doing grad school, just got a few weeks over winter,
Henry Howard
You can have a job and go to school at the same time you spoiled piece of shit.
Gavin Turner
i have a job at school, there is no school now, it's winter break at berkeley
nice assumptions though, brailet
John Jenkins
If you want to make real money, go post an ad on craigslist about SAT Math tutoring, kids will shill out bank if you go to a big name university
Samuel Myers
Real jobs don't stop just because it's winter. Maybe try getting a real job.
Cooper Reed
your reading comprehension is awful. please stop posting.
youre definitely right, i did that during the semester. it's bank and you get to meet a lot of asian moms too
Brandon Johnson
You're "working" at a school "job" that only exists during two semesters a year. Maybe try getting a real job instead.
Samuel Kelly
i'd rather get a phd, then get a "real" job.
Zachary Bell
You're allowed to do both.
Luis Thompson
Taking a c++ class next quarter. The school I go to assigns fair projects but often times youll get stuck and have no real idea what to do next. Mind if I ask you simple question if I ever get stuck??
Austin Nguyen
learn to stack exchange and google, they're really useful skills.
Assuming you've mastered that already feel free to make a thread. i'm a regular.
Charles Cook
I'd like to write a neural net to classify short audio clips as coming from a cow, rooster, pig, or sheep. The idea is that once it's been properly trained, a human can do their best impression of an animal and try to get it to guess correctly. Note that the learning objective is to correctly classify animal sounds, not to correctly classify a human's imitation of an animal.
Can you recommend any literature/links on neural net architecture for audio classification?
I'm also afraid of imbalanced classes - the net might tend to classify "cow" for everything.
Asher Nelson
Thanks for answering my question about random forest last time
I didn't realize that your situation was so immediate last time. Sports betting is definitely not a good idea, since it's risky as fuck. You'd want to take a year to see if you'd actually make a profit on paper using whatever method you come up with. Also, your objective is not to predict the point spread, but to beat the house's prediction, and the house has more information, experience, and phd statisticians at their disposal than you.
I know it doesn't have much to do with your skills, but if you have a car, maybe become an uber driver?
Joseph Cox
I did the tutorial on Tensor Flow and sort of understand how neural networks work. My question is: How the fuck do they train a NN work on a general image? What happens if you change the size of the image? Since the example in the Tensorflow tutorial was detecting numbers drawn on images and classifying them, what happens if I draw the number really small? I don't think it would still work.
How do people make robust neural networks that can identify things despite such changes?
Noah Morris
not OP, but if you take a picture with a higher-res image, you can just lower the resolution. You don't need many pixels to describe a digit, and if you were to train your CNN on higher-res photos, you'd be shooting yourself in the foot by forcing yourself to include a zillion more parameters than you could possibly need
I think this is an implementation of the LeNet architecture. It's pretty clear that it's relatively robust against relatively small changes: scs.ryerson.ca/~aharley/vis/conv/
I *think* what happens in practice is that we use another method to first draw a bounding box around a blob that is a digit (idk what though), then use a CNN to classify the image inside the box as a 0, 1, 2,...
Liam Jones
I have no idea how neural networks work, BUT
The first time you knew what a cow sounded like was when you were given a definitive "this is a cow" from your dad. Could you give it a "certain criteria of cowness" and have it work from there?
Matthew Turner
>I'm also afraid of imbalanced classes
You're not experienced enough to be afraid of potential problems if you're asking about a program you haven't even written yet, on Veeky Forums of all places, stop being so pretentious.
Isaiah Lopez
LSTM brah. They're the easiest to use out of the box and best for sequences.
Parker Collins
the idea is for the neural net to approximate the criteria from labeled data You give the net some functional form and learn the parameters in that function through gradient descent
kek
Mason Myers
wait, why LSTM? Where's the long-term dependency? I've always heard about it in the context of labeling words in a sentence
i was honestly expecting to use a deep feedforward NN, maybe even a CNN, and limit all input to be exactly ~2 seconds or something.
OP sorry to take the attention away from your topic. I already replied and was hoping for an answer from you
Oliver Roberts
Just pick a model and program it already you autist. Questions come after.
Christian Wilson
lol calm down
Charles Howard
Audio clips, they are sequential. You don't need LSTMs of course, but they will work better than CNNs af far as i know.
Nathaniel Watson
What is the appeal of mentally masturbating over details of hypothetical programs you haven't even tried beginning to write yet? Your concerns are inane as fuck to be having BEFORE trying to write something. You're talking about only having three classes, why would you think you would have all your test records get classed as only one of them? That would be extremely obvious, your training wouldn't even converge. Which you would realize if you stopped talking about things you don't understand on the internet and started actually programming instead.
Hudson Brooks
A cow goes MOO. It doesn't go OMO, it doesn't go OOM, always M then the OO. Now lets say you used a CNN of input size 3 to pick the sound waves up. But then the cow goes MMMOOOOOOOOOOOOOOOOOO. Now what? LSTMs brah.
Gavin Mitchell
ah okay. thanks a lot, I would never have thought of that
I'm worried about not getting good generalization holy fucking shit dude why are you so mad
Luke Gomez
I'm mad because it's painful to watch people write pretentious things like what you're writing. Why won't you just try programming something in the neighborhood of what you're describing? It would help yourself more than anyone else. What is your aversion to real content?
David Myers
also not sure if youre the user worried about imbalanced classes, but if you are make sure you balance that shit though. imbalanced classes fucks up training pretty good.
Michael Thomas
Yep, that link pretty much displays exactly what I'm talking about. It'll detect numbers that take up the whole space of the box, but it won't detect small numbers or small numbers that have been translated to some corner of the box.
I can see how the bounding box thing might work. I'd use OpenCV to do a preliminary feature detection test and pump that through the NN.
Zachary Flores
>Also the same guy who posted about betting on horse races a week ago. The AI told me to bet the lead apprentice jockey to show. They get decent horses, plus they have something to prove.
David Taylor
lmao thanks. i'll start reading
glad that helped!
okay this isn't funny anymore and i feel bad
From my experience talking to ML masters students and faculty at my school, machine learning in practice is rarely done by just hacking shit together and then tweaking it. The critical part is often done on paper. It's not OOP, it's algorithmically complex but otherwise straightforward to implement.
Look at all the shit that various anons and I have been talking about. Short comments, but they've been super helpful, and they definitely aren't late-stage tweaks. It's best to think about this now. Why spend a lot of time implementing a poorly thought-out model that gets trash accuracy, when I can just ask somebody on Veeky Forums for a short tip?
You can also look at it this way: I'm an idiot who needs his hand held through the big decisions