Connectionism, symbolic AI, and machine learning in AI

Where do the predictive techniques of machine learning come from? How do they represent and organise the social? Join us for a public talk with’s Dominique Cardon (@Karmacoma), Wed 24th April 5pm: https://t.co/oBplRQvRgZ

— Digital Humanities (@kingsdh) April 19, 2019

If you follow the link you'll see this abstract:

Neurons spike back. The invention of inductive machine and the Artificial intelligence controversy - Dominique Cardon (Sciences Po Médialab)

Since 2010, machine learning based predictive techniques, and more specifically deep learning neural networks, have achieved spectacular performances in the fields of image recognition or automatic translation, under the umbrella term of “Artificial Intelligence”. But their relation to this field of research is not straightforward. In the tumultuous history of AI, learning techniques using so-called "connectionist" neural networks have long been mocked and ostracized by the "symbolic" movement. This talk retraces the history of artificial intelligence through the lens of the tension between symbolic and connectionist approaches. From a social history of science and technology perspective, it seeks to highlight how researchers, relying on the availability of massive data and the multiplication of computing power have undertaken to reformulate the symbolic AI project by reviving the spirit of adaptive and inductive machines dating back from the era of cybernetics.

The hypothesis behind this communication is that the new computational techniques used in machine learning provide a new way of representing society, no longer based on categories but on individual traces of behaviour. The new algorithms of machine learning replace the regularity of constant causes with the "probability of causes". It is therefore another way of representing society and the uncertainties of action that is emerging. To defend this argument, this communication will propose two parallel investigations. The first, from a science and technology history perspective, traces the emergence of the connexionist paradigm within artificial intelligence techniques. The second, based on the sociology of statistical categorization, focuses on how the calculation techniques used by major web services produce predictive recommendations.

This talk will be partly based on the article (in French): Cardon (Dominique), Cointet (Jean-Philippe), Mazières (Antoine), «La revanche des neurones. L’invention des machines inductives et la controverse de l’intelligence artificielle», Réseaux, n°211, 2018, pp. 173-220.

Wall Street's looking for peole who know about culture

Wall Street Is Desperate for Wonks Who Can Explain the Rise of Populism https://t.co/Qfp13tmLW4

— Charlie Tapp (@charleswtapp) June 5, 2019

Untouchability and open defecation in rural India

Dean Spears and Amit Thorat, "The Puzzle of Open Defecation in Rural India: Evidence from a Novel Measure of Caste Attitudes in a Nationally Representative Survey," Economic Development and Cultural Change 0, no. 0 (-Not available-): 000. https://doi.org/10.1086/698852

Abstract

Uniquely widespread and persistent open defecation in rural India has emerged as an important policy challenge and puzzle about behavioral choice in economic development. One candidate explanation is the culture of purity and pollution that reinforces and has its origins in the caste system. Although such a cultural account is inherently difficult to quantitatively test, we provide support for this explanation by comparing open defecation rates across places in India where untouchability is more and less intensely practiced. In particular, we exploit a novel question in the 2012 India Human Development Survey that asked households whether they practice untouchability, meaning whether they enforce norms of purity and pollution in their interactions with lower castes. We find an association between local practice of untouchability and open defecation that is robust; is not explained by economic, educational, or other observable differences; and is specific to open defecation rather than other health behavior or human capital investments more generally. We verify that practicing untouchability is not associated with general disadvantage in health knowledge or access to medical professionals. We interpret this as evidence that the culture of purity, pollution, untouchability, and caste contributes to the exceptional prevalence of open defecation in rural India.

H/t Tyler Cowen.

A tale of two cities, among other things

Why do people die in novels?

Olivier Morin, Alberto Acerbi, Oleg Sobchuk, Why people die in novels: testing the ordeal simulation hypothesis, Palgrave Communications 5, Article number: 2 (2019)

Abstract

What is fiction about, and what is it good for? An influential family of theories sees fiction as rooted in adaptive simulation mechanisms. In this view, our propensity to create and enjoy narrative fictions was selected and maintained due to the training that we get from mentally simulating situations relevant to our survival and reproduction. We put forward and test a precise version of this claim, the “ordeal simulation hypothesis”. It states that fictional narrative primarily simulates “ordeals”: situations where a person’s reaction might dramatically improve or decrease her fitness, such as deadly aggressions, or decisions on long-term matrimonial commitments. Experience does not prepare us well for these rare, high-stakes occasions, in contrast with situations that are just as fitness-relevant but more frequent (e.g., exposure to pathogens). We study mortality in fictional and non-fictional texts as a partial test for this view. Based on an analysis of 744 extensive summaries of twentieth century American novels of various genres, we show that the odds of dying (in a given year) are vastly exaggerated in fiction compared to reality, but specifically more exaggerated for homicides as compared to suicides, accidents, war-related, or natural deaths. This evidence supports the ordeal simulation hypothesis but is also compatible with other accounts. For a more specific test, we look for indications that this focus on death, and in particular on death caused by an agent, is specific to narrative fiction as distinct from other verbal productions. In a comparison of 10,810 private letters and personal diary entries written by American women, with a set of 811 novels (also written by American women), we measure the occurrence of words related to natural death or agentive death. Private letters and diaries are as likely, or more likely, to use words relating to natural or agentive death. Novels written for an adult audience contain more words relating to natural deaths than do letters (though not diary entries), but this is not true for agentive death. Violent death, in spite of its clear appeal for fiction, does not necessarily provide a clear demarcation point between fictional and non-fictional content.

Comment: FWIW, or various reasons, color me skeptical. For one thing, the adaptive hypothesis implies that death shows up in fiction because it is something we all must confront, but that we can't prepare for by rehearsal. But is homicide more prevalent in 20th C. American novels because Americans want to prepare themselves for the threat of murder? Seems unlikely to me. I rather suspect that murder shows up because of the moral and psychological issues it raises about the murderer. Fans of The Sopranos, for example, weren't preparing for the possibility of being murdered by a mob boss. They're interested in how a mob boss thinks and feels about the murders he orders and the ones he commits.

Addendum: 9.20.19: Moreover I don't see how simply reading about this or that violent death at the hands of someone else can provide any simulation that would be of value in an actual physical confrontation. Fighters don't prepare for a match by reading about matches, or even watching film – though they may do some of that. Their primary preparation takes the form of sparing, actual physical practice. Reading about fights does nothing to train your physical actions and reactions.

Extreme counterpoint

#BachUpsideDown #1: Here’s Bach’s 1st Goldberg Variation, normal, then upside down! Been meaning to do this for a while. My computer‘s doing a chromatic inversion of what I just played, so the intervals are preserved. You get incredible harmonic movement this way. Whoo! pic.twitter.com/8Ul9jNZjn8

— Dan Tepfer (@tepferdan) May 31, 2019

Kids will sing and dance

why is this video of a sea of elementary school kids going HAM to old town road literally the funniest video i’ve ever seen in my whole life pic.twitter.com/iXItFq50QG

— ari ❿ (@wasteIandbaby) May 30, 2019

Flower, with car in background

Farmers work harder than hunter-gathers

Farmers have less leisure time than hunter-gatherers, ScienceDaily, May 21, 2019

Hunter-gatherers in the Philippines who adopt farming work around ten hours a week longer than their forager neighbours, a new study suggests, complicating the idea that agriculture represents progress. The research also shows that a shift to agriculture impacts most on the lives of women.

For two years, a team including University of Cambridge anthropologist Dr Mark Dyble, lived with the Agta, a population of small scale hunter-gatherers from the northern Philippines who are increasingly engaging in agriculture.

Every day, at regular intervals between 6am and 6pm, the researchers recorded what their hosts were doing and by repeating this in ten different communities, they calculated how 359 people divided their time between leisure, childcare, domestic chores and out-of-camp work. While some Agta communities engage exclusively in hunting and gathering, others divide their time between foraging and rice farming.

The study, published today in Nature Human Behaviour, reveals that increased engagement in farming and other non-foraging work resulted in the Agta working harder and losing leisure time. On average, the team estimate that Agta engaged primarily in farming work around 30 hours per week while foragers only do so for 20 hours. They found that this dramatic difference was largely due to women being drawn away from domestic activities to working in the fields. The study found that women living in the communities most involved in farming had half as much leisure time as those in communities which only foraged.

The statistical mechanics of musical harmony

Jesse Berezovsky, The structure of musical harmony as an ordered phase of sound: A statistical mechanics approach to music theory, Science Advances 17 May 2019: Vol. 5, no. 5, eaav8490 DOI: 10.1126/sciadv.aav8490

Abstract

Music, while allowing nearly unlimited creative expression, almost always conforms to a set of rigid rules at a fundamental level. The description and study of these rules, and the ordered structures that arise from them, is the basis of the field of music theory. Here, I present a theoretical formalism that aims to explain why basic ordered patterns emerge in music, using the same statistical mechanics framework that describes emergent order across phase transitions in physical systems. I first apply the mean field approximation to demonstrate that phase transitions occur in this model from disordered sound to discrete sets of pitches, including the 12-fold octave division used in Western music. Beyond the mean field model, I use numerical simulation to uncover emergent structures of musical harmony. These results provide a new lens through which to view the fundamental structures of music and to discover new musical ideas to explore.

Introduction

The ubiquity of music throughout history and across cultures raises a fundamental question: Why is this way of arranging sounds such a powerful medium for human artistic expression? Although there are myriad musical systems and styles, certain characteristics are nearly universal, including emergent symmetries such as a restriction to a discrete set of sound frequencies (pitches). Historically, the theory of music has followed an empirical top-down approach: Patterns are observed in music and generalized into theories. Recent work has aimed to generalize these generalized theories to uncover new potential patterns that can lead to new theories of music (1–3). Here, instead, we observe patterns that emerge naturally from a bottom-up theory. We start from two basic (and conflicting) principles: A system of music is most effective when it (i) minimizes dissonant sounds and (ii) allows sufficient complexity to allow the desired artistic expression. Mathematical statement of these principles allows a direct mapping onto a standard statistical mechanics framework. We can thereby apply the tools of statistical mechanics to explore the phenomena that emerge from this model of music. Just as in physical systems where ordered phases with lower symmetry (e.g., crystals) emerge across transitions from higher-symmetry disordered phases (e.g., liquids), we observe ordered phases of music self-organizing from disordered sound. These ordered phases can replicate elements of traditional Western and non-Western systems of music, as well as suggesting new directions to be explored.

The basis for a bottom-up approach was provided by discoveries in the field of psychoacoustics originating with Helmholtz (4) and further developed in the 20th century, which established a quantitative understanding of how sound is perceived. This leads to the idea that the structure of music is related to a minimization of dissonance D, as explored by Plomp and Levelt (5), Sethares (6, 7), and others. Minimization of D cannot be the only criterion for an effective musical system, however, or we would all listen to “music” composed from just a single pitch. Instead, an effective system of music must have some degree of complexity to provide a sufficiently rich palette from which to compose. A recognition of this idea has led to work on quantifying complexity in music, including by computing the entropy S of music in the context of information theory (8) or by considering musical systems to be self-organizing via an evolutionary process (9).

The model I present here combines both the minimization of D and the maximization of S. I draw an analogy to thermodynamic systems with energy U and entropy S, whose macrostate is determined by minimizing Helmholtz free energy F = U − TS. The fixed temperature T is a parameter that specifies the trade-off between decreasing U and increasing S. Here, I similarly introduce a parameter T that specifies the trade-off between decreasing D and increasing S. A musical system in equilibrium will then be found by minimizing F = D − TS, allowing us to exploit the powerful array of tools developed for studying physical systems in statistical mechanics.

The remainder of this paper is organized as follows: I next describe the general model presented here, including how dissonance is quantified. Then, we study the behavior of the model in the mean field approximation and observe phase transitions between disordered sound and ordered distributions of pitches that reproduce commonly used musical systems. Last, we turn to a more realistic model with fewer assumptions and use numerical simulation to explore the patterns that emerge on a lattice of interacting tones.

Kinda' Cool: Miles Davis Meets Mr. Rogers

Dang, I never knew that Miles Davis was a guest on Mister Rogers’ Neighborhood. pic.twitter.com/yJDSeUYIzR

— Ted Gioia (@tedgioia) May 21, 2019

Irises

"Foxes" are better at predicting the future than "hedgehogs"

David Epstein, The Peculiar Blindness of Experts, The Atlantic, June 2019:

One subgroup of scholars, however, did manage to see more of what was coming. Unlike Ehrlich and Simon, they were not vested in a single discipline. They took from each argument and integrated apparently contradictory worldviews. They agreed that Gorbachev was a real reformer and that the Soviet Union had lost legitimacy outside Russia. A few of those integrators saw that the end of the Soviet Union was close at hand and that real reforms would be the catalyst.

The integrators outperformed their colleagues in pretty much every way, but especially trounced them on long-term predictions. Eventually, Tetlock bestowed nicknames (borrowed from the philosopher Isaiah Berlin) on the experts he’d observed: The highly specialized hedgehogs knew “one big thing,” while the integrator foxes knew “many little things.”

Hedgehogs are deeply and tightly focused. Some have spent their career studying one problem. Like Ehrlich and Simon, they fashion tidy theories of how the world works based on observations through the single lens of their specialty. Foxes, meanwhile, “draw from an eclectic array of traditions, and accept ambiguity and contradiction,” Tetlock wrote. Where hedgehogs represent narrowness, foxes embody breadth.

Incredibly, the hedgehogs performed especially poorly on long-term predictions within their specialty. They got worse as they accumulated experience and credentials in their field. The more information they had to work with, the more easily they could fit any story into their worldview. [...]

In Tetlock’s 20-year study, both the broad foxes and the narrow hedgehogs were quick to let a successful prediction reinforce their beliefs. But when an outcome took them by surprise, foxes were much more likely to adjust their ideas. Hedgehogs barely budged. Some made authoritative predictions that turned out to be wildly wrong—then updated their theories in the wrong direction. They became even more convinced of the original beliefs that had led them astray. The best forecasters, by contrast, view their own ideas as hypotheses in need of testing. If they make a bet and lose, they embrace the logic of a loss just as they would the reinforcement of a win. This is called, in a word, learning.

Scrimshawed whales's tooth (by my uncle, Erik "Rune" Ronnberg, Sr.)

Ridding Facebook of "bad activity" – AI isn't up to the task

Cade Metz and Mike Isaac, Facebook’s A.I. Whiz Now Faces the Task of Cleaning It Up. Sometimes That Brings Him to Tears. NYTimes, 17 May 2019.

Mr. Schroepfer — or Schrep, as he is known internally — is the person at Facebook leading the efforts to build the automated tools to sort through and erase the millions of such posts. But the task is Sisyphean, he acknowledged over the course of three interviews recently.

That’s because every time Mr. Schroepfer and his more than 150 engineering specialists create A.I. solutions that flag and squelch noxious material, new and dubious posts that the A.I. systems have never seen before pop up — and are thus not caught. The task is made more difficult because “bad activity” is often in the eye of the beholder and humans, let alone machines, cannot agree on what that is.

In one interview, Mr. Schroepfer acknowledged after some prodding that A.I. alone could not cure Facebook’s ills. “I do think there’s an endgame here,” he said. But “I don’t think it’s ‘everything’s solved,’ and we all pack up and go home.”

Dealing with images (e.g. nudity) is one thing. Words are more difficult:

Identifying rogue images is also one of the easier tasks for A.I. It is harder to build systems to identify false news stories or hate speech. False news stories can easily be fashioned to appear real. And hate speech is problematic because it is so difficult for machines to recognize linguistic nuances. Many nuances differ from language to language, while context around conversations rapidly evolves as they occur, making it difficult for the machines to keep up.

Delip Rao, head of research at A.I. Foundation, a nonprofit that explores how artificial intelligence can fight disinformation, described the challenge as “an arms race.” A.I. is built from what has come before. But so often, there is nothing to learn from. Behavior changes. Attackers create new techniques. By definition, it becomes a game of cat and mouse.

“Sometimes you are ahead of the people causing harm,” Mr. Rao said. “Sometimes they are ahead of you.”

On that afternoon, Mr. Schroepfer tried to answer our questions about the cat-and-mouse game with data and numbers. He said Facebook now automatically removed 96 percent of all nudity from the social network. Hate speech was tougher, he said — the company catches 51 percent of that on the site. (Facebook later said this had risen to 65 percent.)