Summer of Math Exposition

Presented by 3Blue1Brown 3blue1brown

The Universe if F=mv

Audience:

Tags: physicsmachine-learning

What if F=mv instead of F=ma? The video follows how this new law would change physical motion, then uses that setup to explain a surprising connection between physics and machine learning optimization. In particular, it shows why gradient descent behaves like strongly damped motion, how learning rate relates to how strongly a system responds to force, and how momentum brings back some of the inertia that the F=mv universe removes.



Analytics

3.17 Overall score*
118 Rank
6 Votes
4 Comments

Comments

1.2

The entire premise of the video was unclear. Primarily because the sound just did not work. From what I could understand: you were exploring what happens when F=ma is swapped with F=mv as the physical law, your work was also unclear.

3

Interesting. Voice over would improve it, you should let more people watch it though. Post it on YouTube!

1.9

Ok so clearly I’m looking at an unfinished project here (had to download the video from google drive, it is called ‘superroughcut’ and has no audio), but from what I’m getting I can’t judge whether there is something really intriguing going on here or whether it is just nonsense. It might be that you’re on to something very interesting, but it doesn’t come across. At the very least the video would require some narration with motivation and an explanation of what’s happening.

4.5

The idea is interesting, particularly the way the video uses the hypothetical (F=mv) law to show how the usual second-order equation of motion turning into a first-order equation. I also liked the connection to machine learning and the comparison between gradient descent, damping, and momentum.

However, I felt that the explanation could be developed further. The video presents an interesting setup and makes some intriguing connections, but it does not quite explain why those connections are important or what deeper insight we should take away from them. In particular, the relationship between the modified physical law and optimization could be made more explicit, so that the machine-learning connection feels less like an analogy and more like a meaningful consequence of the mathematics.

Overall, it is a nice idea with good potential, but I would have liked a clearer explanation of the motivation and the main takeaway.