Meeting topic suggestions

[bmcfee]

Some ideas floated in our meeting today:

• Reading on speech models
• Reading on hip-hop, lyrics, and language
• Reading on attention mechanisms in deep learning
• Workshopping figures or code ideas
• Progress updates on various projects

How about other topic ideas? Anything's fair game, let's start a discussion here.

[bmcfee]

@lostanlen and I are both interested in reading up on spectral graph theory for directed graphs. This might not be of sufficiently broad interest to the rest of the group, so we can do this solo unless folks are really interested.

[mcartwright]

Papers on attention, e.g. https://arxiv.org/pdf/1409.0473.pdf Other suggestions for attention papers? @pli1988 ?

[bmcfee]

Some more attention-related papers:

• Chiu and Raffel 17: https://arxiv.org/abs/1712.05382
• Mnih 14: http://papers.nips.cc/paper/5542-recurrent-models-of-visual-attention
• Gregor 15: https://arxiv.org/abs/1502.04623 (aka DRAW)
• Xu 15: http://proceedings.mlr.press/v37/xuc15.html (show-attend-tell)
• Arandjelovic 17: https://arxiv.org/abs/1705.08168 (look-listen-learn)

[tomxi]

A tutorial/workshop/overview of the applications of Online Learning or Reinforcement Learning in common MIR tasks?

[bmcfee]

@tomxi sounds interesting -- OL and RL are pretty different though, do you have a preference or something more specific in mind?

[tomxi]

I don't have a preference really, specific examples I'm thinking of: Online Learning: How to make my guitar transcriber better if I'm willing to sit there and "baby sit" it by giving immediate feedback/corrections after each "prediction" of some scope. How to frame this problem, etc.

Reinforcement Learning: Still hex-guitar transcriber example; if I'm already applying online learning, how does the computer pick what to annotate and get feedback next to gain the maximal amount of information?

The above two would extend to crowdsourcing stuff?

These examples are obviously very specific... but some more easy examples so we can see how these things work with MIR?

[pli1988]

Here are some attention that I liked:

• I really like this survey. It starts with seq2seq first which I think is important. https://arxiv.org/pdf/1507.01053.pdf
• This blog covers a bunch of different flavors of attention and is easy to read. http://ruder.io/deep-learning-nlp-best-practices/index.html#attention
• Attention is all you need. https://arxiv.org/pdf/1706.03762.pdf

[lostanlen]

I have finished the draft of a 13-page journal paper (IEEE TASLP) with Joakim Andén (Simons Foundation) and Stéphane Mallat (ENS / Collège de France) on time-frequency scattering. We have new results on audio texture synthesis, speech recognition, urban sound classification, acoustic scene classification, and musical instrument classification. There is almost no machine learning in it though, as it's mostly time-frequency analysis, so I'm not sure if it's within the scope of the reading group. Would you be interested in reading it later this month?

[rowe0002]

I'd be interested in discussing this roadmap: http://openaccess.city.ac.uk/2763/1/MIRES_Roadmap_ver_1.0.0.pdf and then the broader issue of how to increase the usability of MIR tools for musicologists and musicians in general

[mcartwright]

I haven't tried it yet, but I'm interested in using this model in my work. Could be interesting to others and worth discussing: http://papers.nips.cc/paper/6310-phased-lstm-accelerating-recurrent-network-training-for-long-or-event-based-sequences.pdf

[bmcfee]

Jumping back on this after the 6/12 meeting, some ideas relating to non-local means and kernel additive modeling:

• Fitzgerald et al. Kernel additive models for source separation. Gives a general framework for a bunch of related methods (HPSS, repet-sim, etc).
• Taemin Cho's dissertation uses non-local means as a way of smoothing chroma vectors in chord recognition. Matthias Mauch had a similar idea in chordino/segmentino, though I'm not sure it got written up anywhere.
• Other suggestions? Robust PCA can also be shoe-horned into this framework, and that gets us to vocal separation.

The general theme is to model the representation as a combination of "simple"/"predictable" and "sparse"/"random" signals. In the case of rhythm and microtiming (@magdalenafuentes), these sparse/random deviations from the predictable signal should be indicative of expressive timing. There's probably some NMF-related work we could dig into here as well.

[lostanlen]

Hi everyone, I'm re-upping the idea of having a session on time-frequency scattering (see my comment above). I wrote to my coauthor Joakim Andén (Flatiron Institute) and he'd be OK to give a talk at MARL on a Tuesday morning in late July: 7/16 or 7/23. I'll make sure to circulate the paper before the talk. How does this sound?

[lostanlen]

Can we decide now whether it's July 16th or July 23rd for Joakim Andén, so that I can forward the invitation to him?

[bmcfee]

My vote is for 7/16, but only because I'll be away on the 23rd.

[lostanlen]

Joakim Andén just confirmed his presence on Tuesday 16th. I'll send out the paper tomorrow on the mir-marl NYU mailing list.

[justinsalamon]

@lostanlen you mean Tuesday 17th, right? Note that Juan's proposing to move the meetings to CUSP for the summer, starting next week.

[lostanlen]

Yes, Tuesday 17th. My bad.