Causal reasoning and machine learning
In a recent paper titled "The Seven Pillars of Causal Reasoning with Reflections on Machine Learning", JUDEA PEARL, professor of computer science at UCLA and author of Causality, writes:
“Current machine learning systems operate, almost exclusively, in a statistical or model-free mode, which entails severe theoretical limits on their power and performance. Such systems cannot reason about interventions and retrospection and, therefore, cannot serve as the basis for strong AI. To achieve human level intelligence, learning machines need the guidance of a model of reality, similar to the ones used in causal inference tasks. To demonstrate the essential role of such models, I will present a summary of seven tasks which are beyond reach of current machine learning systems and which have been accomplished using the tools of causal modeling."
The tasks include work on counterfactuals, and new approaches to handling incomplete data. Link to the paper. A vivid expression of the issue: "Unlike the rules of geometry, mechanics, optics or probabilities, the rules of cause and effect have been denied the benefits of mathematical analysis. To appreciate the extent of this denial, readers would be stunned to know that only a few decades ago scientists were unable to write down a mathematical equation for the obvious fact that 'mud does not cause rain.' Even today, only the top echelon of the scientific community can write such an equation and formally distinguish 'mud causes rain' from 'rain causes mud.'”
Pearl also has a new book out, co-authored by DANA MCKENZIE, in which he argues for the importance of determining cause and effect in the machine learning context. From an interview in Quanta magazine about his work and the new book:
"As much as I look into what’s being done with deep learning, I see they’re all stuck there on the level of associations. Curve fitting. That sounds like sacrilege, to say that all the impressive achievements of deep learning amount to just fitting a curve to data. If we want machines to reason about interventions ('What if we ban cigarettes?') and introspection ('What if I had finished high school?'), we must invoke causal models. Associations are not enough—and this is a mathematical fact, not opinion.
We have to equip machines with a model of the environment. If a machine does not have a model of reality, you cannot expect the machine to behave intelligently in that reality. The first step, one that will take place in maybe 10 years, is that conceptual models of reality will be programmed by humans."
- The interview caused much discussion among researchers. Twitter machine learning researcher Ferenc Huszár responded to the discourse with a post on his blog introducing and explaining the basics of Pearl's do-calculus. "It's basic hygiene for people working with data and conditional probabilities to understand the basics of this toolkit, and I feel embarrassed for completely ignoring this throughout my career. Pearl's comments may be unhelpful if interpreted as contrasting deep learning with causal inference. Rather, you should interpret it as highlighting causal inference as a huge, relatively underexplored, application of deep learning." Link.
- A 2015 paper by Harvard's Samuel Gershman on causal modeling in reinforcement learning. (NB: Huszár suggests that Pearl's criticisms "conveniently ignore" reinforcement learning). Link. ht Sara
- For more context on Pearl and causality debates, see his critique of Don Rubin's framework, and Andrew Gelman's post about the differences between the two (don't miss the comments section). ht Margarita
A legacy program at CUNY is exceptionally helpful to academically disadvantaged students
New work from Raj Chetty, John Friedman, et al reveals the startling success of a CUNY college success program that has been running for decades.
"The economists tracked more than 10,000 students who had been admitted to a CUNY program called SEEK that helps academically borderline, low-income students complete a four-year degree. They found that SEEK students earned more money as adults in their late 20s and early 30s than academically similar students from higher-income families.
'Giving them access to the SEEK program is giving them $4,000 a year,' said Friedman. 'Maybe there are more impactful things that could be done. But those data suggest that something does seem to be working well in the SEEK program.'
The results are interesting because the SEEK program, which stands for Search for Education, Elevation and Knowledge, has been around for more than 50 years and was an early attempt at CUNY to give low-income students a boost. Today, CUNY invests more resources in newer and sometimes more expensive programs to help struggling low-income students. However, this research shows that old ideas can still work."
Overview in the Hechinger Report here.
ht Will, who comments:
“The disadvantages of poverty aren't erased on the schoolhouse steps. Increasing postsecondary enrollment among low-income students is a noble objective, but too few of America's colleges and universities are prepared to serve those students when they walk into the classroom. In this context, the SEEK program is remarkable. The students who enter CUNY through SEEK come from families that earn an average of $27,000 per year. Low grades and test scores foreclose any other pathway for them onto CUNY's campuses. Yet as alumni, SEEK graduates earn just as much as low-income students with stronger academic profiles - and they earn more than wealthier students with the same grades and test scores. All at a per-student program cost that is a fraction of the total increased earnings generated for participants. Programs like SEEK are a pivotal component of a postsecondary education ecosystem that is broadly accessible and oriented toward economic mobility. And by connecting SEEK student data with Mobility Report Card data, Raj and John have given us a new way of illustrating the significant effects that programs like these can have on the trajectories of low-income Americans."
- More publications on this will be available in the future—for now, you can view a few slides from the researchers hosted on our Dropbox here.
- The CUNY page on the history of the program. Link.
- From Peter S. Goodman in the New York Times, a long new story about the basic income demonstration in Stockton, California. "As the first American city to test so-called universal basic income, Stockton will watch what happens next. So will governments and social scientists around the world as they explore how to share the bounty of capitalism more broadly at a time of rising economic inequality." Link.
- In a NYMag interview, Yanis Varoufakis expresses support for a Universal Basic Dividend: "We need to face up to the fact that [corporate] profits are increasingly produced socially; a result of cooperative, collaborative large-scale projects, many of them financed by the state and state institutions and many of them generated surreptitiously by each one of us. Every time you search something on your Google engine, you’re contributing to the capital of Google. A part of their shares [would] be contributed to a public growth fund so that we can produce a universal basic dividend, distributed to members of society that have contributed — maybe inadvertently — to the production of that social capital." Link. ht Lauren
- An excellent post at Andreessen Horowitz's blog on mortgages. Link.
- Further work on the "industrious revolution" theory by Douglas Hay: "The law of master and servant, to use the technical term, shifted markedly between 1750 and 1850 to advantage capital and disadvantage labour. Medieval in origin, it had always been adjudicated in summary hearings before lay magistrates, and provided penal sanctions to employers (imprisonment, wage abatement, and later fines), while giving workers a summary remedy for unpaid wages. The law always enforced obedience to employers’ commands, suppressed strikes, and tried to keep wages low. Between 1750 and 1850 it became more hostile to workers through legislation and judicial redefinition; its enforcement became harsher through expansion of imprisonment, capture of the local bench by industrial employers, and employer abuse of written contracts. More work in manuscript sources is needed to test the argument, but it seems likely that intensification of labour inputs during industrialisation was closely tied to these legal changes." Link.
- A new report from The Century Foundation finds a 29 percent increase in "'borrower defense' claims—applications for loan relief from students who maintain that they have been defrauded or misled by federally approved colleges and universities." Link. ht Will
- On faculty hiring networks, prestige, and epistemic inequality: "Relatively little is known about how structural factors influence the spread of ideas, and specifically how where an idea originates can influence how it spreads." Link.
- At Developing Economics, a post reflecting on Marx's legacy, the "dismal science", and the published responses to his recent 200th birthday. Link.
- A paper addresses how machine learning in medicine can make uncertainties falsely appear to be objectively settled. "We cope with some form of uncertainty when we cannot pinpoint a phenomenon exactly or when we cannot measure it precisely (i.e., approximation, inaccuracy); when we do not possess a complete account of a case (incompleteness, inadequacy); when we cannot predict what it will come next (unpredictability for randomness or excessive complexity); when our observations seem to contradict each other (inconsistency, ambiguity); and, more generally, when we are not confident of what we know. In clinical practice, all of these phenomena occur on a daily basis, several times." Link.
- From Simon-Wren Lewis at Mainly Macro: nominal wages are not real wages, and why that matters for ongoing debates regarding labor and immigration in the UK. Link.
- Chris Bertram at Crooked Timber outlines the arguments in his just-released new book: Do States Have Rights to Exclude Immigrants? Link.
- Crowdsourcing neuroscience: "Since Eyewire’s launch in 2012, more than 265,000 people have signed onto the game, and they’ve collectively colored in more than 10 million 3-D 'cubes,' resulting in the mapping of more than 3,000 neural cells, of which about a thousand are displayed in the museum." Link.