I liked that they are doing this with the simplest and most well-understood techniques, matrix factorization+PCA.
Though I'm sure they're also trying end-to-end extra-deep programmatic multi-media networks with new kinds of convolutional layers, fancy residual connections, an experimental batchnorm variant, etc... I'd love to see if they squeeze more juice out of that.
As we make more of these semantic meaning-spaces like word2vec, or the meaning space of "photos of airbnb rooms", I think we could use some foundational UI design work to navigate them better.
Like, now you can navigate everything on stitchfix through a small 50-d space where every component is nearly orthogonal and can kind of be interpreted ahead of time+labelled.
Is "a binary tree in the form of playing N questions" the best way to go through them, or are there other options? Should we allow navigation in more than one dimensions at a time? Could we start with a few landmark, prototypical elements in each dimension? Have the user progressively clamp down the range in components, and show a PCA of the remaining components each time? Would a 3rd dimension (with VR?) help separate more dimensions at a time, giving you 50% more dimensions to show items on?
Computers are increasingly understanding our world and so far they understand it through these meaning spaces, so I think this would be incredibly important for the future of UI design and computer interpretability.
It will be interesting to watch as they improve this process. I use Stitch Fix, and I don't feel like it's all that personalized. The problem is that as a consumer you don't see all the products they could have selected for you. Maybe there's three options for male jeans, maybe there's fifty, it's opaque to the user. The result is a little disappointing because as an American consumer my options are nearly unlimited (only my time is constrained).
I really like how well they write their technical articles, but their service leaves something to be desired. I became interested in Stitch Fix a few years ago when I saw a talk at a tech conference. It was intriguing to me because a person's choice of fashion is very complex and nuanced. I've also done a lot of experimenting with my own fashion tastes over the years. I'm also an ML junkie, so Stitch Fix scratched that itch as well.
I used them for 4 months pretty much right after they began offering men's clothing. I was pretty disappointed with every delivery. I returned anywhere between 50%-80% of the items. After a time I threw out some of the items because I had only worn them once and needed closet space. Out of 20 items they sent me only 4 remain in my wardrobe.
Why was I dissatisfied?
1) I'm a guy who shops anywhere from Barney's to thrift stores. A good deal of Stitch Fix's picks had inconsistent quality vs. price.
2) I'm a gay man. One reason for my fashion choices is to make me attractive to other men. The choices made by the stylist (always a woman's name on the insert card) made me feel like I was man being dressed by his girlfriend/wife. There is no choice in your profile to choose the type of stylist you want.
3) No sales. You did get a decent discount if you bought the entire delivery. I never wanted to do this, because it never worked out money-wise.
I think that's the point of stitchfix - you're outsourcing your shopping to them to save time. It's been years since I went to a store and bought clothes. Clicking around on Amazon is really hit or miss - maybe the actual product looks like you thought, and maybe the colors were a bit off, plus you have to put together outfits that go together.
I'm also a Stitch fix user and I think that they've done a great job of finding clothes that I like and the note from the stylist says "hey you can pair these new jeans with that shirt you got last month". And I didn't have to spend anytime thinking about it.
However, it's reached the point where they may be TOO dialed in. The clothes I got this month were almost the same as last month. Do I really need another polo?
I'll probably change to their "delivery every 2-3 months" option, since after 6 months of using them I have enough new clothes.
I'd love to use stitchfix, but they continue the tired trend of clothing retailers not selling clothes in my size. I have a friend who wears a 30/36 jeans, and he buys all his clothes from a cowboy store in Arizona. I'd rather not have to go down that path but clothes that fit me seem harder to come by every year
Engineering is clearly reaching a point where it is and will be much more integrated and coupled with pure and applied mathematics. I think it's more than time to start taking this implicit project seriously by making it explicit. We aught to think back to the foundations of the way we think about programming in the first place, and we should try to build a new mindset which takes the best parts of the hacker spirit and the pure mathematics mindset and create something new, something more fitting to the enviourment we've found ourselves in, where essentially the only limits to what we can produce is our ability to imagine them.
Eh—sometimes I feel like we're repeating many of the same stupid mistakes our parents did only with AI instead of mechanized production. They had ugly, mass produced shirts. Big ugly homes. Big ugly cars on big ugly highways.
I don't need AI to dress me. Looking good takes work, but it's nice to look sexy and to feel like all the effort you put toward finding the perfect shades or shirt was worth it. I don't need an algorithm to feed me or to determine exactly when I should wake up or how long I should exercise.
I need algorithms to distinguish a tumour from a cyst at the hospital. I'd like to do the human stuff.
That's an optimistic idea, and I don't want to discourage optimism. But a common trait shared by both engineering and mathematics is the notion of constraints. The limits of what we can produce in such an idealized math/engineering framework are likely not to be our imaginative limits, but the inherent and necessary limits of the objects and abstractions that math/engg rely upon. You might discover that axioms, rules, boundary conditions, and formal abstractions can't produce the liberating blank-canvas that an imaginative hacker/artist would dream of -- at least not without throwing the baby out with the bathwater.
From one perspective, coding is just dry formal automata theory. But that's not how most coders think about it. We've developed more playful ways to think and learn about it. And the more experience one has with coding, the more one can ignore the details and just think with feelings and intuition. The same effect occurs after a mountain of experience with pure mathematics. After a while, it starts to change the way you think. The brain internalizes it, and makes it accessible on a much more natural plane than how you think about maths when you're starting out, when everything is detail and hard work. It's these two worlds I want integrated. I want the abstractions coders think about on a daily basis to be enriched by a rejuvenating syringe of math juice.
This is a great article, and it reveals some interesting new ideas to test out. However, I wish they included examples that qualitatively demonstrate how effective each improvement was. Some of us don’t have a team of almost 100 high-end data scientists like they do. Since I only have a couple of people, I need to always be careful to focus on techniques with ROI (low implementation cost, (relatively) high effectiveness). Machine learning can be labor intensive.
I think the audience is more the quantitative "data science"/ machine learning crowd.
The nice friendly web design makes this look deceptively simple (and it is if you have the right background), but this type of material is not aimed at a general audience.
At the same time, as noted by another commenter, this isn't arXiv material either.
I think it's a recruiting tool, so they are okay if only the jargon is indecipherable by most readers. In fact, it's probably signalling to their targets that they are serious about machine learning.
I found it enjoyable and understandable, but I guess I was already slightly familiar with matrix factorization for recommendation/characterization problems.
Stitch Fix is a profitable company that recently had an IPO. Their business is mostly subscription: each month they send a small selection of clothes chosen by a stylist. The customers keeps (and pays for) the ones they like, and returns the rest. Here's an overview: https://www.sramanamitra.com/2017/06/15/billion-dollar-unico...
So I have mixed feelings about this. I think it is good that StichFix is publishing how they work, and I have complete admiration for what they have achieved.
I still use their post about the ability to add a "pregnant" vector to an item of clothing and find a similar item in a maternity version to explain how objects can be embedded in a vector space.
But this post seems both too complicated for the general reader, and not deep enough for a paper.
I liked that they are doing this with the simplest and most well-understood techniques, matrix factorization+PCA.
Though I'm sure they're also trying end-to-end extra-deep programmatic multi-media networks with new kinds of convolutional layers, fancy residual connections, an experimental batchnorm variant, etc... I'd love to see if they squeeze more juice out of that.
As we make more of these semantic meaning-spaces like word2vec, or the meaning space of "photos of airbnb rooms", I think we could use some foundational UI design work to navigate them better. Like, now you can navigate everything on stitchfix through a small 50-d space where every component is nearly orthogonal and can kind of be interpreted ahead of time+labelled.
Is "a binary tree in the form of playing N questions" the best way to go through them, or are there other options? Should we allow navigation in more than one dimensions at a time? Could we start with a few landmark, prototypical elements in each dimension? Have the user progressively clamp down the range in components, and show a PCA of the remaining components each time? Would a 3rd dimension (with VR?) help separate more dimensions at a time, giving you 50% more dimensions to show items on?
Computers are increasingly understanding our world and so far they understand it through these meaning spaces, so I think this would be incredibly important for the future of UI design and computer interpretability.