zeroxfe

23 ms·

241.
by zeroxfe 7y ago
Yes, these are two very different things. Differential programming is about building software that is differentiable end-to-end, so that optimal solutions can be calculated with gradient descent. Probabilistic programming (which is a bit mo
242.
Show HN: VexFlow has a new font stack, and supports W3C SMuFL(github.com)
1 points by zeroxfe 7y ago|0 comments
243.
by zeroxfe 7y ago
This is probably the most heartwarming comment I've read on the web in ages!
244.
by zeroxfe 7y ago
Gotcha -- specificity sounded like tightness to me. Thanks for the explanation.
245.
by zeroxfe 7y ago
Sounds like "Sensitivity" and "Specificity" are analogous to "Accuracy" and "Precision" typically used in data sciences.
246.
by zeroxfe 7y ago
> It also assumes out of the gate that everything needs to be able to do HA, scale for 1,000,000 instances/s etc... k8s makes no assumptions about your workloads, it just gives you tools. And it's super useful even if you don&#
247.
by zeroxfe 7y ago
Thanks for the kinds words :-) Re: resources, really Google was my biggest resource as I was learning stuff. One book I really loved was "Signals, Sound, and Sensation" by William Hartmann which connected signal processing to huma
248.
by zeroxfe 7y ago
Slowly hacking away on my open-source projects for music and audio: - VexFlow - https://vexflow.com - VexTab - https://vexflow.com/vextab - Pitchy Ninja - https://pitchy.ninja Also been working on ex
249.
by zeroxfe 7y ago
I think many of these don't apply today (bugs, fallback support, flexibility.) My current project uses (almost) entirely rem and em units, and it is so much cleaner and easier to manage than the alternatives.
250.
by zeroxfe 7y ago
It's a single-page-app, so it uses the history API to route between paths. Unfortunately, there's a bug in my implementation (which I haven't been able to nail down yet), which causes the back button not to work consistently
251.
by zeroxfe 7y ago
Nice writeup! I did a bunch of work to make one of my hobby projects ( https://pitchy.ninja ) work off-line. The trickiest part was data synchronization between IndexedDB and the database on the server, managing and resolving conf
252.
by zeroxfe 7y ago
Just FYI, that's not a spectrogram, it's a frequency histogram. :-)
253.
by zeroxfe 7y ago
> The reverse is true, embeddings are both the performance and memory-footprint bottleneck of modern NN models. They may be a bottleneck, but the alternative is worse -- you can't fit complex models with large vocabularies into GPU
254.
by zeroxfe 7y ago
> If you look at language modeling, then things there are even sparsier - typical neural language model has 1-from-several-hundredths-of-thousands for full language Most real-world models don't use one-hot encodings of words -- they
255.
Generating Spectrograms with Neural Networks(0xfe.blogspot.com)
1 points by zeroxfe 7y ago|0 comments
256.
by zeroxfe 7y ago
You're talking about the magnitude, which is spread evenly across the frequency range. The real component of the freq domain representation is in fact sinusoidal. Time-Frequency duality is a well-known phenomenon in signal processing.
257.
by zeroxfe 7y ago
Sorry to hear that. I know how it feels -- anytime I share anything online (my blog posts, my open source projects, etc.), I get beaten up by a vocal subset of readers. Sometimes they're well meaning, and sometimes they're just je
258.
by zeroxfe 7y ago
I don't think I've ever heard people say that ideas are entirely worthless. Good ideas have enormous potential value. The argument is that, in general, unlocking this potential requires great execution. You can't (in general)
259.
Time Frequency Duality(0xfe.blogspot.com)
2 points by zeroxfe 7y ago|2 comments
260.
by zeroxfe 7y ago
Thanks, that's very helpful, and explains why mel-cepstrum didn't work so well. So far, I found that I'm getting the best results with FFT + log scale, particularly against audio with fundamentals stripped.
261.
by zeroxfe 7y ago
Have not tried it -- def worth investigating. Thanks.
262.
by zeroxfe 7y ago
Thanks yes, this is totally correct -- the FFT uses a Tukey window, which should be able to match the main lobe. I got some other feedback that perceptual training tasks have a very long tail, so it's possible that the network will lea
263.
by zeroxfe 7y ago
> Using synthesized audio from midi seems like an atrocious way to train a neural net for pitch detection. Not sure why you think so. Almost everything you suggested (missing fundamentals, noise, vibrato, reverb, velocity, distortion, et
264.
by zeroxfe 7y ago
See the section "On Pitch Estimation" which addresses exactly that: --- Pitch detection (also called fundamental frequency estimation) is not an exact science. What your brain perceives as pitch is a function of lots of different
265.
by zeroxfe 7y ago
That's definitely interesting, never thought about it. Prob hard to find "angry voice" training data though :-)
266.
by zeroxfe 7y ago
I didn't look very hard for deep learning approaches to pitch detection, mainly because I was really interested in chord recognition (which is a far more interesting problem, IMO.) I didn't find any good research here (would love
267.
by zeroxfe 7y ago
I have training data which includes examples with missing fundamentals (synthetically removed), and the network does learn to recognize this. (The heavy regularization also helped a lot.) Although I was able to test with real instruments (a
268.
by zeroxfe 7y ago
About 10ms/sample (including pre-processing inputs, and generating spectrograms) on a current-gen GPU. It's a bit too high for real-time detection (compared to well-known pitch estimation approaches.)
269.
Pitch Detection with Convolutional Networks(0xfe.blogspot.com)
83 points by zeroxfe 7y ago|27 comments
270.
K-Means Clustering and Art(0xfe.blogspot.com)
4 points by zeroxfe 7y ago|0 comments

More ›