Machine Learning Landscape

Machine Learning Landscape

• Machine learning is the field of study that gives computers the ability to learn without being explicitly programmed (Arthur Samuel, 1959).

• A computer program learns from experience (E) with respect to a task (T) and a performance measure (P). Its performance on T as measured by P improves with experience E (Tom Mitchell, 1997).

• The diagram shows an example of a spam detector which takes in mail and then identifies spam or inbox.

• A traditional approach includes studying the problem, writing rules, evaluating, and analyzing errors. Repeatedly iterating over the steps until the desired result is met.

• A machine learning approach uses data to train the model which is validated through repeated testing and evaluation. The model is updated and trained once errors are analyzed.

• Machine learning can adapt to changing data; models are updated with new data as it becomes available.

Types of Machine Learning Systems

• Machine learning can be categorized by whether they use human supervision. This includes supervised, unsupervised, semi-supervised, and reinforcement learning
• Batch versus online learning – whether the model can learn incrementally on the fly.
• Instance-based versus model-based learning – whether the model compares new data to known data points or detects patterns in training data and builds a predictive model.

Supervised Learning

• Classification: A labeled training set for spam classification is an example. A training set is used to classify new data into the correct category (e.g. spam or inbox)

• Regression: Predict a value given an input feature. The example illustrates a problem in which a value is predicted based on an input feature.

• Algorithms: Some key algorithms include k-Nearest Neighbors, Linear Regression, Logistic Regression, Support Vector Machines (SVMs), Decision Trees, and Random Forests.

Unsupervised Learning

• Clustering: Identifies groups of similar data points in a training set in a diagram.

• Visualisation: Visualization algorithms, like those using t-SNE, are used to represent high-dimensional data as a clear diagram and provide valuable insights. This is illustrated with an example that plots different types of animals on a multi-dimensional chart that shows the different groupings.

• Anomaly detection: Identifying data points that deviate significantly from the norm in a training set in a two-axis chart.

• Association rule learning: Finds relationships and patterns between different data points in a training set in a diagram.

Semi-supervised Learning

• Combines labeled and unlabeled data for training. Diagram shows an example where there is a mixture of labeled classes and unlabeled data.

Self-supervised Learning

• A form of unsupervised learning where the dataset is used to create its own labels. The example illustration contrasts the unlabeled and labeled datasets.

Reinforcement Learning

• This type of learning involves an agent learning by interacting with an environment and receiving rewards for its actions in the example a robot is depicted interacting with an environment and receiving rewards while making decisions.

Batch and Online Learning

• Batch learning: The model learns from the entire dataset at one time before making predictions. Diagram shows the model training and updating with new data.

• Online learning: The model learns incrementally from new data as it arrives. This allows the model to adapt to the new data. The model is trained iteratively with new data added to the existing dataset until a satisfactory model is developed.

Learning Rate

• High learning rate: Rapid adaptation but quickly forgetting.

• Low learning rate: More inertia but less sensitive to noise. Adaptation to changing data, inertia or momentum, and sensitivity to noise when learning with machine learning.

Instance-based vs. Model-based Learning

• Instance-based: Considers similar instances in the training data set to classify new data.

• Model-based: Creates a model of the data's structure to predict new instances. This is visualized using a diagram to distinguish between the different approach.

Overfitting the Training Data

• Problem: Occurs when a model learns the training data's details, including noise, leading to poor generalization to new data.
• Possible solutions: Simplify the model, gather more data, reduce noise in the training data (e.g., fixing data errors and removing outliers), regulate model complexity. This is illustrated with a graph.

Underfitting the Training Data

• Problem: Occurs when the model is too simple to learn the underlying structure of the data.
• Possible solutions: Select a more complex model, feed better features to the learning algorithm, decrease model constraints.

Testing and Validating

• Generalization error: Measures the accuracy of a model on unseen data.
• Training set: Used to train the model.
• Validation set: Used to tune and select the best hyperparameter.
• Testing set: Used to assess the model’s performance on unseen data.
• Cross-validation: A technique to evaluate a model's performance using different subsets of the data for training and validation.

Hyperparameter Tuning and Model Selection

• Validation set: Used during model selection to evaluate the performance. Diagram presents a model of the data flows and analysis that are involved.

No Free Lunch (NFL) Theorem

• No single learning algorithm is universally better than others for all possible learning problems.

Homework

• The user is presented with homework assignments or tasks relating to the materials presented with exercises on the concept provided by the presentation.

Explore the fundamental principles and systems in machine learning. This quiz covers topics such as the operational definitions, approaches, and the iterative nature of machine learning models. Test your understanding of how machines learn from data and adapt to new information.