Page 24 - Toucpad robotics C11

Semi-Supervised Learning
Let’s go back to our classroom example. Your teacher gives you a mix of solved and unsolved problems. There are only a
few solved ones, but a huge number of unsolved ones. You first study the solved examples to get a basic understanding.
Then, you use that knowledge to try and solve the other problems. As you work through them, you gain a deeper
understanding of the entire topic.
Self-supervised learning operates on this principle. It uses a small amount of labelled data to get a head start, and then
uses a much larger amount of unlabelled data to complete the learning process. This is particularly useful when getting
labelled data is difficult or expensive, like in many real-world scenarios.

Example: Consider a medical imaging system designed to detect a rare disease. Getting a doctor to manually label
u
thousands of X-ray images for the disease is extremely time-consuming and costly. Instead, you can use a small
set of labelled images (those already diagnosed by a doctor) to train the model. The model then uses this initial
knowledge to analyse a huge dataset of unlabelled X-rays, using patterns it learned to identify potential cases on its
own, which can then be verified by a doctor.
Reinforcement Learning
Think about learning to ride a bicycle. You don’t have a textbook with step-by-step instructions. Instead, you learn by
doing. You pedal a little, and if you stay balanced, you feel a sense of achievement (a reward). If you wobble and fall, it’s
a negative experience (a penalty). Over time, you learn which actions (like turning the handlebars or leaning your body)
lead to rewards and avoid the actions that lead to penalties, until you can ride perfectly.
This is the core of Reinforcement Learning. An agent (the machine) learns to make decisions by interacting with an
environment. It receives rewards for ‘good’ actions and penalties for ‘bad’ actions. Its ultimate goal is to learn a strategy,
or a policy, that maximises its total reward over time.

Example: Have you ever seen a robotic arm in a factory? Using reinforcement learning, you can train a robotic arm
u
to pick up an object and place it in a box. You don’t program every single movement. Instead, you give it a reward
when it successfully places the object and a penalty if it drops it. The robot learns through a process of trial and error,
adjusting its movements and grip until it masters the task with a high success rate.
Self-Supervised Learning How Does Machine Learning Work?

Finally, let’s consider a new way of learning. Imagine
you’re given a book with half the words missing from
each sentence. You’re asked to fill in the blanks. By
understanding the context of the words around the blank Training
space, you can figure out what the missing word should data
be. You’re essentially creating your own learning task from (Unacceptable) (Acceptable)
the data itself.
This is the essence of Self-supervised learning. The model
generates its own labels or ‘tasks’ from the unlabelled Train ML Accuracy Successful
input data. It learns by solving these created tasks, algorithm model
which helps it understand the underlying structure and
representations of the data. This pre-training makes the
model very powerful and ready for other, more specific
tasks later on. Model input
Example: A great example is data
u
training a model to understand a
language. The model is given vast
amounts of text from the internet.
New input ML algorithm Prediction
data
22
Touchpad Robotics - XI