November 2016 News
This is the November 2016 edition of the Gwern.net newsletter; previous, October 2016 (archives). This is a collation of links and summary of major changes, overlapping with my Changelog; brought to you by my donors on Patreon.
If you have not been receiving issues, please check your email account’s spam folder. (Gmail in particular has been flagging as spam.) If you prefer to subscribe as an RSS feed, you may be interested in the new service “Kill the newsletter”.
The December issue will likely be late as I am traveling most of that month: I will be in Oxford, England 11–22 December for the Cambridge Conference on Catastrophic Risk (12–14 December) and FHI workshop on AI Safety and Blockchain (16–17 December), and in London 24–25 December.
Writings
-
switched Gwern.net to HTTPS
Media
Links
Genetics:
-
Recent Evolution:
-
dysgenics: “Genome-wide analysis identifies 12 loci influencing human reproductive behavior”, et al 2016 ( supplement; genetic correlations with fewer later offspring: rg = -0.236 and 0.712 respectively. Cross-sectional confirmation of et al 2016’s discovery of longitudinal dysgenics.)
-
“A time transect of exomes from a Native American population before and after European contact”, et al 2016 (Rapid soft selection sweeps on immune-system genes post-colonization. Presumably ancient genomes would show even more dramatic population collapses & selection in South/Central America.)
-
“Kinship, Fractionalization and Corruption”, et al 2019
-
-
Everything Is Heritable
-
“Genome-Wide Association Study Reveals First Locus for Anorexia Nervosa and Metabolic Correlations”, et al 2016 (see also “The Thin Gene”)
-
“Lung cancer, genetic predisposition and smoking: the Nordic Twin Study of Cancer”, et al 2016 ( R. A. Fisher would be pleased to see that someone’s finally checked.)
AI:
-
“Scaling Memory-Augmented Neural Networks with Sparse Reads and Writes”, et al 2016
-
“Neural architecture search with reinforcement learning”, 2016 (800 GPUs!)
-
“Designing Neural Network Architectures using Reinforcement Learning”, et al 2016
-
“Learning to reinforcement learn”, et al 2016
-
human-level lipreading (human-level lipreading + MSR’s transcription + Google’s WaveNet could have major implications for the deaf & hard-of-hearing and the role of sign language):
-
“Lip Reading Sentences in the Wild”, et al 2016 ( vide)
-
“LipNet: Sentence-level Lipreading”, et al 2016 ( video)
-
-
“Tracking the World State with Recurrent Entity Networks”, et al 2016
-
“DeepCoder: Learning to Write Programs”, et al 2016; “Learning to superoptimize programs”, et al 2016; DeepMath/ “HolStep: A Machine Learning Dataset for Higher-order Logic Theorem Proving”, et al 2016 (continuing the AlphaGo theme of tree search+NN heuristics; see also “Deep Learning for Real-Time Atari Game Play Using Offline Monte Carlo Tree Search [MCTS] Planning”, et al 2014)
-
“Reinforcement Learning with Unsupervised Auxiliary Tasks”, et al 2016
-
“Outrageously large neural networks: the sparsely-gated mixture-of-experts layer”, et al 2016 (This is cool. The main net is a little NN which chooses a couple out of a few thousand other NNs to do the real computation. Because only a couple subnets are activated at any time and the others are silent, they are effectively small specialized NNs—which accordingly fit in the GPU and are very fast forwards/backwards. Instead of the complications of learning to use an external memory, huge datasets can effectively be sharded & memorized by individual specialized sub NNs.)
-
“Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation”, et al 2016; “Found in translation: More accurate, fluent sentences in Google Translate” (the long-awaited high quality translation RNNs finally get rolled out to the masses)
-
“Image-to-Image Translation with Conditional Adversarial Nets”, et al 2016 (as has often been noted, mapping images to images is a very common type of operation, and there’s no real reason to have a whole menagerie of DCGANs for upscaling and colorizing and inpainting and…)
-
“DeepMind and Blizzard to release StarCraft II as an AI research environment”
-
“miles-deep: Deep Learning Porn Video Classifier/Editor with Caffe”
Statistics/Meta-Science:
Politics/religion:
Psychology/biology:
-
“Neural correlates of specific musical anhedonia”, Martínez-et al 2016
Technology:
Fiction:
-
“Design”, Robert Frost
Film/TV
Live-action:
-
_Five Element Ninjas (good stupid fun; each fight is more ridiculous than the last)
-
Arrival (see my review of Stories of Your Life and Others)
Music
Touhou:
-
“二人のエンドレス秘封活動” (MasamiT; ゴツまさ倶楽部 {C79}) [folk]
Doujin:
-
“美しくも醜いのなら” (君の音。; 少女の海から {M3-2016}) [classical]
-
“DEEP FOREST” (Kirin feat.Sarii Yoshida; Altar Fragments {C90}) [electronic]
-
“INFINITY” (Kokuchou feat. Misa Kikouden; Altar Fragments {C90}) [electronic]