Weekend Reading Round-Up (Nov. '17)
Ian Hellström | 17 November 2017 | 3 min read
How easy is it to deceive a deep neural network? Does gender of leaders affect team cohesion? Can music be classified by looking at entropy alone?
DeepXplore: Automated Whitebox Testing of Deep Learning Systems
Voted the best paper at SOSP 2017, DeepXplore is not about adversarial attacks. Instead it’s about differential testing, that is, comparing different systems and looking at how what they spit out differs. DeepXplore requires a fully trained model and a single second on a commodity laptop to identify corner cases that are indicative of bugs inside the neural network. As the authors note, this is especially important for safety- or security critical systems, especially since the occlusion by a small rectangle or specs of dirt have been shown to cause neural networks to mislabel images. Related research has recently demonstrated that even though the same label may be assigned to, say, similar images, the reasoning behind these classifications may well be very different, even when done by the same neural network. Both the prediction and interpretation of deep neural networks can therefore be iffy.
One-Pixel Attack for Fooling Deep Neural Networks
This arXiv preprint describes a single-pixel attack that can be used to throw off deep neural networks for image recognition. This is of course important work to ensure the stability of such networks. The authors uses differential evolution, a gradient-less meta-heuristic, to obtain the minimal perturbations to the original images that cause these to be misclassified.
Best Ever Algorithm Found for Huge Streams of Data
An article posted in Quanta Magazine discusses recent research into streaming algorithms for the Big Data era. The key to the current best-in-class streaming algorithm for the frequent-items problem is to break up large objects (i.e. lists of items or rather their numerical encoding) using a divide-and-conquer strategy, but including tags so that the split chunks can be pieced back together correctly. In fact, the researchers use expander graphs to ensure that links between multiple chunks are maintained even when an individual link between two chunks is severed. This ensures the correct recovery of the data from its constituents.
Leader Evaluation and Team Cohesiveness in the Process of Team Development: A Matter of Gender?
This research published in PLOS ONE looks at how (binary) gender stereotypes affect team cohesion. The researchers observed 45 small teams, some led by women, and others headed by men, for the duration of a nine-months’ engineering project. While they noticed differences in self- and team evaluations of the leaders between the beginning and the end of the project, there were no significant differences between the sexes at the end of the project. For the cohesion they studied closure in social networks, that is, the existence of triangles in the interactions graph; no differences between genders were visible. The paper’s authors did find that teams led by women tend to cluster more around the leader in the team development phase, which, they believe, may aid groups with conflicts or communication issues.
Music Viewed by Its Entropy Content: A Novel Window for Comparative Analysis
Published in PLOS ONE, the article focuses on the question whether it’s possible to identify musical genres from entropy alone. The authors grab 450 MIDI files and come up with a set of symbols for each piece with the lowest possible entropy, purely by looking at the blob of text inside each file, that is, without knowing anything about the way the sound is encoded. Based on mostly visual inspections and basic statistics to validate their hypotheses, they are able to identify types and styles of MIDI-synthesized music.