Machine Learning: AI Hierarchy and Types
Key Points
- Machine learning (ML) is a subset of artificial intelligence (AI) that uses algorithms to learn patterns from training data and make predictions on new, unseen data, while deep learning (DL) is a further subset of ML that employs multi‑layered neural networks.
- The core process of ML involves training a model on a representative dataset so it can perform accurate inference—running the trained model on fresh inputs to generate predictions.
- ML learning paradigms fall into three main categories: supervised learning (using labeled data), unsupervised learning (discovering structure in unlabeled data), and reinforcement learning (optimizing behavior through trial‑and‑error rewards).
- Supervised learning includes regression models that predict continuous values (e.g., linear and polynomial regression) and classification models that predict discrete categories (e.g., binary fraud‑vs‑legit classification).
- Unsupervised techniques such as clustering, dimensionality reduction, and anomaly detection enable models to uncover hidden patterns without human‑provided labels.
Sections
- Machine Learning, AI, and Deep Learning Explained - The speaker outlines how AI encompasses machine learning, which in turn includes deep learning, and describes model training and inference.
- Clustering Techniques and Dimensionality Reduction - The speaker explains k‑means and hierarchical clustering for segmenting data and introduces dimensionality‑reduction methods that compress feature spaces while preserving essential information.
Full Transcript
# Machine Learning: AI Hierarchy and Types **Source:** [https://www.youtube.com/watch?v=znF2U_3Z210](https://www.youtube.com/watch?v=znF2U_3Z210) **Duration:** 00:10:27 ## Summary - Machine learning (ML) is a subset of artificial intelligence (AI) that uses algorithms to learn patterns from training data and make predictions on new, unseen data, while deep learning (DL) is a further subset of ML that employs multi‑layered neural networks. - The core process of ML involves training a model on a representative dataset so it can perform accurate inference—running the trained model on fresh inputs to generate predictions. - ML learning paradigms fall into three main categories: supervised learning (using labeled data), unsupervised learning (discovering structure in unlabeled data), and reinforcement learning (optimizing behavior through trial‑and‑error rewards). - Supervised learning includes regression models that predict continuous values (e.g., linear and polynomial regression) and classification models that predict discrete categories (e.g., binary fraud‑vs‑legit classification). - Unsupervised techniques such as clustering, dimensionality reduction, and anomaly detection enable models to uncover hidden patterns without human‑provided labels. ## Sections - [00:00:00](https://www.youtube.com/watch?v=znF2U_3Z210&t=0s) **Machine Learning, AI, and Deep Learning Explained** - The speaker outlines how AI encompasses machine learning, which in turn includes deep learning, and describes model training and inference. - [00:05:23](https://www.youtube.com/watch?v=znF2U_3Z210&t=323s) **Clustering Techniques and Dimensionality Reduction** - The speaker explains k‑means and hierarchical clustering for segmenting data and introduces dimensionality‑reduction methods that compress feature spaces while preserving essential information. ## Full Transcript
What is machine learning? Well, I think you have a basic idea. Right? It's the tech that
recommends which video to watch next on YouTube. The machine is learning your preferences
and its bot powers chatbots. But is ML or machine
learning synonymous with AI? Or what about deep
learning? DL. Are deep learning and machine learning the same thing? Well, no,
it's a hierarchy. So, machine learning is a subset of
AI that focuses on algorithms that can learn the patterns of training data and then make
accurate inferences about new, unseen data. Essentially, the machine is learning pattern
recognition that lets models make decisions or predictions without explicit hard-coded
instructions. And then deep learning is a subset of machine learning that uses neural
networks with many layers to learn hierarchical representations. So, ML sits inside
AI and then DL sits inside ML. Now the central premise of
machine learning is that if you optimize the machine's performance on a dataset of tasks that
adequately resemble the real world, that's through a process called model training.
Well, if you train the model in such a way, that model can make accurate predictions on new
data. In essence, a trained model is applying patterns. It's learned from training data to infer
the correct output for a real-world task. And then the deployment of this fully trained
model, that is called AI inference. this is when we actually run the model.
That's when we feed new data to this trained model. And it makes predictions based on the
patterns it's learned. Now most machine learning is grouped into three learning
paradigms. So there is supervised learning, that trains a model to predict the
correct output for an input using labeled examples or ground truth. For instance, we might
have a bunch of emails and they've been labeled spam or not spam. It's supervised because it
generally requires a human in the loop to provide the ground truth data. Then there is
unsupervised learning, and that uses unlabeled data to
discover structure on its own. So that's the tasks like clustering and dimensionality reduction and
anomaly detection. And then there is also reinforcement learning, which optimizes a
policy through trial and error with rewards and penalties. Now there are several types of
supervised machine learning models. Let's take a look at some of them. So one type of
model is called a regression model. And that predicts continuous
numerical values so think like price prediction or think temperature forecasting. Now there
are different types of regression. So one type of regression is called
linear regression that finds the best fit line through data points. Another
type is called polynomial regression. And that captures
nonlinear relationships. So that's regression. There's also
classification. That's another type of supervised learning. And that predicts
discrete classes. So one type of classification might be binary
classification. So fraud or legit it's gonna be one or the other. It might be
multi-class classification. That's where we have one
of many categories. Or it might be multi-label
classification as well. That's where we have multiple tags all at once. And modern
supervised learning. It actually frequently uses a combination of this called ensemble
methods which means combining models for better accuracy. Oh, and before we get to
unsupervised learning, I should mention there's something that kinda sits between here called
semi-supervised learning that sits between these two because you train with a small
labeled data set plus a large unlabeled pool to improve the supervised model. it
uses the labeled examples to generalize over the unlabeled data. So you need far fewer costly
labels. All right. Now unsupervised learning that works with unlabeled data to uncover structure on
its own. So let's look at two common families. One of those families is called
clustering. Now clustering groups similar items so things that behave alike end up together.
A well-known clustering method is called k-means clustering, so you
choose k-groups. Repeatedly assign each item to the nearest group average. Then you recompute the
averages until they're stable. So for example, you could split customers into k equals four segments.
And those four segments might be ah bargain hunters, loyal repeaters, big ticket buyers and just
browsers. And then you could target offers to each of those groups. Now another clustering method is
called hierarchical. And hierarchical is where we start with every item by
itself. Then keep merging the most similar groups to build a tree, and you cut the tree later for
let's say, three or 5 or 12 clusters. So for example, you could cluster IT ticket text. So
that's the subject and the description of the ticket. You could cluster those into themes like
like password reset and laptop won't boot. And you could get broad buckets, or you can have
fine-grained sub-buckets, depending on where you cut the tree. And that's useful for autorouting
and spotting new service issue types. Now another type of unsupervised learning is
called dimensionality reduction. Now dimensionality
reduction are algorithms that reduce the complexity of data points by representing them
with the smaller number of features that is fewer dimensions while still kinda keeping all of the
meaningful characteristics. And they're often used for preprocessing data, as well as for tasks such
as data compression or data visualization. Now, some examples of dimensionality reduction are
PCA. That's principal component analysis. And we also have
encoders as another example. Then there's reinforcement learning. So an
agent interacts with an environment. And it does that by observing its state.
That's what it sees now, and then by choosing an action
to take. So what to do next. And then the environment either will
reward that action, which is to say it moves correctly, ah or
it will punish that action with a penalty for the incorrect
action. And over time, the interactions with the agent teaches it a policy that
maximizes long-term rewards because rewards can be delayed. An agent might also have to balance
exploration of trying new actions with exploitation, which is repeating what works. So as
an example, consider a self-driving car. So the state that comes from the from the
GPS and the cameras, and well, for most cars the lighter as well. And the actions
are to steer and to break and to accelerate and the model rewards the safe and
move ah the smooth progress like, well, staying in its lane and obeying traffic signals. it will
penalize things like hard braking, and it will heavily penalize the really bad stuff like
collisions. The the rewards and the penalties help the model learn. Now, most of what
we've covered here regression, classification, clustering and reinforcement learning that stuff
all kind of falls under what you might call classic machine learning. These techniques have
been around for years and are still everywhere in business today, from predicting prices to
segmenting customers. But today, we're also seeing these old ideas applied in new
ways. And of course, the most famous example of that is LLMs or large
language models. And those are built on top of something called a
transformer. And these transformer architectures are a newer neural network design, but they still
rely on the same ML principles of pattern recognition and data and model training and
inference. And in fact, even reinforcement learning has made a bit of a comeback with
RLHF. That's reinforcement learning with human feedback. So instead of teaching an
agent to say, play a game, RLHF trains LLMs to better align with human
preferences. So human annotators, they provide rewards and penalties on model outputs, which fine
tunes the system's behavior. So, while the buzz today is all around LLMs and generative AI and
agentic AI, the foundations are still classic ML concepts. just scaled up
and combined in new ways, which just goes to show that human learning keeps finding new ways to
apply machine learning.