Martin Thoma A blog about Code, Politics, and my life

ML Review 2

· by

Contents

This Review gives an overview of interesting stuff I stumbled over which is related to machine learning. Most of it was posted in KIT's machine learning group (on Facebook).

A lot of stuff can be found in my article about NIPS 2016.

New Developments

Live Demos and Websites

Quickdraw

Quickdraw is a program which tries to guess what you drew. However, it is difficult to check if they really apply machine learning, because it tells you what to draw and then tries to recognize it.

Quick, Draw! result: six doodles, of which the neural net recognized the chandelier, the suitcase and the candle
I had to draw a piano, a floor lamp, a chandelier, a suitcase, a candle and a lipstick each in under 20 seconds.

It looks very much like an attempt to get lots of training data. However, this plan might not work that well: Interesting Quickdraw Fails

You might find more stuff like Quickdraw on aiexperiments.withgoogle.com.

Loss Functions

lossfunctions.tumblr.com is a blog created by Andrej Karpathy where he collects - well, let's call them "interesting" - loss functions.

Eyescream

Have you heard about GANs?

Eyescream is a demo for the generator.

Publications

Deep Neural Networks are Easily Fooled

The input of CNNs for image classification can be manipulated in two ways:

  1. An image, on which a human does not recognize anything (e.g. white noise) gets a high score for some object class.
  2. An image on which a human is certain to recognize one class (e.g. "cat") is manipulated in a way that the CNN classifies with high certainty something different (e.g. "factory").

See also:

Breaking Linear Classifiers on ImageNet

Andrej Karpathy has once again written a nice article. The article describes the problem that linear classifiers can be broken easily.

Hinton commented something similar on Reddit.

Where am I?

Google Unveils Neural Network with “Superhuman” Ability to Determine the Location of Almost Any Image

One gives the neural network a photo and it tells you where it was taken.

LIME

"Why Should I Trust You?": Explaining the Predictions of Any Classifier deals with the problem of analyzing black box models' decision making process.

Lip Reading

See the paper LipNet: Sentence-Level Lipreading for details.

More

Software

Seaborn

Seaborn hexbin plot of two variables with their histograms on the margins
Example plot generated by Seaborn

Seaborn is a Python package for the visualization of data and statistics.

See stanford.edu/~mwaskom/software/seaborn.

RecNet

Jörg made recnet publicly available. It is a framework based on Theano to simplify the creation of recurrent networks.

Image Segmentation Using DIGITS 5

I haven't tried it yet, but the images in the article Image Segmentation Using DIGITS 5 look awesome. I would be happy to hear what you think about it.

Keras.js

Run Keras models (trained using Tensorflow backend) in your browser, with GPU support. Models are created directly from the Keras JSON-format configuration file, using weights serialized directly from the corresponding HDF5 file.

See github.com/transcranial/keras-js for more.

Interesting Questions

Miscellaneous

Meetings