July link roundup.

Incredible followup to Belsky et al 2016: "Genetic analysis of social-class mobility in five longitudinal studies", Belsky et al 2018a: http://www.pnas.org/content/early/2018/07/03/1801238115

"*Significance*: Genome-wide association study (GWAS) discoveries about educational attainment have raised questions about the meaning of the genetics of success. These discoveries could offer clues about biological mechanisms or, because children inherit genetics and social class from parents, education-linked genetics could be spurious correlates of socially transmitted advantages. To distinguish between these hypotheses, we studied social mobility in five cohorts from three countries. We found that people with more education-linked genetics were more successful compared with parents and siblings. We also found mothers’ education-linked genetics predicted their children’s attainment over and above the children’s own genetics, indicating an environmentally mediated genetic effect. Findings reject pure social-transmission explanations of education GWAS discoveries. Instead, genetics influences attainment directly through social mobility and indirectly through family environments.

Abstract: A summary genetic measure, called a “polygenic score,” derived from a genome-wide association study (GWAS) of education can modestly predict a person’s educational and economic success. This prediction could signal a biological mechanism: Education-linked genetics could encode characteristics that help people get ahead in life. Alternatively, prediction could reflect social history: People from well-off families might stay well-off for social reasons, and these families might also look alike genetically. A key test to distinguish biological mechanism from social history is if people with higher education polygenic scores tend to climb the social ladder beyond their parents’ position. Upward mobility would indicate education-linked genetics encodes characteristics that foster success. We tested if education-linked polygenic scores predicted social mobility in >20,000 individuals in five longitudinal studies in the United States, Britain, and New Zealand. Participants with higher polygenic scores achieved more education and career success and accumulated more wealth. However, they also tended to come from better-off families. In the key test, participants with higher polygenic scores tended to be upwardly mobile compared with their parents. Moreover, in sibling-difference analysis, the sibling with the higher polygenic score was more upwardly mobile. Thus, education GWAS discoveries are not mere correlates of privilege; they influence social mobility within a life. Additional analyses revealed that a mother’s polygenic score predicted her child’s attainment over and above the child’s own polygenic score, suggesting parents’ genetics can also affect their children’s attainment through environmental pathways. Education GWAS discoveries affect socioeconomic attainment through influence on individuals’ family-of-origin environments and their social mobility.

New SOTA for lipreading: "Large-Scale Visual Speech Recognition", Shillingford et al 2018: https://arxiv.org/abs/1807.05162

"This work presents a scalable solution to open-vocabulary visual speech recognition. To achieve this, we constructed the largest existing visual speech recognition dataset, consisting of pairs of text and video clips of faces speaking (3,886 hours of video). In tandem, we designed and trained an integrated lipreading system, consisting of a video processing pipeline that maps raw video to stable videos of lips and sequences of phonemes, a scalable deep neural network that maps the lip videos to sequences of phoneme distributions, and a production-level speech decoder that outputs sequences of words. The proposed system achieves a word error rate (WER) of 40.9% as measured on a held-out set. In comparison, professional lipreaders achieve either 86.4% or 92.9% WER on the same dataset when having access to additional types of contextual information. Our approach significantly improves on other lipreading approaches, including variants of LipNet and of Watch, Attend, and Spell (WAS), which are only capable of 89.8% and 76.8% WER respectively."

"A comprehensive meta-analysis of the predictive validity of the Graduate Record Examinations: Implications for graduate student selection and performance", Kuncel et al 2001: https://pdfs.semanticscholar.org/5c69/875be977c1de9392dd8392937ebbf606dd18.pdf

"This meta-analysis examined the validity of the Graduate Record Examinations (GRE) and undergraduate grade point average (UGPA) as predictors of graduate school performance. The study included samples from multiple disciplines, considered different criterion measures, and corrected for statistical artifacts. Data from 1,753 independent samples were included in the meta-analysis, yielding 6,589 correlations for 8 different criteria and 82,659 graduate students. The results indicated that the GRE and UGPA are generalizably valid predictors of graduate grade point average, 1st-year graduate grade point average, comprehensive examination scores, publication citation counts, and faculty ratings. GRE correlations with degree attainment and research productivity were consistently positive; however, some lower 90% credibility intervals included 0. Subject Tests tended to be better predictors than the Verbal, Quantitative, and Analytical tests."

"Genetics & the Geography of Health, Behavior, and Attainment", Belsky et al 2018b: https://www.biorxiv.org/content/early/2018/07/25/376897

"People's life chances can be predicted by their neighborhoods. This observation is driving efforts to improve lives by changing neighborhoods. Some neighborhood effects may be causal, supporting neighborhood-level interventions. Other neighborhood effects may reflect selection of families with different characteristics into different neighborhoods, supporting interventions that target families/individuals directly. To test how selection affects different neighborhood-linked problems, we linked neighborhood data with genetic, health, and social-outcome data for >7,000 European-descent UK and US young people in the E-Risk and Add Health Studies. We tested selection/concentration of genetic risks for obesity, schizophrenia, teen-pregnancy, and poor educational outcomes in high-risk neighborhoods, including genetic analysis of neighborhood mobility. Findings argue against genetic selection/concentration as an explanation for neighborhood gradients in obesity and mental-health problems, suggesting neighborhoods may be causal. In contrast, modest genetic selection/concentration was evident for teen-pregnancy and poor educational outcomes, suggesting neighborhood effects for these outcomes should be interpreted with care.
...Children in the least-disadvantaged neighborhoods and those in the most-disadvantaged neighborhoods had polygenic scores for educational attainment that were on average about 0.4 standard-deviations apart. This genetic difference corresponded to a predicted 4% increase in risk for poor educational qualifications and a 1.5% increase in risk for NEET status."

Sariaslan notes that while presented as contradicting earlier results on schizophrenia, it may well be underpowered to prove the absence of drifts.

"The Origins of WEIRD Psychology", Schulz et al 2018: https://psyarxiv.com/d6qhu/

"Recent research not only confirms the existence of substantial psychological variation around the globe but also highlights the peculiarity of populations that are Western, Educated, Industrialized, Rich and Democratic (WEIRD). We propose that much of this variation arose as people psychologically adapted to differing kin-based institutions — the set of social norms governing descent, marriage, residence and related domains. We further propose that part of the variation in these institutions arose historically from the Catholic Church’s marriage and family policies, which contributed to the dissolution of Europe’s traditional kin-based institutions, leading eventually to the predominance of nuclear families and impersonal institutions. By combining data on 20 psychological outcomes with historical measures of both kinship and Church exposure, we find support for these ideas in a comprehensive array of analyses across countries, among European regions and between individuals with different cultural backgrounds."

Whole genomes hit a new low: $350.

"Immortals guide history by choosing players for “special roles”—nobles, templar-priests, powerful merchants, military personnel, and the like, who are loosely overseen by immortal-“animated” NPC superiors—and by moderating HRPTs, highly recommended role-playing times: festivals, battles, catastrophes, riots, and other large-scale events that make a sizable impact on the game-world.11 They also act for the silent majority of “virtual” Zalanthans, who are represented neither by players nor by scripted NPCs.12 In a war, staff might be called upon to rewrite the rooms of a city after a general sets fire to its market quarter, or to rule on what portion of the virtual populace expires from hunger in a siege. At a bardic competition, immortals might act as the audience, improvising cheers and jeers from bystanders. No situation is too large or too small for their intercession.

Games are infamous for their invisible walls, boundaries where movement stops and a distant landscape reveals itself to be nothing more than wallpaper. These obstacles are a synecdoche for the larger limitations of the genre, which thrives on creating an illusory sense of completeness.13 The immortals lift Zalanthas above this condition by acting (in a way that is only truly possible for text-based games) as “gods of the gaps,” improvising the terra incognita beyond the world’s edge.

No player-characters live in the remote village of Cenyr, thinly glossed by Armageddon’s publicly available documentation. But when one of my characters arrived there after a risky journey through the Red Desert, the NPCs in the marketplace began to speak. I asked questions and learned an incredible wealth of information: the date of the town’s foundation, the rituals of the local wind cult, and the qualities of Cenyri music—free use of syncopation, rudimentary phrasing, a diatonic scale with blue fourth and fifth notes. What I had thought to be a small collection of rooms, fewer than a dozen among more than thirty thousand, turned out to contain an encyclopedic wealth of background—secret lore that might have been written by the immortals, lived out by players in prior years and recorded, or written on the spot by a staff member for me. Part of Armageddon’s seeming limitlessness is the impossibility of telling the difference between the three.

Armageddon’s immortals imbue Zalanthas with a marvelous plasticity, an alternative to “procedural generation” that is creative and human rather than algorithmic. Their primary role, however, is not revealing the world to players, but enabling them to create it, in both its largest and smallest aspects. Of countless objects in the Armageddon universe, some few thousand are “mastercrafts,” designed by characters who after attaining a certain level of skill (and, of course, roleplaying that attainment, emoting each painstaking step of whittling, flint knapping, chiseling, or leather-curing) are allowed to make their permanent mark on the game.14 In three years of playing, I created one: a carved ivory pipe. Like most special items, it exists primarily as a simple string, difficult to tell from any other equipment in a busy shop’s inventory or the description of a room. But a further command, look pipe, reveals it at a deeper level of granularity:

> A beautifully polished ivory horn has been carved smoothly into a slender pipe with a wide, deep bowl. Just over a cord in length, it is long and thin from the mouthpiece, curving gently downwards and growing thicker towards the base. At bottom, the ivory widens suddenly and curves up into a wide, deep bowl. Etched across the shaft is a rough and impressionistic map of Zalanthas, from Gol Krathu to Vrun Driath, backgrounded by a repeating pattern of dunes. Subtle and detailed carvings of caravans, gith, argosies, and other figures are set against the material, as well as square, fingernail-sized etchings of all the major settlements of the Known World. At the very middle of the pipestem sits a small image of Luir’s Outpost, complete with horn-crested walls. A dark, nearly black sapphire, tumbled to a rough polish, is set above the gates. A leather strap allows the pipe to be worn around the neck.

Seti, my merchant-carver, carried this pipe with him for months. For a while it was his signature accessory, the peacock feather he used to accentuate his presence and a distillation of the world as he knew it. But ultimately I gave it away, not knowing where it would end up—my errant contribution to the mosaic of Zalanthas."

PPO-LSTM+domain-randomization in MuJuCo/Unity for sim2real transfer in a robotic hand grasper: Dactyl, "Learning Dexterity" (OpenAI):

Videos: https://youtu.be/jwSbzNHGflM https://www.youtube.com/watch?v=DKe8FumoD4E
Paper: "Learning Dexterous In-Hand Manipulation", OpenAI 2018: https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/learning-dexterity/learning-dexterity-paper.pdf

"We use reinforcement learning (RL) to learn dexterous in-hand manipulation policies which can perform vision-based object reorientation on a physical Shadow Dexterous Hand. The training is performed in a simulated environment in which we randomize many of the physical properties of the system like friction coefficients and an object’s appearance. Our policies transfer to the physical robot despite being trained entirely in simulation. Our method does not rely on any human demonstrations, but many behaviors found in human manipulation emerge naturally, including finger gaiting, multi-finger coordination, and the controlled use of gravity. Our results were obtained using the same distributed RL system that was used to train OpenAI Five [43]. We also include a video of our results: https://youtu.be/jwSbzNHGflM
...The vast majority of training time is spent making the policy robust to different physical dynamics. Learning to rotate an object in simulation without randomizations requires about 3 years of simulated experience, while achieving the same performance in a fully randomized simulation requires about 100 years of experience. This corresponds to a wall-clock time of around 1.5 hours and 50 hours in our simulation setup, respectively.
...our default setup with an 8 GPU optimizer and 6144 rollout CPU cores reaches 20 consecutive achieved goals approximately 5.5 times faster than a setup with a 1 GPU optimizer and 768 rollout cores. Furthermore, when using 16 GPUs we reach 40 consecutive achieved goals roughly 1.8 times faster than when using the default 8 GPU setup. Scaling up further results in diminishing returns, but it seems that scaling up to 16 GPUs and 12288 CPU cores gives close to linear speedup"

Nothing really new here but it is yet another demonstration of what brute force + systems engineering can do with existing DRL.