A Basic Introduction to Machine Learning Algorithms
Machine learning is a field in computer science, which gives computers the ability to learn without explicitly being programmed to do so, and has widespread applications across several fields, such as diagnosis of diseases, optical character recognition, computer vision and email filtering. Many machine learning techniques are in-fact being used on a day-to-day basis for technologies like smart advertisements, friend recommendations as well as suggested search results, and we interact with them all the time.
Machine learning emerged from similar artificial intelligence fields like pattern recognition, and computational learning, and it heavily relies on statistics and mathematical optimizations. Many classification problems including anomaly-detection problems, can be solved using different machine learning algorithms put together. These algorithms form the backbone of the many technologies for use of artificially intelligent systems.
Machine Learning algorithms are broadly divided into three categories that take on different approaches to help computers learn how to solve tasks on their own. These are: Supervised Learning, Unsupervised Learning and Reinforcement Learning.
This approach to machine learning involves giving a set of inputs and outputs to a computer, and training it based on a dataset of ‘training examples’ we gave. Then it can be further used to predict new values when new inputs are provided to it.
Typically the dataset consists of pairs that consist of an input vector, and a corresponding output. The task of the machine is generally to determine a function such that on taking an argument of a new input vector, it should be able to output a value which has a high rate of accuracy of being the correct output.
This criterion leads to the necessity what is known as a bias-variance trade-off. A high-bias, low-variance machine learning model will overfit the data to even accommodate the outliers and the anomalies, and will not be able to adjust to the new inputs that are provided to it. Meanwhile, a high-variance, low-bias model will underfit the data, and though it will generalize itself well to different data, its accuracy will be low and not fit any data particularly well.
Thus most algorithms provide a parameter that adjusts this bias-variance tradeoff so that the model generalizes well to other data outside the training data with higher accuracy.
This is used to infer results from unclassified data, where the data is provided in the form of unlabelled points without any output values given. Then, this data is grouped into separate categories based on different criterion such as the proximity of the data points, etc.
A notable difference between unsupervised and supervised machine learning algorithms is the absence of an accuracy feature to check the viability of the unsupervised model. This is useful when we need to find hidden structures in unlabelled data, and can draw inferences from it.
Data can in general, be grouped into clusters which are similar in some respect, and these hence create a new structure for finding patterns within the data. This is especially useful in data where such information or structure would have usually gone unnoticed.
This is an area in machine learning which is inspired by behaviourist psychology, in which the program takes actions depending on which it is either rewarded or penalized. It utilises dynamic programming techniques and is often studied in many other disciplines, such as game theory, statistics as well as information theory. This approach helps the model make the ideal choices given a particular situation.
There exist many different solutions to such problems, however, common ones include those which will provide greater reward in the long run, as opposed to immediately after. Its applications are manifold, ranging from controlling robotic arms to programming robots to avoid obstacles by creating penalties for every obstacle hit. Logical games, such as chess, can also be programmed to be played by such models.
Machine Learning is a wide and varied field with plenty of potential uses that can improve not only our technology, but also our lives. In a world filled with pessimistic warnings about the potential impact of AI on our lives, it is important to remember that such models have several advantages, too, and have led to groundbreaking research in many fields, ranging to diagnosis of diseases such as cancer simply by clicking a photo to cybersecurity and financial analysis. As in any such debate, understanding how they work is often one step closer to arriving at a compromise or a solution.
Perhaps we can avoid Judgement Day, after all.