Missing data hinder replication of artificial intelligence studies | Science

7 bulan ago kesimpulan 0
Bringing you the latest Science news from the world.

What is happening in our world? Who is discovering what? What is going on now? These are questions that will be answered. Enjoy.

Missing data hinder replication of artificial intelligence studies | Science

The same algorithm can learn to walk in wildly different ways.

YUVAL TASSA

Last year, computer scientists at the University of Montreal (U of M) in Canada were eager to show off a new speech recognition algorithm, and they wanted to compare it to a benchmark, an algorithm from a well-known scientist. The only problem: The benchmark’s source code wasn’t published. The researchers had to recreate it from the published description. But they couldn’t get their version to match the benchmark’s claimed performance, says Nan Rosemary Ke, a Ph.D. student in the U of M lab. “We tried for 2 months and we couldn’t get anywhere close.”

The booming field of artificial intelligence (AI) is grappling with a replication crisis, much like the ones that have afflicted psychology, medicine, and other fields over the past decade. AI researchers have found it difficult to reproduce many key results, and that is leading to a new conscientiousness about research methods and publication protocols. “I think people outside the field might assume that because we have code, reproducibility is kind of guaranteed,” says Nicolas Rougier, a computational neuroscientist at France’s National Institute for Research in Computer Science and Automation in Bordeaux. “Far from it.” Last week, at a meeting of the Association for the Advancement of Artificial Intelligence (AAAI) in New Orleans, Louisiana, reproducibility was on the agenda, with some teams diagnosing the problem—and one laying out tools to mitigate it.

The most basic problem is that researchers often don’t share their source code. At the AAAI meeting, Odd Erik Gundersen, a computer scientist at the Norwegian University of Science and Technology in Trondheim, reported the results of a survey of 400 algorithms presented in papers at two top AI conferences in the past few years. He found that only 6% of the presenters shared the algorithm’s code. Only a third shared the data they tested their algorithms on, and just half shared “pseudocode”—a limited summary of an algorithm. (In many cases, code is also absent from AI papers published in journals, including Science and Nature.)

Researchers say there are many reasons for the missing details: The code might be a work in progress, owned by a company, or held tightly by a researcher eager to stay ahead of the competition. It might be dependent on other code, itself unpublished. Or it might be that the code is simply lost, on a crashed disk or stolen laptop—what Rougier calls the “my dog ate my program” problem.

Assuming you can get and run the original code, it still might not do what you expect. In the area of AI called machine learning, in which computers derive expertise from experience, the training data for an algorithm can influence its performance. Ke suspects that not knowing the training for the speech-recognition benchmark was what tripped up her group. “There’s randomness from one run to another,” she says. You can get “really, really lucky and have one run with a really good number,” she adds. “That’s usually what people report.”

In a survey of 400 artificial intelligence papers presented at major conferences, just 6% included code for the papers’ algorithms. Some 30% included test data, whereas 54% included pseudocode, a limited summary of an algorithm.

CREDITS: (GRAPHIC) E. HAND/SCIENCE; (DATA) GUNDERSEN AND KJENSMO, ASSOCIATION FOR THE ADVANCEMENT OF ARTIFICIAL INTELLIGENCE 2018

At the AAAI meeting, Peter Henderson, a computer scientist at McGill University in Montreal, showed that the performance of AIs designed to learn by trial and error is highly sensitive not only to the exact code used, but also to the random numbers generated to kick off training, and to “hyperparameters”—settings that are not core to the algorithm but that affect how quickly it learns. He ran several of these “reinforcement learning” algorithms under different conditions and found wildly different results. For example, a virtual “half-cheetah”—a stick figure used in motion algorithms—could learn to sprint in one test but would flail around on the floor in another. Henderson says researchers should document more of these key details. “We’re trying to push the field to have better experimental procedures, better evaluation methods,” he says.

Henderson’s experiment was conducted in a test bed for reinforcement learning algorithms called Gym, created by OpenAI, a nonprofit based in San Francisco, California. John Schulman, a computer scientist at OpenAI who helped create Gym, says that it helps standardize experiments. “Before Gym, a lot of people were working on reinforcement learning, but everyone kind of cooked up their own environments for their experiments, and that made it hard to compare results across papers,” he says.

IBM Research presented another tool at the AAAI meeting to aid replication: a system for recreating unpublished source code automatically, saving researchers days or weeks of effort. It’s a neural network—a machine learning algorithm made of layers of small computational units, analogous to neurons—that is designed to recreate other neural networks. It scans an AI research paper looking for a chart or diagram describing a neural net, parses those data into layers and connections, and generates the network in new code. The tool has now reproduced hundreds of published neural networks, and IBM is planning to make them available in an open, online repository.

Joaquin Vanschoren, a computer scientist at Eindhoven University of Technology in the Netherlands, has created another repository for would-be replicators: a website called OpenML. It hosts not only algorithms, but also data sets and more than 8 million experimental runs with all their attendant details. “The exact way that you run your experiments is full of undocumented assumptions and decisions,” Vanschoren says. “A lot of this detail never makes it into papers.”

Psychology has dealt with its reproducibility crisis in part by creating a culture that favors replication, and AI is starting to do the same. In 2015, Rougier helped start ReScience, a computer science journal dedicated to replications. The large Neural Information Processing Systems conference has started linking from its website to papers’ source code when available. And Ke is helping organize a “reproducibility challenge,” in which researchers are invited to try to replicate papers submitted for an upcoming conference. Ke says nearly 100 replications are in progress, mostly by students, who may receive academic credit for their efforts.

Yet AI researchers say the incentives are still not aligned with reproducibility. They don’t have time to test algorithms under every condition, or the space in articles to document every hyperparameter they tried. They feel pressure to publish quickly, given that many papers are posted online to arXiv every day without peer review. And many are reluctant to report failed replications. At ReScience, for example, all the published replications have so far been positive. Rougier says he’s been told of failed attempts, but young researchers often don’t want to be seen as criticizing senior researchers. That’s one reason why Ke declined to name the researcher behind the speech recognition algorithm she wanted to use as a benchmark.

Gundersen says the culture needs to change. “It’s not about shaming,” he says. “It’s just about being honest.”

Source link

Iklan