IEEE BYTE Volume-3 Issue-2 | Page 23

    9      guess coffee. However if an entry had 25mg caffeine and cost 150​ ₹ ​ , our machine might have difficulty in guessing tea or coffee and could possibly make a wrong choice. Now, since the correct answer deviates from the line, the computer refines the model slightly by adjusting the line and bringing it closer to the correct answer. The computer thus ‘learns’ from its mistake and improves itself to not repeat it again. The computer repeats the process over and over with different data, tweaking the model a bit each time until it is confident of guessing the correct answer with a high degree of probability. Now, one might say - why not train the model more and more to get the ultimate guessing machine? Here’s where overfitting comes into play. Overfitting is where the computer has been fed too much data pertaining to a certain case that it starts predicting answers incorrectly. For example, if you start including too many tea varieties in the dataset that have very high caffeine content, it might incorrectly start assuming even some coffee varieties to be tea. Here, the deviations(called noise) start affecting the computer’s ability to ‘learn’ from the correct answers(called signals). Underfitting is the opposite, where our model is too general and flexible in it’s parameters. After testing, if necessary, some parameter tuning is done. Finally, we can use our ML model for what it was intended - real-world predictions. While this may not sound all that difficult, actual ML models can be incredibly intricate structurally and functionally. Today, the applications of ML are exponentially snowballing. While we cannot ascertain when AI will overtake human intelligence, it is unequivocally clear that Machine Learning will pave the way.