Elements and Challenges in Machine Learning

• Machine learning uses datasets divided into training, validation and testing, training the system to optimize learning and system parameters, then testing the trained system.

Main elements of Machine Learning

• Training examples are used to provide feedback to the system, this feedback can be direct or indirect from the working environment.
• Training examples can be supervised or unsupervised.
• Training examples should accurately represent future test examples.
• The target function to be learned includes F: X -> {0,1} and F: X -> A set of class labels.
• Also F: X -> R+ (i.e., a domain of positive real values) can be used.
• Possible representations of target functions include polynomial functions, a set of rules, decision trees, or artificial neural networks.
• ML algorithms can learn to approximate the target function.
• Some ML Algorithms include: Regression-based, rule induction, decision tree learning (e.g., ID3 or C4.5), back-propagation.

Challenges in Machine Learning

• Finding which learning algorithms can learn a given target function.
• Finding the conditions under which a selected learning algorithm converges towards the target function.
• Determining which learning algorithm performs the best for a given representation and problem.
• Figuring out how many training examples are sufficient for training.
• Evaluating how the size of the training set affects learning accuracy.
• Determining how errors and missing values affect accuracy.
• Finding the best way of use order of training examples.
• Evaluating how the order of training examples affects the complexity of the ML problem.
• Evaluating how specific knowledge of the application problem contributes to learning.
• Selecting which target function the system should learn.
• Considering the representation capability (e.g., linear vs. non-linear) against the complexity of the algorithm.
• Generalization depends on limits, i.e. underfitting vs. overfitting.
• Self-adaptation can improve the system's capability of representation and target function learning.
• Retraining depends on whether the trained model performs poorly on new examples after performing well initially.
• Retraining helps the model adapt to newly coming examples.

Generalization Capability

• The model's ability to maintain high accuracy for future unseen data.
• Test examples should not be used during model selection/training.
• The validation set (extracted from the original training set) is used as unseen data.
• It is assumed that the data characteristics are similar between test and validation sets.
• Under-fitting occurs when low accuracy happens across training, validation, and test sets resulting in "high bias".
• Over-fitting means high accuracy on training but low accuracy on validation/test sets, leading to "high variance."

Over-fit Learning

• A learned target function over-fits a training set if another target function exists with:
  • Lower accuracy on the training set.
  • Higher accuracy on the entire dataset, also including examples evaluated after training.
• The target function h is over-fit to D_train if another target function h' exists, where Err_D_train(h) < Err_D_train(h′), and Err_D(h) > Err_D(h′).
• Occurs due to errors or noise in the training set.
• Lack of representative training examples.
• Too high accuracy (~100%) on the training set, resulting in a function ideal for training examples, but unsuitable for future examples.
• Goal is to achieve high levels of machine learning accuracy in predictions for future examples, not just for the training set.
• Occam's razor is used to select the simplest suitable target function to achieve a better generalization, which is easier to explain, and at lower computing cost.
• Decreasing accuracy on the test set, even with the training set’s accuracy increasing, exemplifies the over-fit learning.

Frameworks and Tools for Machine Learning

• TensorFlow
  • OS: Linux, Mac OS, Windows, Android
  • Programming Language: Python, C++, Java
• Caffe
  • OS: Linux, Mac OS, Windows
  • Programming Language: Python, Matlab
• Caffe2, PyTorch
  • OS: Linux, Mac OS, Windows, iOS, Android, Raspbian
  • Programming Language: C++, Python
  • In March 2018, Caffe2 and PyTorch merged into a single platform
• Keras
  • OS: Linux, Mac OS, Windows
  • Programming Language: Python
• Theano
  • OS: Linux, Mac OS, Windows
  • Programming Language: Python
• CNTK
  • OS: Windows, Linux
  • Programming Language: Python, C++, C#
• Deeplearning4j
  • OS: Linux, Mac OS, Windows, Android
  • Programming Language: Java, Scala, Clojure, Python
• Apache Mahout
  • OS: Any OSs with JVM installed
  • Programming Language: Java, Scala
• MLlib of Apache Spark
  • OS: Any OSs with JVM installed
  • Programming Language: Java, Python, Scala, R
• Weka
  • OS: Any OSs with JVM installed
  • Programming Language: Java

Machine learning involves training systems with datasets divided into training, validation, and testing sets, using training examples. Target functions, represented in various forms, are approximated using algorithms such as regression and decision trees. Challenges include accurately representing future test examples.