Machine Learning

SYED AYAAN
SJR PUBLIC SCHOOL
ROLL NO. 25
XA
An Overview
● Machine Learning is a subset of AI that focuses on developing
algorithms capable of learning from data without being explicitly
programmed.

● Artificial neural networks have notably improved performance in

various tasks within ML.

● ML finds applications in natural language processing, computer

vision, speech recognition, and other fields.

● In business, ML is often used for predictive analytics.

● Mathematical optimization and data mining are foundational

principles that support ML techniques.
 Theory

PAC Model and Generalization Bias-Variance Decomposition Time Complexity and Feasibility
Computational learning theory, The bias-variance decomposition in Learning theorists evaluate time
particularly the PAC model, assesses learning theory highlights the complexity and the feasibility of
machine learning algorithms and their importance of matching hypothesis learning, determining if computation
performance with probabilistic bounds. complexity to data underlying function can be done in polynomial time and
for optimal performance. the efficiency of learning certain
functions.
Models and Types of Machine Learning Models

● Supervised Learning: Models are trained on labeled data to make predictions

or classifications.

● Unsupervised Learning: Models find patterns in unlabeled data, clustering

similar data points or reducing the data's dimensionality.

● Reinforcement Learning: Models learn to make sequences of decisions by

interacting with an environment, receiving rewards or penalties.

● Other Types: This includes semi-supervised learning and supervised learning

with limited labeled data, among other specialized techniques.
Pros and Cons of Machine Learning

Pros Cons
✓ Easily Identifies Trends: Machine learning can sift through ✕ High Error Susceptibility: Errors in machine learning can cause
large amounts of data to spot patterns humans might miss, problems, especially since machines work independently.
making it great for tasks like data mining. Detecting and fixing errors can be time-consuming and
challenging.
✓ Improves Over Time: With more data, machine learning
gets better at making decisions or predictions. For example, ✕ Takes Time to Bring Results: Machine learning needs time and
weather forecasts become more accurate with more resources to develop properly. Handling huge amounts of data
historical data. requires computing power, which can be costly.

✓ Adapts Without Human Intervention: Machine learning can ✕ Lack of Transparency: Machine learning models can sometimes
adapt quickly to new situations without needing human operate like "black boxes," meaning it's hard to understand how
input. This is particularly useful in security software, where they make decisions. This lack of transparency can be a
it can respond rapidly to emerging threats. challenge for users trying to understand or audit the system.
1. Artificial Neural Networks

● ANNs learn tasks through examples, mimicking how humans

learn.

● They are made up of interconnected artificial neurons processing

signals.

● Neurons transmit signals to connected neurons via weighted

edges.

● ANNs typically consist of layers of neurons performing different

transformations on inputs.

● The structure of ANNs is designed to process information and

learn patterns.
 2. Decision Trees

● Decision tree learning uses a tree-like structure for making

predictions based on observations.

● Classification trees are used to predict discrete outcomes, while

regression trees predict continuous values.

● Decision trees are valuable for visualizing and comprehending

complex decision-making processes.
3. Support Vector Machines

● SVMs are supervised learning methods used for classification

and regression.

● SVM training algorithm builds models to predict category

membership for new examples.

● SVMs can perform linear classification and employ the kernel

trick for non-linear classification.

● The kernel trick allows SVMs to map inputs into high-dimensional

feature spaces efficiently.
 4. Regression analysis
● Regression analysis estimates the relationship between input
variables and their associated features.

● Linear regression is the most common form, utilizing a straight line to

fit data based on a mathematical criterion like ordinary least squares.

● Ridge regression is an extension that uses regularization methods to

reduce overfitting and bias in linear regression.

● For non-linear problems, regression analysis includes models like

polynomial, logistic, and kernel regression to fit data and classify
variables.

● Regression analysis is widely used in various fields, including finance,

economics, and machine learning, to make predictions and
understand relationships between variables.
5. Genetic algorithms
● Genetic algorithms are based on the principles of natural
selection and evolution.

● They iteratively refine a population of solutions, favoring those

with desirable traits.

● Mutation and crossover are employed to generate new candidate

solutions in pursuit of optimal answers.
Applications
Machine learning is utilized in healthcare for medical diagnosis, predicting
Healthcare diseases, drug discovery, and personalized medicine. For example, ML
algorithms analyze medical images like MRI scans to detect tumors or anomalies.

In the finance sector, machine learning is applied for financial market analysis,
credit card fraud detection, and insurance risk assessment. For instance,
Finance
algorithms analyze transaction data to identify fraudulent activities or predict
stock prices.

Machine learning techniques are used in NLP for sentiment analysis, text
summarization, and machine translation. For example, NLP algorithms power
NLP
virtual assistants like Siri or Google Assistant to understand and respond to
spoken language.
 Applications
Computer vision enables machines to detect objects, faces, and drive
Computer
autonomously. For example, in self-driving cars, it detects pedestrians and
Vision
vehicles.

Machine learning powers recommendation systems for products, movies,

Recommendation
Systems and content based on user preferences. For instance, Netflix recommends
movies based on viewing history.

Machine learning optimizes online advertising by targeting ads to users

Online
interested in the products or services. Google Ads uses ML for bidding and
Advertising
ad placements.
 Relationships to other fields
Machine Learning
Data Compression
• Machine learning originated from the quest for artificial • Compression algorithms map strings into implicit feature space
intelligence (AI). vectors, facilitating similarity measures within these spaces.
• Early AI researchers explored symbolic methods and neural • Representative lossless compression methods like LZW, LZ77,
networks to enable machines to learn from data. and PPM are examined for their feature space implications.
• A shift towards logic-based AI caused a separation between AI • AIXI theory suggests the best compression of data is the
and machine learning. smallest software that generates it.
• Symbolic learning continued within AI, while machine learning • AI-powered audio/video compression software includes VP9,
embraced statistical methods, fuzzy logic, and probability theory. NVIDIA Maxine, AIVC, and AccMPEG.
• Machine learning shifted its focus from achieving AI to • Data compression improves storage efficiency and data
addressing practical problems. transmission speed by reducing file sizes.
• Techniques like backpropagation in the mid-1980s revitalized
machine learning as its own field.
 Relationships to other fields

Machine Learning Data Mining

• Data mining aims to uncover new insights, while ML
• Statistics draws inferences from samples to
emphasizes reproducing known knowledge.
populations, while machine learning finds predictive
• - ML's performance is evaluated on reproducing
patterns.
known knowledge, while KDD focuses on discovering
• Machine learning has roots in statistics and introduces
new knowledge.
the term "data science" to encompass the broader field.
• - ML involves optimization, where learning problems
• - Conventional statistics relies on selecting a model
minimize loss functions to align predictions with actual
beforehand and including theoretically relevant
instances.
variables, while machine learning lets the data shape
• - Loss functions measure the disparity between model
the model by identifying patterns.
predictions and real-world observations, crucial in
tasks like classification.

KDD - knowledge discovery in databases

Limitations to Machine Learning

Failure to Deliver Expected Black Box Theory Bias

Results

Machine learning programs may Some machine learning algorithms Machine learning systems can
struggle due to lack of suitable function as 'black boxes,' perpetuate biases existing in their
data, data bias, privacy issues, complicating the understanding of training data, resulting in
poorly chosen tasks and their decision-making process. This discriminatory outcomes. Biases
algorithms, and insufficient opacity is particularly concerning in have been observed in language
resources for evaluation. critical applications like self-driving models and image recognition
cars and healthcare. algorithms against certain
demographic groups.
Limitations to Machine Learning

Overfitting Learning the Wrong Explainability

Lesson

Overfitting happens when a Machine learning models Explainable AI (XAI) strives

model learns noise instead of can grasp onto spurious to enhance the
general patterns, potentially interpretability of machine
correlations or irrelevant
leading to poor performance learning models for humans,
features, causing
on new data.
promoting trust and
inaccurate conclusions or
understanding in AI
predictions.
decisions.
THANK YOU!
SYED AYAAN
XA
ROLL NO. 25