Stitch Fix Multithreaded

I think the article overestimates how much we understand about neural networks, and the novel ways we might use them.

Take word2vec for example. It's a two-layer (i.e. not deep) neural network that uses a relatively simple training algorithm. Yet, after 30 minutes it can learn impressive relationships between English words. The method is only 3 years old and businesses continue to find novel ways to apply word2vec. Check out this article:

http://multithreaded.stitchfix.com/blog/2015/03/11/word-is-worth-a-thousand-vectors/

They are using word2vec to find clothing items that resemble other products by adding or subtracting descriptors such as "stripes", "blue", or even "pregnant."

For my company, we are attempting to normalize medical terminology using word2vec. Here are a couple examples:

1. "myocardial" + "infarction" = "MI"

2. "ring" + "finger" = "fourth" + "metacarpal"

We never would have thought to normalize the names of different body parts, but with word2vec, it does it for us. There's a great deal yet to learn about how this model alone works, let alone deeper networks.

It's true that we (as in everyone working on these problems - not just IBM) don't have this working.

Two points, though:

1) This is probably the most active area of research in machine learning/AI at the moment. The same type of progress that was made in image understanding 2 years ago using deep learning techniques is now being seen in NLP. See for example[1].

2) It isn't at all clear that "making connections" of the type he talks about is needed to generate the types on insights that he claims. For example, the work StitchFix is doing combining analogy reasoning on words along with other non-textual data[2] indicates to me that the current state-of-the-art absolutely could "understand" that Bob Dylan wrote about themes that were associated with anti-war sentiment.

[1] "Teaching Machines to Read and Comprehend" http://arxiv.org/abs/1506.03340

[2] http://multithreaded.stitchfix.com/blog/2015/03/11/word-is-worth-a-thousand-vectors/, https://github.com/stitchfix/context2vec,

When I think "successful social media marketing" I think of some kind of podcast or blog... say

https://www.flightradar24.com/blog/avtalk-podcast/

or

https://multithreaded.stitchfix.com/blog/
https://www.stitchfix.com/men/blog/
https://www.stitchfix.com/women/blog/

(e.g. it was reading about machine learning at Stitch Fix (SF) that I felt solidarity for them and became a subscriber)

That aviation podcast is like a freight train, but that is because they have been doing it so long.

If there is a royal road, it is to pay someone and ride on their momentum. For instance the SF company will hire pro bloggers, you work with an "influencer", etc.

This San Francisco entrepreneur staged a fake protest to launch Salesforce.com (SF)

https://www.businessinsider.com/marc-benioff-salesforcecom-chief-has-pulled-some-crazy-stunts-2012-3#he-hired-fake-protesters-to-disrupt-a-siebel-conference-and-drew-the-cops-in-1

I'm the author of this article.

If you find it interesting, here are some links to a couple other articles I've really enjoyed on the topic:

# Queues don't fix overload

http://ferd.ca/queues-don-t-fix-overload.html

# There is no happy path in programming

http://multithreaded.stitchfix.com/blog/2016/01/29/no-happy-path-in-programming/

Also, if you're looking for a whole book on systems thinking, albeit in a more general context, it's worth checking out "Thinking in Systems: A Primer" by Donella H. Meadows.

Once you start thinking about this stuff it's impossible to not see it everywhere. But for programmers, ops people, etc. that's a good thing... because our jobs depend on it.

Same, the difference I had in my implementation was an added an 8 week moving average. This worked for most of the year and I think even holidays. What was considered normal for the week of Thanksgiving would not be considered normal for a summer week. This was relatively easy using SQL CTEs and windowing functions.

I also separately used a package in R called bsts (Bayesian Structural Time Series) for a different way of projecting seasonality on a trend to find an acceptable normal range. If the actual fell out of range then it was an anomaly. Great write up on the technique by Kim Larsen here: https://multithreaded.stitchfix.com/blog/2016/04/21/forget-arima/

Yeah, the Rails community noticed this a while ago. Put the business logic doing transformations on models and interacting with external dependencies in service classes, which accomplishes three things:

1) Avoids bloating model objects.

2) Avoids creating complicated controllers.

3) Makes it much easier to move to a microservices architecture when the time comes, because business logic has already been grouped into isolated components.

https://multithreaded.stitchfix.com/blog/2015/06/02/anatomy-of-service-objects-in-rails/

https://www.toptal.com/ruby-on-rails/rails-service-objects-tutorial