Table of Contents

• A brief history of machine learning
• Main challenges
• Why predictions fail?
• Machine learning project checklist
• Resources
  • Presentations
  • Tutorials
  • Online content
  • Papers
  • Books
  • Datasets
  • Others
• Glossary

A brief history of machine learning

Machine Learning is the study of computer algorithms that improve automatically through experience. --Tom Mitchell

Adapted from the section 1.2 of the book "Deep Learning with R".

• Probabilistic modelling was one of the earliest forms of machine learning, where the principles of statistics were applied to data analysis.
• The Naive Bayes algorithm is one of the best known methods for carrying out probabilistic modelling.
• Logistic regression is an algorithm with roots that date back to the early 19th century.
• The core ideas of neural networks were investigated back in the 1950s.
• In the 1970s, the backpropagation algorithm was originally introduced.
• Kernel methods, especially Support Vector Machines (SVMs), become popular in the 90s.
• Decision trees date back in the 1960s.
• The first algorithm for random decision trees was created in 1995 and an extension of the algorithm, called Random Forests, was created in 2001.
• Gradient boosting is a machine-learning technique based on ensembling weak prediction models, generally decision trees, which originated in 1997.
• In 2012, the SuperVision team led by Alex Krizhevsky and advised by Geoffrey Hinton was able to achieve a top-five accuracy of 83.6% on the ImageNet challenge.

Main challenges

• Most machine learning algorithms require a lot of data to work properly; if your sample is too small, sampling noise will have a larger effect.
• In addition, training data needs to representative, i.e. equally sampling from different labels or range of values, for a model to generalise. Non-representative data can result from a flawed sampling method (but sometimes it is just difficult to collect the necessary data).
• Poor quality data will make it more difficult to detect underlying trends and a lot of time is required to "clean up" the data, e.g. dealing with missing values and removing or fixing outliers.
• Models are only capable of learning if there are enough relevant features. A critical part of machine learning is feature engineering, which typically involves feature selection (selecting the most useful features) and feature extraction (combining existing features to produce more useful ones).
• Avoiding overfitting, which is when the model performs well on the training data but does not generalise and this usually occurs when the model is too complex. Constraining a model to make it simpler and reducing the risk of overfitting is called regularisation. The amount of regularisation to apply during learning can be controlled by a hyperparameter, which is a parameter that is not part of the model.
• Avoiding underfitting, which is when the model is too simple to learn the underlying structure of the data.

Why predictions fail?

Reasons machine learning models fail to make correct predictions despite having enough data:

1. Inadequate pre-processing of data
2. Inadequate model validation (the procedure where a trained model is assessed with testing data)
3. Inappropriate model was used
4. Unjustified extrapolation (making predictions on new data that is characteristically different from the training data)
5. Over-fitting the model on existing data

Machine learning project checklist

A useful checklist for machine learning projects from Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow 3rd edition.

1. Frame the problem and look at the big picture.
2. Get the data.
3. Explore the data to gain insights.
4. Prepare the data to better expose the underlying data patterns to machine learning algorithms.
5. Explore many different models and shortlist the best ones.
6. Fine-tune your models and combine them into a great solution.
7. Present your solution.
8. Launch, monitor, and maintain your system.

Adapt this checklist to your project needs!

Resources

A list of useful resources for learning about machine learning with an emphasis on biological applications.

Presentations

• Some Things Every Biologist Should Know About Machine Learning by Robert Gentleman

Tutorials

• How to Perform a Logistic Regression in R
• A gentle introduction to decision trees using R
• A gentle introduction to random forests using R
• Random Forest Regression and Classification in R and Python
• Fitting a Neural Network in R using the neuralnet package

Online content

• A Tour of The Top 10 Algorithms for Machine Learning Newbies
• Comparing supervised learning algorithms
• How to get better at data science

Papers

• What is Bayesian statistics by Sean Eddy
• What is a support vector machine? by William Noble
• What is a hidden Markov model? by Sean Eddy
• What are artificial neural networks by Anders Krogh
• Deep learning for computational biology by Angermueller et al.
• Machine learning applications in genetics and genomics by Maxwell Libbrecht and William Noble
• Conditional variable importance for random forests by Strobl et al.
• Ten quick tips for machine learning in computational biology by Davide Chicco
• Improving neural networks by preventing co-adaptation of feature detectors by Hinton et al.

Books

• The Elements of Statistical Learning by Trevor Hastie, Robert Tibshirani, and Jerome Friedman
• An Introduction to Statistical Learning by Gareth James, Daniela Witten, Trevor Hastie, and Robert Tibshirani
• Introduction to Machine Learning by Ethem Alpaydin
• Deep Learning with R by François Chollet with J. J. Allaire

Datasets

• Medical Data for Machine Learning by Andrew Beam
• UCI Machine Learning Repository
• Kaggle Datasets
• A collection of microarray data by John Ramey

Others

• Tom Mitchell's home page
• My notes on Random Forests
• No Free Lunch Theorems

Glossary

• Algorithm: A step-by-step procedure or formula for solving a problem or achieving a particular goal in machine learning, such as predicting outcomes based on input data.

• Backward Pass: The backward pass, also known as backpropagation, is the process of computing gradients of the loss function with respect to the parameters (weights and biases) of the neural network.

• Cross-Validation: A technique used to assess the performance of a machine learning model by splitting the dataset into multiple subsets (folds) and evaluating the model on each fold.

• Deep Learning: A subset of machine learning that uses neural networks with multiple layers (deep architectures) to learn hierarchical representations of data.

• Dropout layer: A dropout layer is a regularisation technique used in deep learning to prevent overfitting and improve the generalisation performance of neural networks. In a dropout layer, a random subset of neurons in the network is temporarily "dropped out" or ignored during training. This means that their contributions to the forward pass (computation of activations) and backward pass (computation of gradients) are temporarily removed. Dropout is typically applied after the activation function in each hidden layer.

• Ensemble Learning: A technique that combines multiple machine learning models to improve predictive performance. Examples include bagging, boosting, and stacking.

• Feature: A measurable property or characteristic (variable) of the data that is used as input for a machine learning model. Features are typically represented as columns in a dataset.

• Feature Engineering: The process of selecting, transforming, and creating new features from raw data to improve the performance of a machine learning model.

• Forward Pass: In a neural network, the forward pass occurs when input data is passed through the network, layer by layer, to compute the predicted output. Each layer in the network performs two main computations: a linear transformation followed by a non-linear activation function.

• Gradient Descent: An optimisation algorithm used to minimise the loss function by iteratively adjusting the parameters of the model in the direction of the steepest descent of the gradient.

• Hyperparameter: A parameter whose value is set before the training process begins. Examples include learning rate, regularisation strength, and the number of hidden layers in a neural network.

• Label: The output or target variable in supervised learning. It represents the value that the model aims to predict based on input features.

• Learning rate: The learning rate is a hyperparameter that controls the size of the steps taken during the optimisation process in deep learning. It determines how much the parameters of the neural network are adjusted in the direction opposite to the gradient of the loss function with respect to those parameters. In other words, it governs the speed at which the model learns during training.

• Loss Function: A function that quantifies the difference between the predicted outputs of a model and the true labels in supervised learning. It is used to measure the model's performance during training.

• Model: A mathematical representation of a real-world process or system used by machine learning algorithms to make predictions or decisions.

• Neural Network: A computational model inspired by the structure and function of the human brain. It consists of interconnected nodes (neurons) organised in layers.

• Optimiser: In deep learning, an optimiser is an algorithm used to update the parameters (weights and biases) of a neural network during the training process in order to minimise the loss function. The goal of optimisation is to find the set of parameter values that result in the best performance of the model on the training data.

• Overfitting: A phenomenon where a machine learning model is over optimised on the training data, capturing noise and irrelevant patterns, instead of generalising from it and leads to poor performance on the testing data.

• Regularisation: A set of techniques used to prevent overfitting and improve the generalisation performance of a model. Regularisation techniques introduce constraints on the model parameters during training, encouraging simpler and more robust models that generalise better to new data. These techniques help strike a balance between fitting the training data well and avoiding overfitting. Some common regularisation techniques include: L1 Regularisation (Lasso), L2 Regularisation (Ridge), and Dropout.

• Reinforcement Learning: A type of machine learning where an agent learns to make decisions by interacting with an environment. The agent receives feedback in the form of rewards or penalties based on its actions.

• RMSprop: RMSprop (Root Mean Square Propagation) is an optimisation algorithm that divides the learning rate for each parameter by a running average of the magnitudes of recent gradients for that parameter. It helps to stabilise the learning process and adaptively adjust the learning rates for different parameters.

• Semi-Supervised Learning: A combination of supervised and unsupervised learning where the model is trained on a dataset that contains both labeled and unlabeled data.

• Supervised Learning: A type of machine learning where the model is trained on labeled data, i.e., data with known inputs and outputs. The goal is to learn a mapping from input features to output labels.

• Testing Data: The dataset used to evaluate the performance of a trained machine learning model. It contains data that the model did not used during training.

• Training Data: The dataset used to train a machine learning model. It consists of input-output pairs used by the algorithm to learn patterns and relationships in the data.

• Underfitting: A phenomenon where a machine learning model is too simple to capture the underlying structure of the data. It performs poorly both on the training data and testing data.

• Unsupervised Learning: A type of machine learning where the model is trained on unlabeled data. The algorithm learns patterns and structures in the data without explicit supervision.