Research overviews  ·  chapter 1 of 14  ·  1 min read

Interesting science

Some interesting papers in science and statistics.

About this chapter

A flat list of papers and links that are worth knowing about but don’t belong to any one topic. Many are from neuroscience and perception: Held & Hein’s moving kittens, decoding 50 face components from 206 neurons, semantic maps that tile the cortex, and the finding that languages trade information per syllable against speaking rate, so that every language ends up with a similar information rate.

The rest mix classic papers (Shannon 1948, Hopfield networks, BOIDS) with machine learning papers on generalization in deep nets, bag-of-local-features models on ImageNet and image synthesis with a single robust classifier. There are no sections; skim it.

• Moravec’s paradox

• Held & Hein 1963 - moving kittens

• Human attention vqa dataset (VQA HAT) ~ 60k images

• Understanding deep learning requires rethinking generalization (zhang et al. 2016)

• Mike the headless chicken

• thompson 96 evolved circuit

• The Code for Facial Identity in the Primate Brain - can decode 50 face components from 206 neurons

• Stimulating memory

• herculano-houzel talk

• TMS - changing people’s minds

• Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche

  • “Languages vary a lot in terms of the information that they pack into a syllable and also in the rate that they are spoken at. But the interesting thing is that the two kind of balance each other, so that more information dense languages are spoken slower, and those that are less informationally heavy are spoken faster. This means that there is a steady information rate that is very similar among languages

• BOIDS: cool simple algorithm for emergent behavior

• Image Synthesis with a Single (Robust) Classifier

• Shrew skulls shrink for winter survival

• Performance-optimized hierarchical models predict neural responses in higher visual cortex (yamins et al. 2014)

• Natural speech reveals the semantic maps that tile human cerebral cortex (huth et al. 2016)

• A Mathematical Theory of Communication (shannon 1948)

• hopfield networks

• benjamini-hoffberg - controlling decision rate

• infogan chen et. al 16

• visual deep reps ganguli 19

• rulefit friedman & popescu 19

• Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet (brendel & bethge 2019)

• statistical learning by 8-month-old infants 1996

• Unpaired Image-To-Image Translation Using Cycle-Consistent Adversarial Networks (zhu et al. 2017)

• cyclegan 17

• oregan 01 vision philosophy

• The Association Between Income and Life Expectancy in the United States, 2001-2014

  • 

    Fig. 149 Figure 1. Expected age at death for 40-year-olds against household income percentile in the US, for women and men. Life expectancy rises steeply across the lower incomes and flattens at the top, and women live longer than men at every income.

• aztecs, mayans predicted eclipses with data

  • Pigeons as Trainable Observers of Pathology and Radiology Breast Cancer Images

Key ideas

• The list has no sections; most entries are single links to results that are surprising or simple to state.

• Several are about perception and the brain, such as Held & Hein’s kittens, decoding faces from 206 neurons and semantic maps that tile the cortex.

• Languages trade information per syllable against speaking rate, so their information rates end up similar.

• The machine learning entries question how deep nets work, from rethinking generalization to bag-of-local-features models doing well on ImageNet.