This was really interesting, I had never really thought about this problem before! Snazzy visuals, you clearly put a lot of effort into those.

I caught that you can use neural nets to narrow down “mathematically possible” solutions to physically reasonable ones. So, as you said, the ocean can’t be pink, and a NN can use this contextual information. Maybe you did address this and I just missed it, but what if there are multiple feasible solutions? For instance, a t-shirt may be blue or pink or any number of colours, so how do you (or a NN) figure out which one it’s supposed to be in that case? If I had to give feedback, I would say to address edge cases like this for completeness.

Question: Which part of the exposition was most engaging? Answer: I really liked to animation! showing the colours in 3d space, giving a geometry intuition into colours.

Overall: The whole video had a very interesting concept of turning grey scale images back to a coloured image. It introduced many topics such as colour value encoding, thinking of colour in a high dimensional space, neural networks and more. The concept was interesting however, I would say the execution could have been better.

Critique:

Motivation

Whilst the concept of turning a black and white image into a coloured one is an interesting premise with definitely many useful implications. The video didn’t really state why we should be interested or care. I hope the introduction at least explained what kind of application this has or why this problem is interesting.

Clarity

The pacing was a bit to fast for my taste. I would have loved some pauses between the sentences. Also some concepts need a bit more like the concept of mapping colours into 3d space/ black and white into 1d space, and also the concept of the same grey scale forming a of triangle. Whilst these concepts seem trivial some viewer may find it confusing and a bit out of nowhere (not the end of the world).

Novelty

I think the concept was really novel and there were lots of new perspective I never thought about! (no complaints here)

Memorability

The whole video seems too fast and didn’t really have one focal point. So I think the memorability of this video suffered as a consequence.

Great video! The pacing was good and the motivation was strong. The topic lends itself very naturally to the chosen visualizations. It was a very high-level overview and I feel that some of the points that were touched upon could have been tied into a more cohesive narrative or deepened, especially towards the end, but I think the video achieves what it aims to do. It would have been nice if you had ended with some other examples where this kind if information loss needs to be overcome to prime the viewer to consider how what they have seen can be transferred.

Everything felt very concise, yet important for the main topic. I don’t know what could be shorter or simpler.

This was an excellent video all-around. The motivation was made really clear, and you set up all the different mathematical pieces quite nicely, while keeping all of the explanations pretty low-tech and without too much jargon (although you could use even less). The pacing was also nice, feel free to go even a bit slower.

This is definitely a video I would show to students when they ask me what the point of linear algebra and high-dimensional spaces is.

Everything was straight to point, that’s it.

As a high schooler I couldn’t understand all the Terms you used and it went a bit fast to fully grasp everything. Understood the main concept but things like 3 million dimensions doesn’t make much sense to me.

This is a very interesting choice of topic, I wasn’t aware of the manifold hypothesis but it does make some sense. That said, I think the shapes that are being referred to in that higher-dimensional space are a little too abstract to really break through and make the concept intuitive - which is a very difficult task, I’m sure, but I think that clarity is very important. Without more details it’s hard to have a takeaway beyond “one part of this system has an idea of what should be possible, one part has an idea of what is probable,” and the math underlying it is sort of up in the air.

This isn’t my area of expertise so I don’t know how hard it might be to construct this, but I think this is where the idea of using a smaller example as an illustration might be powerful: you certainly can’t show what an intersection with a 3-million-dimensional space looks like, but maybe if the problem is pared down to fewer dimensions a more concrete example could present itself? What if there was a version of this method where there was a concept of which colors in the 3-dimensional space showed up more or less often - that’s a space you’ve already visualized, so showing a manifold embedded in that space should be simpler, right?

I don’t know, I’m spitballing a little. My point is, although the abstraction is something that seems to give the manifold hypothesis its power, instantiating it a little more tangibly might make things a bit clearer.

Thank you, I really enjoyed the content - it’s interesting also for its applicability. I remain uncertain about how to introduce it to high school students (many struggle with learning difficulties that affect their understanding of 3D graphs). Which topic does your content connect to? In what way can it be useful? Does it help reinforce prior knowledge? I couldn’t find an answer for high school, especially since the final topics are quite advanced. Perhaps it is more useful for university students.

I like it that the title appears so innocent and the viewer is heated up slowly to boiling temperature :-)

Would have loved some references and sources. For example for the coefficients used around 01:30 for turning colours to grey.

Overall, a very lovely video! But you touch a few too many different topics for such a short video: colour spaces, information theory, image process, machine learning etc. Would have preferred a longer video or a series of dedicated videos exploring all these aspects in more detail.

A nice introduction to the topic, but the video has too little maths.

Great video! I really liked that you focused more on building intuition behind some results, rather than diving into heavy math.

I really liked the “reality manifold” concept. Thanks for introducing me to it!

Very good. Great story, graphics, audio and all in a sub 10 min video. One small piece of constructive criticism: I found that there was a jump in complexity when we got to Unet NN. Before that the video did not assume any prior knowledge and patiently too the viewer along how colour images work. Then suddenly it assumed familiarity with NN architectures and terminology.

The problem outlined in the beginning of the video was fine, but it was never explained why this problem is important or why we should care. Although it’s somewhat self-explanatory in this video, from a pedagogical standpoint, this isn’t great. Also, I found the video moved much too fast, with almost no time between sentences, making it very difficult to understand what is being said.

Unfortunately, the end result is that I got very little out of this video, and I suspect the target audience would also get very little out of this video.

I feel like highly technical and specialized terms like “convolutional” and “inductive bias” without further explanation might be confusing for people who don’t have that background, which is who the video seems to be targeting

But the manifold based intuition is a cool way to start thinking about the structure and regularity of extremely high dimensional data