Ozan Sener ozansener

NIPS Notes

Simple Intro by Blei mostly going over review paper of Jordan
Later introduce SVI (Stochastic VI) as a remedy to solve VI tractably with large dataset.
Review the black box inference -assumption free VI- http://www.jmlr.org/proceedings/papers/v33/ranganath14.pdf
Key idea is replacing gradient and the expectation in VI formulation. Since expectation reqiuires exponential family assumption to work replacing expectation and gradient solves this if overall method is stochastic since your samples are unbiased gradient estimates satisfying Robinson-Monroe conditions however the variance is very large and it requires even further tricks

I hereby claim:

I am ozansener on github.
I am ozansener (https://keybase.io/ozansener) on keybase.
I have a public key whose fingerprint is A427 ACAD 1AAC 0B74 DEB0 F20B 6394 749D 0843 4519

To claim this, I am signing this object:

	def projection2simplex(y):
	"""
	Given y, it solves argmin_z \|y-z\|_2 st \sum z = 1 , 1 >= z_i >= 0 for all i
	"""
	m = len(y)
	sorted_y = np.flip(np.sort(y), axis=0)
	tmpsum = 0.0
	tmax_f = (np.sum(y) - 1.0)/m
	for i in range(m-1):
	tmpsum+= sorted_y[i]

	#!/bin/bash

	# Bash script to install latest version of ffmpeg and its dependencies on Ubuntu 12.04 or 14.04
	# Inspired from https://gist.github.com/faleev/3435377

	# Remove any existing packages:
	sudo apt-get -y remove ffmpeg x264 libav-tools libvpx-dev libx264-dev

	# Get the dependencies (Ubuntu Server or headless users):
	sudo apt-get update

	import numpy

	def get_turn_off(alpha):
	cycle_cnt = (numpy.log(0.4) / numpy.log(1 - alpha) )
	return (cycle_cnt * 50)/ (1000 * 60)

	def get_alpha_from_cut_off(cut_off):
	alpha = 1 - numpy.exp(numpy.log(0.1) / ((float(cut_off) * 60 * 1000) / float(50)))
	return alpha