Oleg Zabluda's blog
Monday, June 18, 2018
 
The Batch Normalization layer of Keras is broken
The Batch Normalization layer of Keras is broken
http://blog.datumbox.com/the-batch-normalization-layer-of-keras-is-broken/
http://blog.datumbox.com/the-batch-normalization-layer-of-keras-is-broken/

Labels:


- posted by Oleg Zabluda @ 12:49 PM | permalink | | | technorati | bloglines | feedster
 
"""
"""
In 2014, 47 of the nation’s 73 inspectors general signed a letter alleging that Obama had stonewalled their “ability to conduct our work thoroughly, independently, and in a timely manner.”
[...]
In 2012, Horowitz recommended that 14 Justice Department and ATF officials be disciplined for their conduct in the “Fast and Furious” gun-walking scandal.

A 2013 IG audit found that the IRS had targeted conservative groups for special scrutiny prior to the 2012 Obama reelection effort.

In 2014, an internal audit revealed that CIA officials had hacked the Senate Intelligence Committee’s computers while compiling a report on enhanced interrogation techniques. CIA director John Brennan had claimed that his agents were not improperly monitoring Senate staff computer files. He was forced to retract his denials and apologize for his prevarication.

In 2016, the State Department’s inspector general found that Hillary Clinton had never sought approval for her reckless and illegal use of an unsecured private email server. The IG also found that other Clinton aides silenced staffers who were worried about national security being compromised by the unsecured server.
"""
https://www.nationalreview.com/2018/06/obama-administration-stonewalled-inspectors-general/
https://www.nationalreview.com/2018/06/obama-administration-stonewalled-inspectors-general/

Labels:


- posted by Oleg Zabluda @ 10:24 AM | permalink | | | technorati | bloglines | feedster
 
Interpretable Discovery in Large Image Data Sets
Interpretable Discovery in Large Image Data Sets
"""
Novelty detection:

clustering: look for observations that fall between clusters
Isolation forest: build a RF, then look for examples that split off early from the main data mas
density-based (e.g., local outlier flow)
Use SVD to create a low-dimensional feature space, project your data down to that space and back, and anything you failed to reconstruct is an anomaly that your model didn’t capture
DEMUD: SVD-based + explanations

Especially good for discovering rare classes. "Explanations justify selections."
"""
https://vimeo.com/252188399
https://vimeo.com/252188399

Labels:


- posted by Oleg Zabluda @ 9:29 AM | permalink | | | technorati | bloglines | feedster

Powered by Blogger