The Line Equation as a Tensor Graph - Topic 65 of Machine Learning Foundations

#MLFoundations #Calculus #MachineLearning
In this video, we get ourselves set up for applying Automatic Differentiation within a Machine Learning loop by first discussing how to represent an equation as a Tensor Graph and then actually creating that graph in Python code using the PyTorch library.
There are eight subjects covered comprehensively in the ML Foundations series and this video is from the third subject, "Calculus I: Limits & Derivatives". More detail about the series and all of the associated open-source code is available at github.com/jonkrohn/ML-foundations
The playlist for the Calculus subjects is here: • Calculus for Machine L...
Jon Krohn is Chief Data Scientist at the machine learning company Nebula. He authored the book Deep Learning Illustrated, an instant #1 bestseller that was translated into six languages. Jon is renowned for his compelling lectures, which he offers in-person at Columbia University, New York University, and leading industry conferences, as well as online via O'Reilly, his UA-cam channel, and the SuperDataScience podcast.
More courses and content from Jon can be found at jonkrohn.com.

[@theseusRJ7]

really excited for next one!

[@JonKrohnLearns]

ahhhhh yeah!!!

[@morhaflababidi6347]

I am still confused why did you start with a 0.9 slope and emphasized on having a value near zero ? Wouldn't it be more easier if we start with a negative approximation of the slope and the y-intercept parameter near 2 ?
And thanks for the series. It is really helpful

[@JonKrohnLearns]

Glad you're finding the series helpful, Morhaf, and good question!
When we randomly initialize the parameters of an ML model, it's typically a good practice to start with near-zero values because values far from zero would imply a strong opinion or that learning had already occurred.
For the purposes of this toy ML demonstration, we aren't truly initializing with a random value. Instead we're manually selecting a "random" value. You are welcome to choose a different value yourself such as one that approximates the negative slope you can see (however, with real-world ML models, we'd rarely be able to "eyeball" parameter values because there could be millions of them). I deliberately chose starting values that *do not* look like they would fit the data well right off the bat so that we can clearly observe a model learning.
Make sense?

[@justsimple2223]

Thanks for this wonderful question.