Sensitive

Steps to Visualize
 Step 1:
• Load the model and prepare the image
 Step 2:
• Create a model that outputs layer activations
 Step 3:
• Generate activations for an input image
 Step 4:
• Visualize the activations
Official (Closed) - Non Sensitive

Let’s try your hands on visualizing our model!


(Practical 4 - Part 1: Visualizing intermediate
activation)
Official (Closed) - Non Sensitive

What was seen


Official (Closed) - Non Sensitive

3. Visualizing Convnets Filters


Official (Closed) - Non Sensitive

What is filter visualization


 A technique to understand and interpret
what individual filter in a convnet are
responding to in input images
 Involves visualizing the patterns or
features that activate each filter
Official (Closed) - Non Sensitive

Why filter visualization


 Interpretability: It helps us understand the inner
workings of convnets by showing us what each
filter detects in an image
 Debugging: By examining filters, we can diagnose
potential issues, such as overfitting or filters not
learning meaningful patterns
 Feature Understanding: Visualization highlights
the progression from low-level to high-level feature
detection as we move deeper into the network
Official (Closed) - Non Sensitive

How filter visualization works


 Filter visualization is generally achieved by
maximizing the activation of each filter, using
techniques like gradient ascent
 This process essentially creates synthetic
images that show which types of patterns cause
a high response in a particular filter
 These synthetic images provide a visual
representation of the filter’s learned features
Official (Closed) - Non Sensitive

Let’s try your hands on visualizing our model!


(Practical 4 - Part 2: Visualizing convnet filters)
Official (Closed) - Non Sensitive

How filter looks like


Official (Closed) - Non Sensitive

4. Visualizing Heatmaps of Class


Activation
Official (Closed) - Non Sensitive

Introduction to Class Activation


Maps
 Understanding Image Classification Decisions
• Useful for debugging convnet decisions
• Helps locate specific objects in an image
 Class Activation Map (CAM) Visualization
• Produces heatmaps of class activation over input images
• 2D grid of scores for a specific output class
• Indicates importance of each location in the image
 Application Example
• Dogs vs. Cats convnet
• Heatmap for class
Official (Closed) - Non Sensitive

Gradient Class Activation Maps


(Grad-CAM)
 Uses output feature map of a convolution layer
 Weigh every channel in the feature map by the
gradient of the class with respect to the channel
 Create a spatial map of how intensely the input
image activates different channels

Ramprasaath R. Selvaraju et al., arXiv (2017), https://arxiv.org/abs/ 1610.02391


Official (Closed) - Non Sensitive

Let’s try your hands on visualizing our model!


(Practical 4 - Part 3: Visualizing heatmaps of class
activation in an image)
Official (Closed) - Non Sensitive

What Grad-CAM generated

Ramprasaath R. Selvaraju et al., arXiv (2017), https://arxiv.org/abs/ 1610.02391


Official (Closed) - Non Sensitive

Wrapping up
 Understanding ConvNet
learning and representation
 Visualization techniques
applied on activation, filter
and heatmaps
 Understand model focus to
provide interpretability
Official (Closed) - Non Sensitive

Q&A
Official (Closed) - Non Sensitive

References
Books:

François Chollet, Deep Learning with Python

Online Resources:

You might also like