Skip to content

Instantly share code, notes, and snippets.

View samuelgoto's full-sized avatar
💭
hello world!

Sam Goto samuelgoto

💭
hello world!
View GitHub Profile
<html>
<head>
<link href="http://fonts.googleapis.com/css?family=Source+Sans+Pro:400,600,700,900,400italic|Droid+Serif:400,700,400italic" media="screen" rel="stylesheet" type="text/css">
<style>
body {
color: #333333;
font-family: "Droid Serif", Georgia, serif;
-webkit-font-smoothing: antialiased;
margin: 0 auto;
max-width: 900px;

Introduction

The scientific paper format was invented in the context of printed media.

This is an exploration of a format for scientific papers optimized to be distributed over computer networks.

Problem

The printed media imposes the following constraints:

Things that could be interesting to control/censor:

Things that "use strict" control:

  • web assembly will always have a 10-12% interpretation cost (anecdotal data from @bnelson, need to sanity check / validate with others), which for computer vision things over a came stream seems like a non starter.
  • composition, portability
  • ergonomics (don't have to care about gpu)
  • cpu vs gpu: memory constrained versus computing contrained
  • concurrency, multithreading: don't you need to transfer video streams back and fourth between worker threads?

Deep Learning Processing Units

10-15% cost of interpreation seems like a non-starter to me when they are trying to squeeze in every level of performance. I'm seeing some convergence in the industry regarding "inference models" file formats / representations

I would challenge that assertion. Your going to get a larger drop on an android phone when thermal throttling kicks in.

Prior Art

Alternatives considered

TODO(goto): go over this.

ARML is an XML-based data format to describe an interact with AR scenes.

ARML

ATOM feeds in JSON-LD.

{
  "@context": "http://www.w3.org/2005/Atom",
  "@type": "Feed",
  "title": "Example Feed",
  "subtitle": "A subtitle.",
  "links": [{