AI Buzzwords: All you need to know! Part 1
Before a deep dive into the world of AI, let us take some time to look at a few jargons in AI. I have tried to explain it as simple as possible, such that it always resides in our conscious memory without the need to learn and relearn hopefully, without extensive training on God’s gifted intelligent network.
Artificial Intelligence: It is the ability of computers and machines to mimic human intelligence (‘learn and think’) for problem-solving and decision making. AI works with a large volume of data as inputs and smart algorithms to learn the patterns in the data in such a way that the system delivers accurate outputs. Machine learning, deep learning, neural networks, computer vision, cognitive computing, and natural language processing are the major sub-fields of AI.
Machine learning: A subset of AI that uses mathematical models and data in such a way that the system can learn and improve from examples and previous experiences without being explicitly programmed. ML models require minimal human intervention to identify the important set of features to understand the difference between the data inputs, i.e., require structured data to learn. The broad categories of ML are supervised learning (linear regression, support vector machine, and random forest), unsupervised learning (k-means clustering, principal component analysis, and Apriori algorithm), and reinforcement learning (Q-learning and deep Q-learning).
Deep learning: A subset of AI that makes use of neural networks to mimic the working of the human brain. DL algorithms use the ‘information processing patterns’ mechanism to process the large volume of raw unstructured data (image, text, etc.), automate feature extraction, identify patterns, and classify the input. A deep learning model can be visualized as a neural network architecture with more number of hidden layers. The broad categories of DL are supervised DL (recurrent neural network and convolutional neural network) and unsupervised DL (self-organizing maps and autoencoders).
1. Supervised learning: The labeled data, i.e., data gathered from experience is fed as an input to the learning model. As the input (X) and output (Y) are already known, the learning model learns and finds the best mapping function (f), such that it can provide predicted output (Y’) closer to the actual output (Y). A labeled dataset is divided into
(i) Train dataset: input for the training process, wherein the model learns the mapping function (f) based on the input (X) and output (Y), and
(ii) Test dataset: a set of unseen observations (without labels) that are fed to the model to make predictions based on the inferred mapping function (f).
The learning stops when the difference between the predicted output (Y’) and actual output (Y) goes beyond a certain threshold where the errors become negligible. Supervised learning can be categorized as classification and regression problems.
Classification: A classification problem identifies the input data as a member of a particular class and the output variable is a discrete value or category. For example, classification of emails as ‘SPAM’ and ‘NOT SPAM’.
Binary classification: A supervised machine learning problem that classifies the input data into two mutually exclusive groups. For example, classify an image as a ‘DOG’ or ‘CAT’.
Multi-class classification: A supervised machine learning problem that classifies the input data into three or more groups. For example, classify a car as ‘SEDAN’ or ‘SUV’ or ‘HATCHBACK’.
Multi-label classification: A supervised machine learning problem that classifies each input data into zero or more labels. For example, classifying movies (Jumanji) based on the genre (Adventure and Comedy) can have more than two possible labels.
Regression: The output variable of a regression problem is a continuous value commonly known as forecast or prediction which is based on the historical data. For example, predict the temperature of the next day or week based on the dependent variables like humidity, air velocity, solar energy, elevation, and proximity of water bodies.
Unsupervised learning: Unsupervised learning happens with the unlabeled data (input (X) is known; output (Y) and mapping function (f): unknown), where labeling is often difficult to obtain due to lack of knowledge or expensive. The learning model tries to find the similarity index among the different input (X) based on distance-based similarity measures and group them based on the similarity index, i.e., output (Y). Unsupervised learning can be broadly categorized into clustering, association and dimensionality reduction.
Clustering: It reveals the underlying patterns in the data. It helps to group the unlabeled data into similar clusters based on some features. For example, grouping books into different genres based on the summary.
Association: A rule-based ML technique that discovers association rules, i.e., finding interesting relationships hidden in large datasets. For example, market basket analysis — a customer who bought bread also bought milk, jam, and butter.
Dimensionality reduction: Detects correlation between the different dimensions of the dataset and helps to find the redundancy in the dataset and reduce it. In simpler terms, reduces the dimensionality of the dataset without loss of information. For example, principal component analysis and linear discriminant analysis.
Semi-supervised learning: Falls in between supervised and semi-supervised algorithms wherein the training dataset comprises smaller labeled samples and more unlabeled samples. Here, learning can be done in two ways
(i) Cluster the similar data using any clustering algorithm and label the unlabeled data using the existing labeled data.
(ii) Train the learning model with the labeled data (true-labeled dataset), the mapping function (Y) learnt from the labeled data is used to predict the output (Y) of the unlabeled data (pseudo-labeled dataset); Train a new model with the true-labeled and pseudo-labeled dataset. For example, document classification and content classification.
Reinforcement learning: A feedback-based ML technique where the learning model that acts as a virtual agent in an interactive environment selects the best action to maximize the reward (positive behavior) and minimize the penalty (negative behavior). The agent learns and explores the environment to perform the task in a better way through a ‘trial and error’ mechanism, feedback, and experience. The agent takes the next action and changes the state based on the feedback of the previous action. RL solves specific-problem with sequential decision-making and long-term goal. Few notable applications of RL are DeepMind, robotic navigation, industrial automation, and gaming.
Part 2 is to be published soon…
