Application of Machine Learning in Artificial Intelligence
The world was watching breathlessly. As the timer ticked its way towards zero, Lee Sedol chewed his lip nervously, his eyes never leaving the board. In the busy streets of South Korea, life had reached a peculiar standstill as crowds gathered before the giant screens, silently willing their champion to amaze them once more. The feeling was shared by viewers all over the world, for this was no ordinary Go match. Across the board was AlphaGo, an artificial intelligence based Go playing program, facing off against the reigning world champion in the ultimate test of man versus machine. Lee would eventually resign, and in the course of the next 4 matches would lose to AlphaGo in a stunning 1-4 defeat.
That was in 2016. Today, 5 years later, we are on the brink of an explosion of artificial intelligence. Originating from a 1956 workshop at Dartmouth College, the field of artificial intelligence (AI) has an eventful history, marked by several waves of intense interest by the scientific community followed by periods of disillusionment known as “AI winters”. Coined by John McCarthy, artificial intelligence, in simple terms, stands for the forms of “intelligence” demonstrated by computers and other digital entities.
Learning abilities of machines (i.e. machine learning) are of critical importance in AI. In the words of computer scientist Tom Mitchell, “machine learning is the study of computer algorithms that allow computer programs to automatically improve through experience”. Computer algorithms aside, this should ring a bell. “Improvement through experience” is precisely the way that we humans learn! Perhaps fueled by this realization, there have been many successful applications of machine learning in the pursuit of artificial intelligence.
A student expecting to sit for the A/L examination could choose to prepare for it in an unusual way. Instead of studying and understanding the concept behind each problem, he could simply collect an enormous quantity of past papers, read each problem and digest the sample answer given. At the exam, he could sift through his mental bank of questions and use previously observed patterns to produce the correct answer. Widely practiced by tuition classes today, this method is not entirely unsuccessful given the repetitive nature of certain questions.
This is a perfect analogy for a type of machine learning algorithm commonly known as “supervised learning”. The procedure involves feeding a “training dataset” to the algorithm (which is coded in the computer using a programming language). The training dataset consists of sample data attached with “labels” which represent the output that the algorithm should learn to predict. This is where machine learning works its magic; the algorithm produces a model that reflects the patterns observed in the training dataset. The model is generally produced by attaching numerical weights to “nodes” in the code of the algorithm such that the input is matched to the output with a minimal amount of error. Having been trained on this dataset, the algorithm’s accuracy could then be tested on a “test dataset”.
Speech recognition software such as Google Assistant, Amazon’s Alexa and Microsoft’s Cortana show this abstract algorithm in action. Since these programs are equipped to receive data from their surroundings, they mimic the fashion in which humans interact with their environment, and thus are prime examples of today’s AI.
Click the YouTube icon on your smartphone. The first page on display is a list of videos that YouTube somehow predicts you will enjoy. Akin to items on display in a supermarket, the video hosting platform has also taken the pains to organize videos under clusters of topics. The science at work here is machine learning’s clustering algorithm.
The clustering algorithm used in artificial intelligence is a common example of a type of machine learning algorithm called “unsupervised learning”. Unlike in supervised learning, this does not involve labelled data and instead feeds the algorithm an array of raw data, which it peruses to predict any patterns using probability densities. Several clustering algorithms are currently in use including k-means clustering and Expectation Maximization. E-commerce companies such as Amazon and streaming services like Netflix benefit heavily from these since the algorithms “interact” with user data to deduce preferences which can be addressed using targeted marketing.
The third major type of machine learning algorithm is “reinforcement learning”. This is the closest form of learning to that of humans since this consists of direct interactions with the environment. Each interaction (or “action”) is associated with a “reward” whose value depends on how effective the interaction was. For an example, a robot learning to navigate through a maze would receive a positive reward for moving in the direction of the exit and a negative reward for bumping into a wall. Eventually, the robot’s algorithm learns to predict a sequence of actions to maximize its cumulative reward.
Reinforcement learning is at the core of some of the most publicized developments in AI such as self-driving vehicles and strategy game programs. Elon Musk’s Tesla has invested millions of dollars in the autonomous vehicle industry and BMW plans to have a fully self-driving car in production by the end of 2021.
Undoubtedly, the future of artificial intelligence is brighter than the headlights on Tesla’s newest electric car. However, the possible existential dangers of AI reaching human-level cognition (Artificial General Intelligence) are widely discussed. Will AI usher in a technological utopia? Or are we in line to see a sequel to Terminator 2? We shall leave the predictions to the algorithms.
Bibliography:
• Dartmouth conference: * McCorduck 2004, pp. 111–136
• Second AI winter: * McCorduck 2004, pp. 430–435
• Hinton, Geoffrey; Sejnowski, Terrence (1999): Foundations of Neural Computation. MIT Press.
• Lloyd, S. (1982). "Least squares quantization in PCM". IEEE Transactions on Information Theory. 28 (2): 129–137
Image Courtesy:
• Fig.1: https://miro.medium.com/max/977/1*DjUEt5--t6lCjYG_MuZlLg.png
• Fig.2: https://static.javatpoint.com/.../k-means-clustering...
• Fig.3: https://miro.medium.com/max/592/1*pEcq5MyHSSO6bgIiWE3GOQ.png
0 Comments