Page 291 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

282 Chapter 10 Deep neural network in medical image processing

2.10 Unsupervised learning
Unsupervised learning is the opposite of supervised learning.
Unsupervised learning works on unlabeled data sets. This type
of algorithm works by finding structures in data and analyzes
vast amount of data to understand critical properties of the
data so that it can learn to divide data into clusters based on
those properties in a way that a human being can make sense
out of the newly formed clusters. What makes unsupervised
learning such an interesting proposition is the fact that it can
work on unlabeled data; therefore, it becomes the algorithm of
choice in cases where data are readily and cheaply available,
but labels are either very expensive or not available, such as
images in image galleries (putting labels on these images is very
time-consuming). Another interesting example would be a collec-
tion of news articles and running an unsupervised algorithm to
categorize the data to find the most important or trending
news topics, or recommendation system used by majority of
e-commerce players recommends products based on available
clusters in the database based on various parameters such as
how they are frequently bought together by different users or
how those items are related to each other. Different types of tasks
that can be performed using unsupervised learning are as follows:
• Recommendation systems
• Grouping user logs
• Grouping images

2.11 Reinforcement learning
Reinforcement learning algorithms are one of the most used
algorithms in modern-day machine learning environment. In
layman terms, it is about learning through one’s own experience.
It uses a reward-based system where the agent starts off doing
random actions, and when it performs the task right, it is
rewarded; therefore, it gets smarter over time. This type of
approach has its own pitfalls; however, one of them being that
agent will perform certain actions once it reaches the action
that gives it maximum reward, which may keep on repeating
the action and not exploring any other better approaches that
may exist. To stop this undesirable outcome, we introduce some-
thing called the greedy approach. In this approach, a randomness
coefficient is added, its value is initially very high as the agent has
no fixed path, but even when the agent discovers the path to
reward, this coefficient still forces it to preform random action
although much lesser than what it did in the beginning.