Where does the knowledge is storage inside machine learning algorithms?
Different paradigms have distinct ways of learning.
Formally, we have the following definitions about Machine Learning (ML):
i. “Field of study that gives computers the ability to learn without being explicitly programmed.” — Arthur Samuel, 1959.
ii. “Well-posed Learning Problem: A computer program is said to learn from experience E with respect to some task T and some performance P, if its performance on T, as measured by P, improves with experience E.” — Tom Mitchell, 1998
Thus, we can conclude that ML has as objective the development of algorithms that aim to improve some tasks through previous experiences. These experiences are the available data, or in other words, the samples of the observed event. At this point, the idea of using samples of data (experience) as a base for the development of an ML algorithm is called inductive reasoning. From this observed data set, we expect that the algorithm can understand the patterns and generalize its conclusions to the entire population. It allows the computer to make decisions and learning without being explicitly programmed for a given task.
For example, instead we explicitly write a code like if(words_number<100){“spam”}; else{“not spam”} for the computer make emails classification, we give a large number of emails, and the computer should be able to, through these samples and some rules, say whether future emails are spam much more accurate then our first approach. Usually, the way to measure our models’ performance is by dividing the samples into two sets: a training set and a test set (actually, it’s possible to add another set called validation set, but we’ll omit it for now to turn the explanation easier). So, the sets are responsible, respectively, for training the model (estimate its parameters) and evaluating its performance (simulating future samples).
Types of learning
There are many ways to classify ML algorithms. Here, we’ll use one of them, the types of learning:
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
The following bubble chart shows some commons fields of applications for each paradigm.
Supervised Learning
In this category, the models have as input a set of labeled data. When the labels are discrete, we say that is a classification problem (for example, medical diagnosis, credit analysis, or our previous example on email classification), on the other hand, when labels are continuous, we say a regression problem (wind speed prediction, houses prices based in features). Here, the algorithms try to calculate an explicit function that maps the labeled data with the correspondent input. The supervised algorithms fit the model by error minimization or accuracy maximization. At the end of the training process, the knowledge is stored in the model itself, that is, in the first place, a graph of operators and coefficients or a set of rules.
The most used supervised algorithms are:
- Linear Regression
- Logistic Regression
- Artificial Neural Networks
- Support Vector Machines
- K-Nearest Neighbors
- Decision Trees
- Naive Bayes
Unsupervised Learning
Mainly applied for clustering, anomaly detection and association problems, the unsupervised learning algorithms don’t need labeled data. Here, the learning is inherent to the algorithm’s capability in distinguishing the input information by means of distances or pattern recognition. Therefore, the knowledge here is somehow in the data itself, but also as the models’ hyperparameters.
The most used supervised algorithms are:
- K-Means
- Gaussian Mixture
- Hidden Markov Models
- Principal Component Analysis
- Isolation Forest
- DBScan
Reinforcement Learning
This paradigm is quite different from the two above. Basically, here the agent (system) is subjected to situations and gets rewards or penalties according to the actions it takes. The most illustrative example is that techniques of reinforcement learning are used to make robots learn how to walk. RL algorithms mixed with deep learning have shown awesome results as well. Here, the knowledge is dynamic, constantly modified.
Thank you for reading.
Special thanks to Gabriel Teotonio (gabrielteotonio.com) and Jair Sales (jps4@cin.ufpe.br).
