Machine Learning and Deep Learning

Machine Learning Overview

Machine learning (ML) is a branch of artificial intelligence where computers learn from data and adapt the computational models to enhance performance. It is a method of analysis that allows computers to reveal information within data.

• The “learning” is not the type of learning that you and I do.
• It is a systematic approach to finding an appropriate data transformation from inputs to output.

Why ML?

• Computers can sort through data faster than humans can.
• Computers can identify patterns quickly and use these patterns for predictions or classifications.
• Machine learning can handle noisy data – it doesn’t find a perfect answer, but rather a “really good” answer.

Applications of ML

• Regression techniques
  • Determines a mathematical model for the relationship among features or attributes so that an outcome can be predicted.
  • Results can be any value within a possible range (e.g., what will the average Earth temperature be in 2050?)
• Classification problem
  • Identifies a combination of attributes that best fits a class or category so that an object can be classified.
  • Results can be from a list of known possibilities (e.g., is the tumor benign or malignant?)

Note: Examples included in this tutorial are all of classification type problems.

Types of ML

• Supervised Learning:
  • A data set exists where the samples can be categorized into two or more classifications.
  • The computer uses the data set to learn how to predict the classification of an unknown sample.
  • Examples include Decision Trees and Deep Learning
• Unsupervised Learning:
  • The collected data has no known classification or pattern.
  • The computer must identify the groups or hidden structures within the data.
  • Examples include Dendograms, K-means clustering, Self-organizing Maps
• Reinforcement Learning:
  • Computer learns from positive or negative feedback
  • Example includes Swarm intelligence

Note: Examples included in this tutorial are all instances of supervised learning.

Data for ML

• For many Machine Learning algorithms, the data is expected to be in a table format, where:
  • each row represents an object, and
  • each column has the measurements for a specific attribute or feature of the object
• For supervised learning, the classifications of the objects must be known.
  • The data with known classifications are divided into a training set and a testing set.
  • The data is then used to develop a model.
    • The training data are submitted to an algorithm that will fit a model to the data.
    • The test data are submitted to the model to produce predicted classifications and determine the accuracy of the model.
  • Finally, the model can be used to predict classifications for “unknown” data.

ML Algorithm

The algorithm determines the best mathematical model for the code. However, you still need to provide a “framework” for the algorithm. The framework provides the algorithm with tools for performing the learning.

Deep Learning vs Machine Learning

• Deep Learning is a subset of Machine Learning that differentiates itself from ML algorithms based on the methods (neural networks) it uses to solve problems.
• Any deep learning algorithm would reiterate and perform a task repeatedly, improving a bit every time, in order to improve the outcome.
• A deep learning program builds the feature set by itself without supervision and domain expertise.
  • Unsupervised learning is not only faster, but it is usually more accurate.
    • Unsupervised learning needs access to immense amounts of data and compute power and takes much longer to train, but it is much faster to run tests.
• ML algorithms have superior interpretability and are favorable for small amounts of data.
• ML works only with sets of structured and semi-structured data, while deep learning works with both structured and unstructured data

Limitations and Challenges

• DL models learn through observations. Outcome is not ”generalizable” if data was small or if its scope was limited.
• If a model trains on data that contains biases, the model will reproduce those biases in its predictions.
• If the learning rate is too high, then the model will converge too quickly, producing a less-than-optimal solution. If the rate is too low, then the process may get stuck, and it will be even harder to reach a solution.
• There are a large amount of data and parameters.