Welcome again to the Illustrated Machine Studying sequence. In case you learn the opposite articles in the series, you already know the drill. We take a (boring sounding) machine studying idea and make it enjoyable by illustrating it! This text will cowl an idea referred to as Incremental Studying, the place machine studying fashions be taught new data over time, sustaining and constructing upon earlier data. However earlier than entering into that, let’s first speak about what the mannequin constructing course of seems to be like right this moment.
We often observe a course of referred to as static studying when constructing fashions. On this course of, we prepare a mannequin utilizing the newest obtainable information. We tweak and tune the mannequin within the coaching course of. And as soon as we’re pleased with its efficiency, we deploy it. This mannequin is in manufacturing for some time. Then we discover that the mannequin efficiency is getting worse over time. That’s once we throw away the prevailing mannequin and construct a brand new one utilizing the newest obtainable information. And we rinse and repeat this similar course of.
Let’s illustrate this utilizing a concrete instance. Take into account this hypothetical state of affairs. We began constructing a fraud mannequin on the finish of January 2023. This mannequin detects whether or not a bank card transaction is fraudulent or not. We prepare our mannequin utilizing all of the bank card transaction information that we had from the previous one-year interval (January 2022 to December 2022) and use transaction information from this month (January 2023) to check the mannequin.
On the finish of subsequent month we discover that the mannequin isn’t doing too nicely towards new information. So we constructed one other mannequin, however this time utilizing information from the previous one-year interval (February 2022 to January 2023) to coach it after which use the present month’s information (February 2023) to check it. And all information outdoors of those coaching and testing intervals is thrown out.