Introduction
How does your machine learning algorithm actually know if it is a good or bad model? How does your machine learning model get the best accuracy for your dataset? Well, the answer is by using a loss function. The model will try to minimize its loss function and that model – according to the computer – is the best model for your dataset. this is just the basis of what Cassie Kozyrkov talks about in this video of her course(the video can be found here).
Summary of Article
Cassie starts out by defining a loss function which she calls a loss function “the actual function that your model will work with[to know if it is doing good things]”. The machine will try to minimize this function since less loss is better. She then touches on the point of for any business they have a metric that they care about(Cassie takes accuracy as an example), and a function that the computer is working to minimize(the loss function). If the functions are moving in opposite directions(as the loss function goes down the metric that you care about gets worse), machine learning will be working against you. So, to start solving this she simulates a fake dataset, and independently makes the computer optimize based on the loss function and on the metric that the business owner cares about(accuracy). She does this in a loop with different parameters(different datasets, different sizes, different proportions of each outcome, etc.). Finally, Cassie plots this on a graph and then shows how, in this case, the two functions are working together so this is an acceptable loss function to use. To end the talk Cassie stresses that you don’t need to choose your loss function yet, but this is just a preliminary step to “make sure you are not asking for the moon”.
My Take
Overall, I liked this video. It was clear, concise, and informative. I learned a lot about the actual applications of the loss function and how it is used in a business sense. I also learned a lot about the steps to get ML to use in a business. The fact that there are so many preliminary steps that seem small but can really add up, didn’t occur to me. I also really enjoyed the visuals, as they really made the talk livelier. The talk was given in a very beginner-friendly way, which was really helpful, and made it so that it was easy to understand while also being extremely educational. One more thing about the talk is that it went into more of the technical aspect of the topic, which I really liked. The talk gave equations and then proceeded to explain those equations in a way that appeals to everyone.
Conclusion
All in all, this was a good video. I really enjoyed its content and the way it was presented. It allowed me to understand how this is applied in a business sense while still giving me insight into the technical aspect of loss functions. I really recommend you go watch it when you get the chance(the video can be found here).