Sorry about all the spam, my twitter was hacked 🙃

[4 of 4] Good paper on this is

[3 of 4] Please 1. report your standard errors from multiple runs and splits. 2. publish your hyperparmeters (and code and data), and how you got them (e.g. finetuning). Just because you can get it published / get your results to happen once does not mean that it's significant.

[2 of 4] If I had a dime for everytime I've read "we've improved the SOTA from an WER of 6.09 to 6.07", I'd be rich. In some cases the .02 improvement is significant, but in other's it's not. We can't tell without properly published metrics. You (should) already know this, but:

[1 of 4] Doing a lit review for some #ML. Unpopular opinion: Every time I do one of these I never cease to be amazed at the lack of rigorous reproducibility in the field and the amount of overclaims and "garbage science" (including from the big labs).

[2 of 2]If we had cut off the test early due to the technical difficulties, we would not have learned how tollerant users are to accuracy drops, latency spikes, etc (apparantly more tollerant than me)… It’s always a good day when you invalidate an assumption #UserExperience

[1 of 2]Interesting experience: due to tech difficulties (wifi, etc…) I thought that several usability tests had bombed, but when we got surprisingly good reviews from the users despite this, I remembered that we run experiments specifically to get these surprises #entrepreneurs

words cannot express my pain when a carefully crafted model that I poured my love and soul into is outperformed by the naive baseline T.T, which happens on a weekly basis #deeplearning #machinelearning

somehow just found out about numba lol, who needs vectorized operations anyway? (jkjk) #numba #python #numpy

2AM thoughts: is engineering possible without crunchtime? #SoftwareDeveloper #programming #Engineering

Just discovered the Bayes Rules book ( https://t.co/rgxlDHXPtL). Wish I had this to lay the framework before delving into Probabilistic Graphical Models. Highly recommend to anyone looking to get into Bayesian learning. #MachineLearning #bayesian #learning

We are so excited for another day at @SummitAmericas hosted by @USChamber! They showcased business from all over the world, including Sign-Speak, a sign language recognition and avatar solution being created in the US! Let's see what the day holds #growwithgoogle #accessibility

Favorite part of having RSI is the random realization that it decided to flare up this week for no apparent reason 🙃. Any advice would be much appreciated. I kinda want to start coding by voice, but it's not gotten bad enough yet to justify the investment. #engineer #technology

The joy of choosing between rolling it yourself or using an large external dependency #tech #engineering #SoftwareEngineering

be me: accidentally sets learning rate to 0 and debugs why the loss won't decrease for a solid 15 minutes

Training a particularly precarious model right now. Making a sacrifice to the god of NaNs.

Why does #ML take so long to train??? I get that I'm asking you to learn an entire semester's worth of content from a monstrous dataset, but can't you hurry it a little up. Maybe I'll just move my learning rate to lr=100 lol

So incredibly proud of my #cofounder @YamiPayano. Unfortunately can’t announce any of the big news yet, but YOU ROCK! #entrepreneur

this is my life now #Entrepreneur

I think both my least and most favorite part of #MachineLearning if the amount of time it takes for jobs to complete. Gives alot of good reading time, but sometimes, I just wish it would hurry up :P