Teaching Machines to Read, and Humans to Understand

Prismatic is dead. Sad!

Prismatic was a news reader that used machine learning to read and recommend content. Despite the fact that queries for “interest rate” often returned blog posts from ZeroHedge or sponsored content from credit agencies, it mostly did a good job.

Machine readers have come a long way. In the early days, neural networks processed data points independently. A translation machine would interpret a string like “Errare humanum est” word-by-word: To err human the.

A recurrent neural network feeds the output back into the hidden layer
A recurrent neural network feeds the outputs back into the hidden layer

Recurrent neural networks (RNNs) brought the ability to propagate past information. Words, with grammatical context, become sentences. When a machine has processed a subject and complement it knows to add a linking verb: To err is human.

A subset of RNNs, Long Short Term Memory networks, can persist early information without distortion from new inputs. By preserving long-term dependencies, a machine can connect sentences into a story.

To err is human. My wife could do no wrong. My wife is therefore…[not human].

Recent research has introduced the concept of attention [1], where a machine can selectively focus on data relevant to the question at hand. This model enables machines to answer queries on large bodies of text without trying to propagate dependencies from beginning to end.

Somewhere along the way, as machines gained the capacity for contextual comprehension, humans lost that same ability.

It used to be that humans could understand intent. We knew that words were clumsy vehicles used to communicate ideas, and so we learned to separate meaning from words. Now we can’t see beyond words.

This is not a memory-bound problem. College students in particular have excellent Long Short Term Memories, emphasis on the long. In fact, students at Harvard have identified that the Law School seal is derived from Isaac Royall, Jr’s family crest of two-and-a-half centuries ago, and that the Royall family owned slaves.

The seal—which adorns all of our buildings, apparel, stationery, and diplomas—honors a slaver and murderer. —Royall Must Fall

A computer might recognize that slave ownership is not Mr. Royall’s most salient feature – at least not where Harvard Law is concerned. A computer could identify that a slightly more relevant feature is Isaac Royall, Jr’s land donation, which led to the founding of the damn school.

By understanding dependencies, a computer can determine intent. It can separate someone else’s ignorance from its own. Maybe computers have more self-awareness than humans.

People say that a general artificial intelligence is still decades away. It’s clear that machines have surpassed human intelligence at least in reading comprehension and acting like a functional adult. Maybe humans should try to be more like computers.

See Also:
1. Hermann, et al. Teaching Machines to Read and Comprehend. Advances in Neural Information Processing Systems, 28 (NIPS 2015)

2 thoughts on “Teaching Machines to Read, and Humans to Understand

  1. Your non-political posts are much more fun, although I completely agree. This is the nirvana fallacy – there are no perfect historical figures, and we’d have to drop honoring all of them if we rejected anyone with a moral failing. If changing, for example, a football team mascot helps build a better sense of community at a university, then do it, but something like a detail on a university crest seems a bit nit-picky. This also brings out the metaphor of social warriors who somehow managed to switch the windshield and the rear-view mirror of their cars. Yes, look back some, but the main focus should be on fixing the problems of today.

Leave a Reply to Elaine Cancel reply