The Same Job, Two Ways of Doing It | AI Learning Guide

Start with what the two approaches share, because the differences only make sense against that backdrop. Both are given examples of inputs paired with correct outputs, and both must produce answers for inputs they have never seen. A spam filter, a price estimator, and an image classifier all have this same shape: data in, a useful answer out. The models are not meant to memorize the training examples; they are meant to generalize beyond them. So when we compare traditional machine learning and deep learning, we are not comparing two different goals. We are comparing two different routes to the same goal, and the interesting question is who decides what matters in the data along the way.

0:00 / 0:00

Auto-play

The shared task

Both approaches take input data and produce a prediction or decision that should hold for cases they were not trained on. This is the job description they have in common.

The same shape of problem

Consider three tasks. A spam filter maps an email to spam or not-spam. A price estimator maps a description of a house to a number. An image classifier maps a photo to a category. In each case the input is data and the output is a useful answer, and in each case the model must work on inputs it has never seen before.

Where the paths diverge

The two approaches agree on the destination and disagree on the route. One relies on a person to decide which properties of the input matter and to encode them explicitly. The other is handed the raw input and works out its own internal representation of what matters. Both routes end in a prediction; they differ in who does the deciding.

Why generalization matters

A model that only reproduces the answers it was trained on is useless. The point of learning from examples is to handle new ones, and this is true for both approaches.