Moving beyond detection of statistical patterns in AI

Gary Marcus and Ernest Davis writing for the NYT:

. . . We recently searched on Google for “Did George Washington own a computer?” — a query whose answer requires relating two basic facts (when Washington lived, when the computer was invented) in a single temporal framework. None of Google’s first 10 search results gave the correct answer. The results didn’t even really address the question. The highest-ranked link was to a news story in The Guardian about a computerized portrait of Martha Washington as she might have looked as a young woman.

Google’s Talk to Books, an A.I. venture that aims to answer your questions by providing relevant passages from a huge database of texts, did no better. It served up 20 passages with a wide array of facts, some about George Washington, others about the invention of computers, but with no meaningful connection between the two.

The situation is even worse when it comes to A.I. and the concepts of space and causality. Even a young child, encountering a cheese grater for the first time, can figure out why it has holes with sharp edges, which parts allow cheese to drop through, which parts you grasp with your fingers and so on. But no existing A.I. can properly understand how the shape of an object is related to its function. Machines can identify what things are, but not how something’s physical features correspond to its potential causal effects.

How to Build Artificial Intelligence We Can Trust

Modern AI’s have no basic understanding of the world, and there’s not much progress.

Does automating the boring stuff leave only the unpleasant stuff?

Fred Benenson, writing for The Atlantic, is concerned that AI automation will leave only the most difficult and unpleasant tasks for humans:

What’s less understood is that artificial intelligence will transform higher-skill positions, too—in ways that demand more human judgment rather than less. And that could be a problem. As AI gets better at performing the routine tasks traditionally done by humans, only the hardest ones will be left for us to do. But wrestling with only difficult decisions all day long is stressful and unpleasant. Being able to make at least some easy calls, such as allowing Santorini onto Kickstarter, can be deeply satisfying.

“Decision making is very cognitively draining,” the author and former clinical psychologist Alice Boyes told me via email, “so it’s nice to have some tasks that provide a sense of accomplishment but just require getting it done and repeating what you know, rather than everything needing very taxing novel decision making.”

AI Is Coming for Your Favorite Menial Tasks

He recognizes that many professions (e.g., lawyers!) may welcome automation of the boring stuff. But he’s particularly concerned about content moderators.

But we may find that as jobs get harder, the benefits get better.

Survey suggests most Americans support police use of facial recognition technology

According to the Pew Research Center, a full 56 percent said that they trust police and officials to use these technologies responsibly. That goes for situations in which no consent is given: About 59 percent said it is acceptable for law enforcement to use facial recognition tools to assess security threats in public spaces.

Police Use of Facial Recognition is Just Fine, Say Most Americans

Black and Hispanic adults approve at lower rates. See the study for details.

UK court approves police use of facial recognition

In contrast to recent U.S. municipal decisions restricting government use of facial recognition technology, a UK court has ruled that police use of the technology does not violate any fundamental rights.

In one of the first lawsuits to address the use of live facial recognition technology by governments, a British court ruled on Wednesday that police use of the systems is acceptable and does not violate privacy and human rights.

Police Use of Facial Recognition Is Accepted by British Court

The UK is of course one of the most surveilled countries in the world.

Language models are improving quickly

Aristo, an AI developed by the Allen Institute for Artificial Intelligence, passed an eighth-grade science test with a score of 90%. They used BERT.

At Google, researchers built a system called Bert that combed through thousands of Wikipedia articles and a vast digital library of romance novels, science fiction and other self-published books.

Through analyzing all that text, Bert learned how to guess the missing word in a sentence. By learning that one skill, Bert soaked up enormous amounts of information about the fundamental ways language is constructed. And researchers could apply that knowledge to other tasks.

The Allen Institute built their Aristo system on top of the Bert technology. They fed Bert a wide range of questions and answers. In time, it learned to answer similar questions on its own.

A Breakthrough for A.I. Technology: Passing an 8th-Grade Science Test

AI for military also means compassion

Phenomenal essay by Lucas Kunce, a U.S. Marine who served in Iraq and Afghanistan, responding to news that 4,600 Google employees signed a petition urging the company to refuse to build weapons technology:

People frequently threw objects of all sizes at our vehicles in anger and protest. Aside from roadside bombs, the biggest threat at the time, particularly in crowded areas, was an armor-piercing hand-held grenade. It looked like a dark soda can with a handle protruding from the bottom. Or, from a distance and with only an instant to decide, it looked just like many of the other objects that were thrown at us. 

One day in Falluja, at the site of a previous attack, an Iraqi man threw a dark oblong object at one of the vehicles in my sister team. The Marine in the turret, believing it was an armor-piercing grenade, shot the man in the chest. The object turned out to be a shoe.

[. . . . .]

When I think about A.I. and weapons development, I don’t imagine Skynet, the Terminator, or some other Hollywood dream of killer robots. I picture the Marines I know patrolling Falluja with a heads-up display like HoloLens, tied to sensors and to an A.I. system that can process data faster and more precisely than humanly possible — an interface that helps them identify an object as a shoe, or an approaching truck as too light to be laden with explosives.

Dear Tech Workers, U.S. Service Members Need Your Help

Interview with John Shawe-Taylor, professor at University College London

I enjoyed this interview and especially the title: “Humans Don’t Realize How Biased They Are Until AI Reproduces the Same Bias, Says UNESCO AI Chair.”

What are some core problems or research areas you want to approach?

People are now solving problems just by throwing an enormous amount of computation and data at them and trying every possible way. You can afford to do that if you are a big company and have a lot of resources, but people in developing countries cannot afford the data or the computational resources. So the theoretical challenge, or the fundamental challenge, is how to develop methods that are better understood and therefore don’t need experiments with hundreds of variants to get things to work.

Another thing is that some of the problems with current datasets, especially in terms of the usefulness of these systems for different cultures, is that there is a cultural bias in the data that has been collected. It is Western data informed with the Western way of seeing and doing things, so to some extent having data from different cultures and different environments is going to help make things more useful. You need to learn from data that is more relevant to the task.

Humans Don’t Realize How Biased They Are Until AI Reproduces the Same Bias, Says UNESCO AI Chair

And of course:

“Solving” is probably too strong, but for addressing those problems, as I’ve said, the problem is that we don’t realise that they are the reflections of our own problems. We don’t realise how biased we are until we see an AI reproduce the same bias, and we see that it’s biased.

I chuckle a bit when I hear about biased humans going over biased data in the hopes of creating unbiased data. Bias is a really hard problem, and it’s always going to be with us in one form or another. Education and awareness are the most important tools for addressing.

It is easy to be fooled if you do not understand how a model works

Benjamin Heinzerling writes in The Gradient that the “Clever Hans effect” is alive and well in natural language processing (NLP) deep learning models:

Of course, the problem of learners solving a task by learning the “wrong” thing has been known for a long time and is known as the Clever Hans effect, after the eponymous horse which appeared to be able to perform simple intellectual tasks, but in reality relied on involuntary cues given by its handler. Since the 1960s, versions of the tank anecdote tell of a neural network trained by the military to recognize tanks in images, but actually learning to recognize different levels of brightness due to one type of tank appearing only in bright photos and another type only in darker ones.

Less anecdotal, Viktoria Krakovna has collected a depressingly long list of agents following the letter, but not the spirit of their reward function, with such gems as a video game agent learning to die at the end of the first level, since repeating that easy level gives a higher score than dying early in the harder second level. Two more recent, but already infamous cases are an image classifier claimed to be able to distinguish faces of criminals from those of law-abiding citizens, but actually recognizing smiles and a supposed “sexual orientation detector” which can be better explained as a detector of glasses, beards and eyeshadow.

NLP’s Clever Hans Moment has Arrived

BERT is a fantastic NLP model, but it’s not displaying deep understanding of the material. For certain tasks, at least, it is exploiting statistical correlations better than you can. And that makes it hard to see what it’s doing.

Reminds me of one of my favorite quotes: “The first principle is that you must not fool yourself – and you are the easiest person to fool.” Richard Feynman

Deep Learning and Astronomy

Anil Ananthaswamy, writing for Scientific American:

[T]the signals being collected by LIGO must be matched by supercomputers against hundreds of thousands of templates of possible gravitational-wave signatures. Promising signals trigger an internal alert; those that survive additional scrutiny trigger a public alert so that the global astronomy community can look for electromagnetic and neutrino counterparts.

Template matching is so computationally intensive that, for gravitational waves produced by mergers, astronomers use only four attributes of the colliding cosmic objects (the masses of both and the magnitudes of their spins) to make detections in real time. From there, LIGO scientists spend hours, days or even weeks performing more processing offline to further refine the understanding of a signal’s sources, a task called parameter estimation.

Seeking ways to make that labyrinthine process faster and more computationally efficient, in work published in 2018, Huerta and his research group at NCSA turned to machine learning. Specifically, Huerta and his then graduate student Daniel George pioneered the use of so-called convolutional neural networks (CNNs), which are a type of deep-learning algorithm, to detect and decipher gravitational-wave signals in real time.

Faced with a Data Deluge, Astronomers Turn to Automation

And they learned a bit about what the neural networks are seeing:

For Ntampaka, these results suggest that machine-learning systems are not entirely immune to interpretation. “It’s a misunderstanding within the community that they only can be black boxes,” she says. “I think interpretability is on the horizon. It’s coming. We are starting to be able to do it now.” But she also acknowledges that had her team not already known the underlying physics connecting the x-ray emissions from galaxy clusters to their mass, it might not have figured out that the neural network was excising the cores from its analysis.

Skepticism on Deep Learning, Reinforcement Learning, and DeepMind

Gary Marcus, NYU professor of psychology and neural science, is skeptical in light of DeepMind’s loss of $572M last year:

My own guess?

Ten years from now we will conclude that deep reinforcement learning was overrated in the late 2010s, and that many other important research avenues were neglected. Every dollar invested in reinforcement learning is a dollar not invested somewhere else, at a time when, for example, insights from the human cognitive sciences might yield valuable clues. Researchers in machine learning now often ask, “How can machines optimize complex problems using massive amounts of data?” We might also ask, “How do children acquire language and come to understand the world, using less power and data than current AI systems do?” If we spent more time, money, and energy on the latter question than the former, we might get to artificial general intelligence a lot sooner.

DEEPMIND’S LOSSES AND THE FUTURE OF ARTIFICIAL INTELLIGENCE

Deep learning has been so hyped that it will be difficult to meet expectations. And reinforcement learning has serious challenges when applied to real-world environments. But they are both revolutions in AI and will alter computing forever.