Image analysis
Working with Convolutional Neural Nets
N Dethlefs (mini) 2/3 Dec 2021 - Understanding AI
Convolutional neural nets
Reduce images into a set of features that makes them easier to
process without losing vital information for classi cation.
fi
Convolutional neural nets
Stage 1 Stage 2
Reduce images into a set of features that makes them easier to
process without losing vital information for classi cation.
fi
Convolutional layer
When computing convolutions, we will normally have more than
one lter (with di erent weights). This leads to multiple activation
maps that get combined later on (in the fully-connected layer).
fi
ff
Pooling Layer
The pooling layer reduces the size of the “convolved feature” and
extracts dominant features. It can do this via max pooling or via
average pooling (normally).
Feature learning
We then have an internal feature representation…
Feature learning
#4 Learning mul
of representation
[
Successive%model%layers%learn%deeper
%
Layer 3
Layer%3%
The feature learning stage is
important as it extracts detail from
Layer 2
the images at di erent level of
Layer%2%
abstraction and granularity.
Layer 1
12% Layer%1%
ff
Fully-connected Layer
attened representation
hidden
layers
outputs
inputs
Flatten the inputs (multiple activation maps, one per kernel/
lter) and feed them into the nal layer/s.
fi
fl
fi
CNNs with Keras
Prior knowledge
… in machine learning
N Dethlefs (mini) 2/3 Dec 2021 - Understanding AI
Prior Knowledge in ML
• Feature extraction
• Using a pre-trained model and extracting features (from
existing weights) to improve performance in a new domain
• Fine-tuning
• Using a pre-trained model and updating its weights for a given
new target domain
Transfer learning
Source Target
domain Learnt weights… domain
Different types of pre-training
Reporting experiments
… Part 1 and 2 of your assessment
N Dethlefs (mini) 2/3 Dec 2021 - Understanding AI
Choosing baselines
• What is a baseline?
• A point of reference that helps you assess the performance of
your learning model (by comparing it against something)
82% - good or bad?
Choosing baselines
• What is a baseline?
• A point of reference that helps you assess the performance of
your learning model (by comparing it against something)
82% - good or bad?
76% 85% 93%
Choosing baselines
• What is a baseline?
• A point of reference that helps you assess the performance of
your learning model (by comparing it against something)
82% - good or bad?
Majority
76% 85% 93%
Choosing baselines
• What is a baseline?
• A point of reference that helps you assess the performance of
your learning model (by comparing it against something)
82% - good or bad?
Majority baseline
50% of getting it right when voting for the majority
Choosing baselines
• What is a baseline?
• A point of reference that helps you assess the performance of
your learning model (by comparing it against something)
82% - good or bad?
Majority baseline
90% of getting it right when voting for the majority
Reporting experiments
• Choose a set of meaningful baselines
• Think about your problem and decide what makes sense
• Report a meaningful metric
• e.g. accuracy, report train and validation / test accuracy
• Report details to make your experiments replicable
• From your report, others should be able to re-run your
experiments and get the same results
Model Train acc Test acc Train loss Test loss
Other “baselines”,
Neural net
Models, …
Alterations etc.
…
Majority
baseee
Critical reflection
… Part 3 of your assessment
N Dethlefs (mini) 2/3 Dec 2021 - Understanding AI
Discussion of Ethics in AI
• Highlight the ethical challenge
• Brief objective description of challenge - in your own words!,
refer to chosen article (cite it), no need to quote, provide
examples (optional)
• Describe researchers’ approach
• Objective description, in your own words
• Discuss areas of applied AI where similar challenges may
occur and what incentives can be provided to address them
• Evaluation and discussion of your own point of view; elements
of critical re ection, use additional academic sources, draw
your own informed conclusions.
fl
Plagiarism
… don’t copy anything without reference
N Dethlefs (mini) 2/3 Dec 2021 - Understanding AI
Use References for everything!
• You cannot use materials from others without referencing
them and citing their original source
• This applies to text, images, code, etc.
• Plagiarised work will automatically
fail (in the best case)
TA holiday…
… aka PhD work!
N Dethlefs (mini) 2/3 Dec 2021 - Understanding AI
