Bayesian Probabilities in Machine Learning
https://www.newyorker.com/magazine/2019/10/14/can-a-machine-learn-to-write-for-the-new-yorker
Bayes’ Rule can be seen as a model for decision-making under uncertainty. A concept that predicts the general shape that a cascade will form with wide applications to social networks, statistics, probability, and more. This sequential nature of decision-making is what enables the highly-accurate suggestions made by Smart Compose, a feature that Google introduced to the one and a half billion people who use Gmail. Suggestions are composed by the helper using a probabilistic algorithm that constantly updates probabilities for the next word to be typed in the sentence based on the words you’ve already written. The “predictive text” guesses your likely thought process and can append the AI’s suggestion to save your fingers the trouble.
In the New Yorker article attached, John Seabrook explores just how far such Bayesian principles could be applied in the written word. Smart Compose does more than simply fix spelling mistakes, it comes up with words using the predictive power of deep learning, an intensive application of probabilistic thinking in large data sets that falls into the field of machine learning. AI advances in the areas of navigation, search engines, autonomous vehicles, and now even writing, make use of these probability calculations to become more and more accurate with their results – entirely independent of human input.
Larger AI systems, neural nets such as the GPT-2 mentioned in the article, could go even further with probabilistic computations to build a machine that effectively “thinks” like a human being. Such machine intelligence based on a concept as simple as Bayes’ Rule shows the importance of conceptual understanding when building one’s foundations in fields like computer and information sciences. This is highly relevant to the topics of information cascades and network effects in how people and companies are able to harness the power of big data to more effectively navigate social network properties and predict consumer behavior. Groups of people make decisions sequentially, and forecast models that apply the predictions of Bayes’ are surprisingly powerful in their ability to determine consumer behavior. But maybe more importantly, machines are becoming better at replicating such human behavior, and the applications of such a development could have far ranging effects on networks of all sizes.