What Are Different Models In Machine Learning?
Artificial Intelligence & Machine Learning
Over 2.5 quintillion bytes of data are created daily, and understanding how to analyze it and use it to draw informed conclusions is an essential facet of work in any field. In the customer service industry, reviews are parsed and studied to determine overall customer satisfaction trends, and in medicine, analysis of historical patient data can provide clues as to the illnesses afflicting new patients. Credit card companies sift through a huge volume of daily transactions to identify a few fraudulent ones, and in finance, the study of a multitude of market indicators can yield insights into future market trends and stock performance.
All these tasks and many more are accomplished using machine learning, a type of artificial intelligence that relies on using algorithms to predict values or classify data with increasing degrees of accuracy by “training” itself using large volumes of input data. However, “machine learning” is an umbrella term often used to describe many different problem-solving methods, each of which can be applied to solving specific types of tasks. What are the main types of machine learning models, and how is each used to solve problems in modern society?
Unsupervised learning
Unsupervised learning is the process through which unlabeled datasets are analyzed and subdivided without human input. Such algorithms identify similarities between data points and uses those to sort data, without the need for identifying information or prior inputs. As such, it is well suited to exploring new datasets comprised of datapoints with unknown relationships.
The two main types of unsupervised learning are clustering and dimensionality reduction.
Clustering
Clustering is a process through which similar data points are organized into subgroups or ‘clusters”. Hence, points within the same cluster share a common characteristic or have similar properties, which also differentiates them from other clusters. Once the data is grouped into clusters, novel conclusions can be drawn and serve as the backbone for further analysis. For example, suppose we had Netflix user data describing login frequency, length of time watched per day, total watch time and more. Clustering could separate these users into categories (active/inactive/active on weekends/active in the evening/etc…), allowing Netflix to target larger groups more accurately with advertising.
Dimensionality reduction
In supervised learning, input data characterized by one or more variables is used to “train” a model to identify trends in data. However, using too many input variables can contribute to higher error rates and decreased overall performance. Dimensionality reduction addresses this issue by identifying groups of variables that are highly correlated amongst themselves and combining them to reduce the number of input variables, effectively reducing the dimensions of the matrix representing said variables. This in turn increases model accuracy, reduces the storage requirements of the data and makes it easier to visualize.
Supervised learning
Supervised learning methods rely on the use of labelled input data, which serves to ‘train’ a model to form associations between different types of input data and specific outputs. Then, when presented with new “testing” data that it hasn’t previously evaluated, the model can compare this new data to its previous inputs and deduce the most likely output type to a high degree of accuracy. The two main types of supervised learning are classification and regression problems.
Classification
In classification problems, input data is provided to a model along with accompanying labels, enabling the model to create and fine-tune a function linking the presence or intensity of certain input parameters to the final output of the model. Then, when provided with new input data, it can classify it according to its pre-defined possible outputs. A particularly popular classification problem is image recognition. For example, one could use labelled pictures of dogs and cats as inputs to such a machine learning model. These pictures are then decomposed into individual pixels, and certain recurring pixel patterns and groupings in each category (dog/cat) are identified by the model. When a new, unlabeled picture is provided, the model searches for the aforementioned patterns and labels the picture as either ‘dog” or ‘cat” based on the identified features.
Regression
Whereas classification problems involve matching certain predefined outputs to each set of input data, regression problems are more complex in that the predicted outputs are often continuous numerical values. Hence, a regression model is used to establish a relationship between several individual input variables and a single output variable, but does not return only predefined outputs (as is the case with classification problems).
Rather, it attempts to fit a function to the data to find the relation between inputs and outputs, then uses this function to deduce the output for a novel set of inputs. For example, suppose the daily temperature and number of ice cream cones sold at an ice cream stand are studied. From studying historical data, we find that for every additional degree of temperature, the ice cream stand sells approximately 10 additional cones. Hence, by identifying the trend in the data, we can deduce that, should tomorrow’s temperature be 5 degrees warmer, the stand will sell approximately 50 more cones. Machine learning regression generalizes such numerical prediction problems to larger scales, as models can include many different input variables studied over years or even decades. More complex regression models are used to approximately predict carbon emissions, stock prices or energy consumption as a function of time.
Deep Learning
Deep learning is not a category of machine learning problems, but rather an approach to solving complex problems that perfectly illustrates the depth of this field. Moreover, Deep learning uses algorithms called artificial neural networks to imitate the human brain using layers of nodes to represent neurons. These interconnected nodes are adept at sorting through large volumes of unstructured data and determining which features of an input dataset are most important when predicting output values. The connections between nodes are characterized by weights, which determine how closely linked an input and output are. Once the model outputs a prediction, it adjusts the weights of each connection depending on the prediction’s accuracy and continues to do so while training data is input.
Deep learning models can possess hundreds of layers of nodes, and require massive amounts of training data and training time relative to other machine learning methods. However, they are capable of feats far beyond the reach of other technologies. For example, DeepMind’s AlphaGo program used deep learning to defeat the Go world champion in a series of 5 games in 2016, a feat previously considered impossible due to the game’s sheer volume of possible board states (approximately 10^170). Another example of deep learning’s versatility is OpenAI’s DALL-E software, which generates extremely realistic images and visual art using simple text inputs.
This picture was generated using OpenAI’s DALL-E tool using the text input
“Neural network, visual art”
Machine learning is an incredibly vast field with increasingly complex applications in modern technologies. ML algorithms can sort and organize new data, simplify complex datasets by reducing the number of inputs, classify inputs into groups based on chosen parameters, predict numerical values, or even generate images, outperform human experts and identify human emotions in text. The applications of machine learning are near limitless, and as computing power continues to increase in scale, every-more complex algorithms can be developed and run to analyze the growing amount of data generated by society. Improving our understanding of the data that surrounds us is key to problem-solving across many disciplines, and machine learning is the ideal tool to do so.
Written by Sebastien Brown
What Are Different Models In Machine Learning?
