Jan 17, 2018

Odd, funny and scary words by a machine

Illustration: Janelle Shane / AI Weirdness

Electrical engineer Janelle Shane's hobby is giving neural networks — a computing system that loosely mimics a brain — somewhat silly datasets to see what they can create. 

Most recently, she input 44,126 thesis titles from MIT. Some of the products —  “Atoms and characteristics of monolithic nanocity,” for example — may not immediately jump out as odd. Others, Shane says, are "completely ridiculous."

Case in point: Spacecraft Coal battery induced by mortgage microcontrol.

How it works:

  • The neural network basically predicts what the next letter in a word or phrase should be based on the probability of it being in that spot across the dataset it was trained on.
  • If there is an 80% chance of an ‘h’ following a ‘c’ and a 20% chance of it instead being an ‘e,’ the responsible neural network would go with ‘h.’
  • But if the creativity setting is turned up, the AI may gamble with the less probable territory and see where it takes it, says Shane.
  • At a certain point, though, it may output words that don't even exist. (For example, "frook" and "nurler.") “It is an art form with its own connections," says Shane. 

Yes but: There is the issue of bias, which is clear in the dataset.

  • The word "control" (a favorite of engineers) and the phrase "of a medical device" (a sign of the times) appeared often.
  • And because these were MIT theses, tokamak — an experimental fusion device that was a source of institutional pride before it was shut down — turned up often.
  • The neural network returns results skewed by the raw materials it was given to start.
  • It wouldn't do a good job of detecting and collecting all of the world's theses into a database given only a sample from a school focused on engineering and science.
  • And, Shane's goal is for this to be funny, so her own sense of humor colors the results.

Chuckle all you will about the neural network's creations — and they are entertainment in Shane’s eyes — but if what's generated resonates with humans and is then replicated or perpetuated by us, the neural network is arguably creative at some level. An opera company used lines generated by Shane’s neural network and a brewery in Michigan took one of the machine’s suggestions of "The Fine Stranger" in naming one of its beers. 

The larger lesson: The AI actually understand very little, says Shane. The neural network she uses is roughly equivalent to the brain of an earthworm in terms of neurons employed. “It is limited and unintelligent when taken outside it's comfort zone," she says. "And, my computer is dedicating 100% of itself to these tasks, but an earthworm has to do other things."

Go deeper

U.S. coronavirus updates

Data: The Center for Systems Science and Engineering at Johns Hopkins; Map: Andrew Witherspoon/Axios. This graphic includes "probable deaths" that New York City began reporting on April 14.

More than 62,300 U.S. health care workers have tested positive for the novel coronavirus and at least 291 have died from the virus, the Centers for Disease Control and Prevention reported on Tuesday. COVID-19 had infected about 9,300 health professionals when the CDC gave its last update on April 17.

By the numbers: More than 98,900 people have died from COVID-19 and over 1.6 million have tested positive in the U.S. Over 384,900 Americans have recovered and more than 14.9 million tests have been conducted.

Coronavirus dashboard

Illustration: Aïda Amer/Axios

  1. Global: Total confirmed cases as of 11:00 p.m. ET: 5,589,626 — Total deaths: 350,453 — Total recoveries — 2,286,956Map.
  2. U.S.: Total confirmed cases as of 11:00 p.m. ET: 1,680,913 — Total deaths: 98,913 — Total recoveries: 384,902 — Total tested: 14,907,041Map.
  3. Federal response: DOJ investigates meatpacking industry over soaring beef pricesMike Pence's press secretary returns to work.
  4. Congress: House Republicans to sue Nancy Pelosi in effort to block proxy voting.
  5. Business: How the new workplace could leave parents behind.
  6. Tech: Twitter fact-checks Trump's tweets about mail-in voting for first timeGoogle to open offices July 6 for 10% of workers.
  7. Public health: Coronavirus antibodies could give "short-term immunity," CDC says, but more data is neededCDC releases guidance on when you can be around others after contracting the virus.
  8. What should I do? When you can be around others after contracting the coronavirus — Traveling, asthma, dishes, disinfectants and being contagiousMasks, lending books and self-isolatingExercise, laundry, what counts as soap — Pets, moving and personal healthAnswers about the virus from Axios expertsWhat to know about social distancingHow to minimize your risk.
  9. Other resources: CDC on how to avoid the virus, what to do if you get it, the right mask to wear.

Subscribe to Mike Allen's Axios AM to follow our coronavirus coverage each morning from your inbox.

Updated 46 mins ago - Politics & Policy

World coronavirus updates

Data: The Center for Systems Science and Engineering at Johns Hopkins; Map: Axios Visuals

There are no COVID-19 patients in hospital in New Zealand, which reported just 21 active cases after days of zero new infections. A top NZ health official said Tuesday he's "confident we have broken the chain of domestic transmission."

By the numbers: Almost 5.5 million people have tested positive for the novel coronavirus as of Tuesday, and more than 2.2 million have recovered. The U.S. has reported the most cases in the world (over 1.6 million from 14.9 million tests).