Machine Learning (ML) is the second of 21 technologies or ideas that will shape the business landscape in the decade of 2020s. The first one is Artificial Intelligence. Let us look at Machine Learning in detail.
What is machine learning?
Machine Learning (ML) is a subset (subfield) of Artificial Intelligence where machines learn from the data presented without explicitly programmed. The objective of ML is to elicit structure from a group of apparently unstructured data and fit that data into models with predictive and problem solving capabilities. The data used to generate the models is called 'Training Data' and the data input for prediction purposes is 'Validation Data'.
For example, training data could be a set of different fruits, bananas, apples and oranges. Based on the different features of the fruits, the algorithm develops a model that learns to identify (classify) different groups of fruits. Once this model is executed on a computer, the machine has learned to identify the fruit. When a new fruit is presented (validation data), the machine correctly identifies it as 'Banana'.
Definition of machine learning.
The term 'Machine Learning' was coined in 1959 by Arthur Samuel. He defined it as "it (ML) gives the machines the ability to learn without being explicitly programmed'
In 1997, Tom Mitchell defined Machine Learning as "A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E."
A simple definition of machine learning is 'Using Data to Answer Questions'. Here 'using data' is part of training and 'Answer Questions' is part of prediction or inference.
Why machine learning now?
To answer the question, let us look at the way human beings learn. We see disparate data and then make inferences. Even before it learns the word, the child sees a goat in his house as well as in his friend's house and learns that these are multiple instances of the 'same' animal. Over a period of time the child may learn more features of the animal, 'Generally White', 'Is domesticated', 'Gives milk' etc. Finally he may learn that it is called 'Goat'.
Or take another example of building a church. The masons who built the first church did not know what it was. They were building 'something', Over generations masons learned that their building had certain features like 'Has bells', 'Has a Cross' etc. Finally they learned that it is called a 'Church'.
Human beings look at data and identify patterns and features. This is how we learn.
However, our brains has a limitation in processing data. With the data explosion of the recent years, we have taken recourse to 'Heuristics' to make our decisions. But we cannot allow this vast amount of data go unutilized. We needed machines and processes that can take this huge amount of data and identify the hidden patterns out of this data.
Hence we can say that the availability of huge amount of data and high end computing resources together ushered in the age of Machine Learning. (Another component is the development of advanced algorithms).
Two types of learning
Human beings learn in two ways. One is called 'Rote Learning' or Memorization or Declarative Learning. Here we memorize facts and data and retrieve it when asked. This is a highly inefficient form of learning.
The other learning approach is 'Generalization' or Inference or Imperative Learning. Here we learn the 'first principles' and learn new facts from old facts. The two aspects of this type of learning is its predictive ability and the assumption of extrapolation of past to future.
Early ML used Rote Learning. The objective was to feed it with vast amount of data so that it can come back with answers when asked a question. This approach is also known as 'Symbolic Learning'. The recent advances in ML focus on Generalization as a learning approach. We feed large amount of data to the machine and let it identify the patterns for example.
Machine learning approaches
There are three main approaches to ML. They are.
1. Supervised Learning: In this method, each instance of a training dataset is composed of different types of input attributes and a corresponding expected output attribute. The input can be any type of date including pixels, images, numbers etc. This approach uses Labeled Data and the objective is to classify and group the data.
It is called 'Supervised' due to the presence of a 'Supervisor / Trainer' which evaluates the output provided by the model and make corrections. This iterative approach continues till model comes up with an acceptable level of performance (expected results over multiple iterations).
The trained model is then used to predict the output based on a new untested input (Validation Dataset)
Business use cases of this include:
2. Unsupervised Learning. In this case there is only input data without any corresponding output variable. The goal is to model the underlying data structure to learn more about the data. This method uses unlabelled data and there is no supervisor to guide the algorithm. The algorithm is left to itself to find interrelationships and patterns within the data. This mode of learning is very valuable since we have more of unlabelled data than labelled ones.
Business use cases include:
3. Reinforcement Learning: In this method the learning is through continuous interaction between the learning agent and the environment. The learning happens through trial and error. The environment provides regular feedback through rewards and penalties to guide the learning agent in the right direction. The reward / penalty acts as a reinforcement signal. The LA improves its knowledge about environment through the rewards mechanism.
Learning agent uses two mechanisms.
- Exploration: When LA acts on trial and error basis
- Exploitation: When LA acts on the basis of knowledge gained from environment.
Business use cases include:
Biases
Machine learning suffers from a number of data biases
References.
No comments:
Post a Comment