Understanding Machine Learning

• Learning involves changes in a system, enabling it to perform tasks more efficiently over time
• Learning is a feature of Intelligence.
• In AI, an agent is referred to as an intelligent agent

What is Machine Learning?

• Machine Learning is a field of study that equips computers with the ability to learn without explicit programming (Arthur Samuel 1959)
• Humans learn from experience
• Computers and machines follow instructions.
• Machine learning enables machines to learn from experiences or past data.

Traditional Programming vs Machine Learning

• Traditional programming involves feeding data and a program into a computer to generate an output.
• Machine learning involves feeding data and the desired output into a computer to generate a program

Information Systems vs. Intelligent Systems

• Information systems include the internet, the world wide web, wireless fax, digital libraries, data mining, and information retrieval
• Intelligent systems include smart cameras, smart appliances, smart elevators, smart robots, intelligent manufacturing, expert systems, smart search engines, and smart quality control

Machine Learning Definition

• ML involves algorithms that learn from data to make predictions about new data.
• y = f(x; θ) where:
  • f is the model.
  • θ represents the parameters of the model.
  • x represents features.
  • y represents the model's prediction

How Machine Learning Works

• Data is processed through a machine learning algorithm to predict, learn, and improve.

When to Use Machine Learning

• Machine learning is useful in situations where human expertise is lacking, such as navigating on Mars
• It also applies when humans struggle to articulate their expertise, like in speech recognition
• ML is useful when data volumes are too large for manual analysis, such as in medical diagnostics
• It is valuable when solutions change over time, like in computer network routing
• It helps when new knowledge is constantly emerging, and continuous system redesign is needed
• Machine learning is helpful when tasks are hard to precisely define, except through examples, like recognizing people
• It can uncover relationships and correlations hidden in large datasets, such as through data mining

Basic Principles of Machine Learning

• Machine Learning is a branch of AI focused on developing algorithms that enable computers to learn automatically from past data and experiences
• Machine learning uses algorithms to build mathematical models and make predictions, using historical data/information
• Historical data samples, known as training data, help ML algorithms construct mathematical models for making predictions or decisions

Machine Learning Systems

• Machine learning systems learn from historical data and then build prediction models
• Upon receiving new data, those systems can then predict its output.
• Prediction accuracy is correlated with the amount of data used so more data generally builds better models
• Machine learning involves feeding data into an algorithm, which then builds a logic to predict the output for new data rather than writing specific problem solving code
• This process has changed the conventional approach to problem-solving

Machine Learning Processes

• Data is input, then analyzed; patterns are found, predictions made, and feedback stored

Features of Machine Learning

• Machine learning identifies patterns in data, learns, and improves automatically from past data
• ML is a data-driven technology that can handle large data volumes

Learning Algorithms

• Learning algorithms write code that discerns patterns in data.
• These algorithms create statistical models approximating the data.

Classic Machine Learning Task

• Recognizing what makes the number "2" as a classic ML task

Challenges of Machine Learning

• High Dimensionality: increased data complexity necessitates larger models requiring more memory/time which in turn might cause overfitting
• Statistical Model Selection: Choosing an effective models with appropriate parameters is crucial
• Noise and Errors:
  • Outliers are observations that are far from the rest of the data
  • Human errors can also cause incorrect measurements
• Insufficient Training Data: the amount of data may be insufficient to build a good process approximation for generated data
• Feature extraction: Converting data to a reduced representation made of a series of features

Overfitting

• Overfitting occurs when a model trains on sample data for too long or is too complex, causing it to learn "noise" or irrelevant dataset information
• When a model memorizes the noise and fits too closely to the training set, it becomes overfitted and is unable to generalize well to new data, limiting its ability to classify or predict tasks.

Types of Machine Learning

• Supervised learning: uses labeled data for training
• Unsupervised learning: uses unlabeled data for training
• Reinforcement learning: the model takes actions in the environment and receives state updates and feedback

1 - Supervised Learning

• Supervised learning is a machine learning method that uses labeled data to train a system in order to predict outcomes
• With supervised learning, the system creates a model using labeled data to understand datasets
• After training, the model is tested with sample data to see if it can output the correct result.
• The intent of supervised learning is to map input data to output data.

Steps for Supervised Learning

• Training: involves giving the learner examples in D train which returns model g(x)
• Testing: which involves reserving some labeled data

Testing and Evaluation

• Apply model to the raw test data.
• Predicted labels are measured against test labels to evaluate.

Supervised Learning Prediction Tasks

• Regression: Examples include determining housing prices based on location and year
• Binary classification: Examples include identifying whether an email is spam

Regression in Detail

• Regression aims to estimate a response
  • Input: {x1, x2,..., xn}, called "features" are numeric values
  • Output: called "target value" are numeric values

Classifications in Detail

• Classification aims to identify group membership
  • Input: {x1, x2,..., xn}, called "features" are categorical values
  • Output: called "target value" are categorical values

Supervised Learning Algorithms

• Supervised learning algorithms include neural networks, SVM, KNN, decision tree, naive bayes, logistic regression, linear regression, multi linear regression, and polynomial regression

Advantages of Supervised Machine Learning

• Supervised Machine Learning allows for precise definition of labels and determining the number of classes
• The method provides well known, labeled input data for more accurate/reliable results

Disadvantages of Supervised Machine Learning

• Supervised machine learning can be complex to understand and label inputs
• The process doesn't take place in real time, as unsupervised learning does
• A lot of computation time is needed for training

2 - Unsupervised Learning

• Unsupervised learning is a method in which a machine learns without any supervision from data that has no labels.
• The goal is to restructure input data into new features or groups of objects with similar patterns
• This approach tries to find useful insights from large amounts of data without a predetermined result

Types of Unsupervised Learning

• Clustering: assigns a set of objects into subsets with similar traits (clusters)
• Association: Discovers relationships between variables in large databases

Clustering

• Clustering assigns a set of objects into subsets called clusters, so that objects in the same cluster have similar characteristics
• The goal of clustering is to segregate groups by traits and assign them to clusters.
• Clustering groups inputs based on similarities.

Clustering Example

• A survey with questions on a scale of 1-10:
  • How much do you like shopping?
  • How much are you willing to spend on shopping?
  • Cluster 1 refers to people who are dedicated to shopping
  • Cluster 2 refers to people who rarely go shopping

Association

• Association discovers interesting relationships between variables in large databases
• Association rules aim to find links amongst data objects within databases
• As an example, people that buy a new house also tend to buy new furniture

Unsupervised Machine Learning Algorithms

• Algorithms include K-means, k-medoids, fuzzy C-means, hidden markov models, neural networks, and Gaussian mixture

Advantages of Unsupervised Machine Learning

• Offers less complexity than supervised learning, takes place in real-time, and is easier to source unlabeled data

Disadvantages of Unsupervised Machine Learning

• You can’t get very specific about the definition of the data sorting and the output
• Less accuracy of the results because the input data is not known and not labeled
• The results of the analysis can't be known

3 - Reinforcement Learning

• Reinforcement learning involves a learning agent that gets rewarded for right actions and penalized for wrong actions
• The agent learns automatically from feedback and improves its performance
• Input raw data -> Agent -> Environment -> Output

Reinforcement Learning Example

• Creating a new game, RL is an efficient/relatively easy resource that programmers can use for testing bugs and defeating levels
• In contract to traditional games' complex code, training an RL model is simpler, allowing agents to learn by themselves within a simulated game environment through navigation, defense, attack, and strategizing
• Through trial and error, actions will be performed to reach the end goal
• RL agents are also used in bug and game testing due to their ability to run testing through countless iterations

Applications of Machine Learning

• Machine learning is used in image recognition, automatic language translation, medical diagnosis, speech recognition, stock market trading, traffic Prediction, online fraud detection, virtual personal assistants, product recommendations, self-driving cars, and email spam/malware filtering

How Machine Learning Works Overall

• Training data -> Train ML algorithm -> Model Input Data -> New Input Data -> ML algorithm -> Prediction -> Accuracy -> Successful Model