love reading the threads, want to challenge myself:
Machine Learning <-> Data visualization
1 like = 1 opinion
Machine Learning <-> Data visualization
1 like = 1 opinion
1. all data is subjective.
data are measurements of systems taken by particular people from a particular perspective
data are measurements of systems taken by particular people from a particular perspective
2. the future of data visualization is machine learning.
3. the future of machine learning requires data visualization
4. data visualization is about finding meaningful representations of data to convey understanding of a system
5. a lot of machine learning is about finding efficient representations of data in order to accomplish a goal.
those representations can (but don't have to) be meaningful
those representations can (but don't have to) be meaningful
6. whatever makes a "thing" a thing is what we are trying to figure out.
e.g. people are posting pictures of their lunch on the internet...
that's a thing?
e.g. people are posting pictures of their lunch on the internet...
that's a thing?
7. this is why ML is the future of datavis, it can point out a bunch of "things" at scales we can't handle the manual way. then gives us numbers about those things we can turn into graphics
8. this is why datavis is essential to the future of ML, we need to be able to talk about these "things". we need to make decisions about them, judge them as good or bad. that's too hard to do by staring at a raw tensor
9. there is no such thing as a black box, just something that hasn't been looked at hard enough. if we can know about black holes and star formation we can figure out what these relatively simple machines are doing
10. patterns are what we usually call those "things" I was talking about. not all patterns are meaningful, and what form they are expressed in matters a lot.
i.e. using polar coordinates can make anything involving angles much easier to see
i.e. using polar coordinates can make anything involving angles much easier to see
11. what we really want is to find and share the patterns that mean something to us. that meaning comes from outside of the system we are examining, but will probably affect it.
12. we find patterns from the data we look at. we give the patterns meaning and either change the system that produces the data or change the minds of people who affect the system that produces the data
13. being limited to a 2D screen is the main reason data visualization is so limited and fleeting.
there are people doing wonderful things within these constraints but we will be able to do much more
there are people doing wonderful things within these constraints but we will be able to do much more
14. focusing on the architecture of models is limiting the application of ML, the focus should be on the data, especially control over the data
15. datavis is an inherently multidisciplinary field. there are strong technical components: working w/ data, UI programming, but design and story telling are necessary. not one person in the world does all of these at the highest level
16. As ML becomes more accessible, we could see it transform the "working with data" aspect of datavis enormously. currently it adds more technical burden but should end up lowering it
17. charting libraries are stupid
18. ML framework doesn't matter
19. front-end framework doesn't matter
20. what matters is being able to combine the elemental building blocks of the discipline while understanding the purpose of what you are trying to create
21. this is why #d3js is so powerful and lasting. it is not a charting library, not even a framework. it is a collection of useful abstractions for authoring data-driven documents.
as you grow fluent in it, you become more expressive.
as you grow fluent in it, you become more expressive.
22. what is the equivalent for ML? I'm still learning but I studied linear algebra in grad school in '09 right before deep learning became "the shit"
a collection of abstractions over linear algebra and statistics that enables the exploration of data spaces is what I want
a collection of abstractions over linear algebra and statistics that enables the exploration of data spaces is what I want
