November 27, 2016: Business Meeting

Revision 43193

Date
Author
SolSeed archive migration
Comment
Create maintainable Markdown archive revision
Size
16270 bytes
Delta
+74 bytes

Sunday Agenda

Check-in

Opening Blessing

As we live in the world, there are times when many things are upsetting. May we remember that even in those times, we have much to be thankful for. May we find a spirit of gratitude in our hearts. --Blessed Be

WW(W)s from last time

Weekly Events Logistics

Service rotation: Eric, Ben, Shelley, Brandon

Strategy Talk

Six Thinking Hats

Red Hatting

Work on ways to preserve good things about civilization through a climate-disrupted dark age

Deepening our practices (internal audience)

Recruiting (external audience)

Place Making (largely internal audience)

Art (internal and external audience)

Science fiction (external audience)

Sharing our ideas (external audience)

Research (external audience)

Practices maintenance

Winter Solstice

Concerning Speech

The Concerning Speech project is meant to provide people who are interested in machine learning/AI (just Brandon to start with) with a problem that is tractable and illustrative of the space of machine learning problems. In particular, it begins to create the sense of what the Curse of Dimensionality is, and how/why feature engineering is still a black art. So Brandon wants to create visualizations that people can play with to tweak parameters and try different features in order to get their feet wet easily.

For example, a speech signal in its native form is a time-varying magnitude of the displacement of the microphone membrane. So to train, you can record a couple seconds of speech and end up with a point in a several-thousand-dimensional space, where each audio sample is one dimension. That would be a feature taken over a 2-second time window at the 20-KHz sampling frequency. That is the most naive feature you can think of, practically, and immediately runs into all sorts of problems.

The next sort of feature that people turn to is energy at various frequencies. This is decomposing the signal using a Fourier transform approach and determining the weights of various frequency bands, with a much lower sampling granularity, maybe 10 samples per second instead of 10,000. So now you have 20 dimensions over a 2-second window.

In order to make it as accessible as possible to the most people, I want the playground/worksheet to be in Web standards so that people can immediately interact with it in their browser. Examples of similar workspaces include CodePin.io, JSFiddle, etc. To visualize the data, we will use the D3 graphics library. To interact with the audio, we will use Aurora.js, which is a Node and browser package for loading, decoding, and playing audio files. There is already a good start on creating a dataset that is called Librispeech, which is using the Librivox audio recordings of public-domain works (books, plays, poems, etc). I want to tie Concerning Speech directly to that as well.

The reason it's called Concerning Speech is because it's going to be launched on the Concerning AI podcast. I'm not concerned that this project could get out of hand and produce an Artificial General Intelligence. But I am concerned that it might launch the career of the person who instigates everything getting out of hand.

Evaluation

Closing Blessing

The world we live in is not always easy to live in. The work we want to do in the world often seems impossible in the face of what we know about the world. But we can be grateful that we have each other and can be together in our highest aspirations and support each other in working to make them real. -- Blessed Be