Fasttext +CommonCrawl. Fight me
Machine learning engineer, AMAA
Zachary Brooks
Gabriel Rodriguez
There is a thread already.
Wyatt Ross
As much I agree that this is a duplicate, this specific thread is actually a good quality tech thread with lots of good knowledge being shared. This thread is far better than the shitting Firefox threads that happen on every Firefox event.
Asher Parker
Fucking posseur you know high tech stuff for white non Pajeets have to be done in ASM right?
Tyler Ortiz
Some debugging ideas:
Log curves of loss function.
Log the activations: sometimes neural networks get stuck in a region where activations of each neuron are always max, and so the network cannot learn as the gradient is always zero. This is usually caused by bad initialization.
Even though you're building this yourself, you can check it tensorboard for inspiration.
Finally check your own code, both for the forward pass or gradients. John Carmack has some useful advice on rolling your own deep learning code facebook.com
Hudson Lopez
That's a good way to get word embeddings but word embeddings are a tiny part of understanding natural language.
Josiah Williams
You've noted that you feel you are acting against mass surveillance (Props I guess) but isn't machine learning the cutting edge of data collection right now?
Don't get me wrong I think the tech is cool and it is one of a vast many things my pleb ass wants to know more about, but there are plenty of applications for it beyond just camera surveillance, including traffic analysis and profiling. How do you think the average user should counteract this?
Eli Ross
I think it's necessary to distinguish between ML and data collection. E.g. the police can operate fake cellphone towers to track people's location, but this doesn't require machine learning. So a lot of issues around data collection are not related to ML.
That said, ML can be used to analyze video and track people, number plates etc. And it can also go through data that was collected (e.g. location data) and extract even more info from it.
The only protection for a person is to not reveal that info in the first place. For video collected in public there is no way to prevent this. I'm not an expert but I think there are ways to prevent or minimize how much location, browsing etc. data is collected about you.
The work I do is general purpose so I can't control how it used. But when I do work on specific applications I avoid things that could be used for the purposes.
Chase Price
Question on feature selection on text classification: TF-RF for supervised, TF-IDF for unsupervised in Neural Network...
In what situations are Information Gain, Odds Ratio or Chi Squared useful (SVM, kNN, Neural Networks)
Base assumption is that there is no feature over-load aaai.org
Caleb Walker
Adagrad vs Adadelta vs RMSProp vs Adam vs adamax vs amsgrad
Which one is the cheapest to run? Which one is the most optimal?