It's simple, you retrain the model several times removing one sample at a time (leave-one-out cross-validation), until the erratic behavior disappears. You repeat until there are no more instances of erratic behavior. Sorry couldn't resist the attempt at humor, but also consider it as an argument by contradiction that what you are asking is not possible. It's like asking which lesson caused me to hit a forehand over the tramlines. The answer is the one I didn't have time for.
I inherited a code base that other people thought they understood. People have very different tolerance levels for complexity and lack of control. People here complain about bad coders but the incentives out there are all aligned against quality: short stints and glorified short-term-thinking (a.k.a. Agile) and proprietary code. Imagine if you were bound to work on a codebase for most of your career, were given all the time you need to do your best and had all your code in the open, forever part of your reputation. Would you write better code?
So I am going to answer it myself. On batched data, it's a lot faster than 200ms per embedding and I'd say on a par with Infersent. On the other hand, I couldn't get statistical performance in the same ballpark as Infersent and I had to backtrack. This was training a logistic regression on the embeddings to filter some text streams according to my preferences. If I had, I would have preferred Basilica as Infersent is py2 only, hard to install and distribute and a battery killer on my laptop. Its vectors are also 4x bigger. I experienced some server errors and the team at Basilica was very responsive and fixed it, very pleased with the interaction. It would be important IMHO to publish some benchmark results for these embeddings, as it's usually done in the universal embedding literature, or serve published embeddings with known performance when licensing terms are favorable.
You quote a target of 200ms per embedding, not sure if it's one type of embedding in particular. I am using Infersent (a sentence embedding from FAIR https://github.com/facebookresearch/InferSent) for filtering and they quote a number of 1000/sentences per second on generic GPU. That's 200 times faster than your number, but it is a local API so I am comparing apples to oranges. Yet it's hard to imagine you are spending 1ms embedding and 199 on API overhead. I am sure I have missed a 0 here or there, but I don't see where, other than theirs is a batch number (batch size 128) and maybe yours is a single embedding number. Can you please clarify? Thanks
The funny thing is that this is actually how R development works. The main package repo is CRAN. If your package is not there, your inbox will fill with "can't install help". But getting your tests to pass on CRAN is an obstacle course. (see point 9 in http://kbroman.org/pkg_primer/pages/cran.html "put on your armor"). Your tests have to pass on Solaris when very few developers have access to that OS anymore. So there are directives to skip examples in the docs and what not (which should in theory run) but that weakens your test suit for everyone. So someone developed wolkswagen for R, for real (https://rdrr.io/cran/testthat/man/skip.html) It's part of a mainstream, widely adopted test package, testthat. It's how the R world goes.
One thing that fascinates me about archeology is how they would come up with new conclusions based on a single observation. Quite the smallest sample size of any natural science I can think of, only thing I can compare to is the case study in medicine and psychoanalysis. One case, but very intensely studied.
Look, email doesn't work, that's clear proof. DNS doesn't either and if the whole Internet hadn't been centralized long ago, it would have stopped working in, like, 1982. Or, maybe, they are not being honest and centralized gives them more control, like FB.
That's unavoidable. He's called manager for a reason. I would replace powerpoint for excel in my experience. Everyone wants a leader they can look up to and admire, not a PM.
If you know how to read through the lines, these documents are pure gold. Bosses who say "work under/for me"? Stay away, egomaniacs. You always work for an organization, not an egomaniac. My job is to "give you context"? Weasel words alert. Fake freedom. The best is this one: "1. you should do what you think is right 2. I can fire you"
This is free AI. If you paid for it, it would serve you. This way, it serves whoever pays for it. Email from top fintech asking me to work on their "personalized recommendation system that balances engagement, revenue, and our partner constraints". Do you see the welfare of their users in there? It's financial clickbait for naive people. Stop saying it's a tech problem, it's the uncanny valley and what not. This problem is created by billions of naive cheapskates and a system that rewards moral flexibility.
"designed to make you come back" ... but that's true of the whole ad-supported internet, broadcast TV etc. If it's free, you are the product. Still a problem, but not just a FB problem.
I got tired to skim hundreds of RSS feeds, some very noisy. So I created an NLP and ML-powered RSS proxy that scores each message for "interestingness" and re-delivers it. "Interestingness" is a purely subjective classification based on explicit and implicit feedback. Still work in progress, but in production (one client, me). It's a new take on an old project. The old one used beautiful soup, bag-of-words features and SVMs and learned too slowly. The new one uses boilerpipe, sentence vector mapping and logistic regression and is good enough. It's not me, it is the whole field that has improved so much in a decade. I haven't thought about anyone else using it. If you want to try, you'll have to fight through the installation a bit. Open an issue if needed. https://github.com/piccolbo/rightload
Why not work on the only two existential threats to mankind currently known, nuclear war and global warming? If global nuclear war or extreme climate change occur, what else matters?
And by the way, I couldn't care less about this post, it's just that I can't stand seeing ego depletion theory repeated forever and the fine minds of HN readers contaminated by it.
As an aside, the research on ego depletion that he quotes in uncritical terms has been completely discredited. The phenomenon does not exist. Of course, to know that you need to read science blogs, not books. But there are books about ego depletion, several, filled with lies. When the container is more important than the content, I suppose truthfulness is a secondary concern.
Main points:
* works with most feeds and feed readers
* filter the feeds you want, not some other stuff on top of what you have to read already
* uses 1 second yes/no feedback for fine tuning
* uses state of the art algos for text analysis
* first public announcement, simple prototype, please be kind ;)