Data Science Terminology - AI / ML / DL
What is Artificial Intelligence (AI), Machine Learning (ML) and Deep Learning (DL) ?
It all started in 1950. Alan Turing, one of the greatest mathematicians of the history, suggested the Turing Test focusing on the following question: Will a computer ever be able to sufficiently imitate a human to the point where a suspicious judge cannot tell the difference between human and machine? This question was the driving force which inspired researchers to explore whether it is achievable to simulate intelligent behaviours in computers. Curious minds assembled at the Dartmouth Conference in 1956 and coined the term Artificial Intelligence, accepting it as a field of study.
In the following decades, despite drought periods in terms of advancements in AI, the focus of the field has started to shift from rule based expert systems to machine learning and neural network models. Especially with the increase of personal computer possession and popular applications like the PageRank algorithm introduced by Google accelerated the evolution of AI. Today, from self-driving cars to sentiment analysis, it occupies a great area, while continuing to create new business spaces.
As the relationship between AI, ML and DL has been clarified, further investigation of these terms from the perspective of organizations may shed light on gray areas. Artificial Intelligence presents great opportunities for the journey of becoming an insight driven organization. Particularly the progress in Machine Learning and Deep Learning have reformed the conventional practices of businesses in fundamental ways. The adoption and application of these revolutionary approaches by leading organizations have amplified the race for exploring and experimenting in these fields.
Artificial Intelligence
Artificial Intelligence means the simulation of intelligent behaviours with computer algorithms. Chat-bots on a website, for example, may respond to the questions like a human being, which in turn can be the best candidate to pass the Turing Test. However, the applications are not limited to the chat-bots. The use of AI is profoundly altering how businesses operate.
PwC’s AI Predictions survey in 2021 [1] presents that 52% of companies have prompted their AI adoption plans due to Covid-19 and they expect to realize the true impacts of these projects in the upcoming years. The survey also records that 86% of the participants claims that AI will be a “mainstream technology” at their organization in 2021.
In a similar aspect, IBM’s Global AI Adoption Index 2021 study [2] suggests that 31% of companies have already deployed at least one AI project and 43% of businesses are actively exploring AI. It means almost every 3 out of 4 organizations are using or experimenting with AI projects. The report also indicates that the shortage of AI skills and expertise is the greatest source of concern at smaller organizations while increasing data complexity and data silos are the most significant impediments to AI adoption among larger companies.
There are also other reports and articles published by respectable institutions like McKinsey [3] or Stanford University [4] supporting the argument that AI is a growing field in both academics and business. Although Machine Learning and Deep Learning are subsets of Artificial Intelligence, the number of survey or article centering around ML and DL is rather limited. On the other hand, the number of academic papers relating to ML and DL is rising everyday. Thus, it might be useful to grasp the fundamentals of ML and DL.
Machine Learning
Arthur Samuel, an American pioneer in AI, defines Machine Learning as the field of study that gives computers the ability to learn without explicitly programmed. This includes learning from data, which in nature, is an iterative process. It is often categorized according to whether or not it is trained with human supervision (or target variable).
- Supervised Learning: A machine learning algorithm is called supervised if the training data includes the desired target, or so called label. If the target variable contains continuous numerical it is called Regression whereas if the target variable includes categories instead, it is named Classification. Most commonly utilized supervised learning algorithms can be listed as: Linear Regression, Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Decision Trees, and Random Forests.
- Unsupervised Learning: If the training data is not labeled, it is classified as an unsupervised learning algorithm. The model or the system tries to learn without a target. Therefore, it tries to solve problems focusing on other objectives: Clustering, Anomaly Detection, Dimentionality Reduction, Visualization etc. Most popular algorithms are: K-Means, DBSCAN, Hierarchical Cluster Analysis and Principal Component Analysis.
- Semi-supervised Learning: Some algorithms use partially labeled training data, which often contains a huge amount of unlabeled data and a small amount of labeled data.This is called Semi-supervised Learning and it usually utilizes the combinations of unsupervised and supervised algorithms.
- Reinforcement Learning: It is a special learning system in which the algorithm (or the agent in this context) can observe its surroundings, select and carry out actions, and receive reward or penalty in doing so. Therefore, in order to minimize the penalties or to maximize the rewards, the algorithm develops the best strategy (or the policy in this context). The prime example of Reinforcement Learning is DeepMind’s AlphaGo program. It increased its popularity by defeating the world champion Ke Jie at the game of Go in 2017. It formulated its winning strategy by analyzing millions of games, and then playing many games against itself.
Deep Learning
Deep Learning, also called Deep Neural Networks (DNN) is a subset of machine learning that makes use of concepts from representation learning where multilayered neural networks learn from vast amounts of data. It is a core enhancement to Artificial Neural Networks (ANN). ANN symbolizes a simple neural network composed of an input layer, a hidden layer and an output layer. DNN contains more deep or hidden layers but there is still no consensus on how many hidden layers makes it deep. Therefore, today Deep Learning has been rebranded as a term to represent all neural network based algorithms.
Deep Learning distinguishes itself from other machine learning algorithms with its hierarchical layered representation of the concepts as well as its effectiveness in performance and scalability with enormous data. However, as the network model constructed deeper, the interpretability of network starts to become an issue. Surely, it is the greatest obstacle for its usefulness in applications that require transparency. Although it is preferred for its great performance in many complicated real world problems, the reasoning behind the connection processes between the layers is still need to be clarified for human understanding. That is why the field is expanding in this direction to develop explanation methods in order to explain the role and importance of the input features.
Who knows, maybe the improvements in this area shed light on the meanings of the complex connections in the brain, resolving the mystery in our heads. I am truly thrilled to see advances in this field.
DISCLAIMER
All views and opinions expressed in this post are my own and do not reflect the official policy or position of any agency, organization, employer, company or entity whatsoever with which I have been, am now, or will be affiliated.
