Cadence on the Beat Blogs

I just came across this article on Vice Motherboard, and it blew me away. The article was reporting on a paper published in Nature magazine on July 3, 2019, about unsupervised word embeddings capturing latent knowledge from materials science literature. Researchers from Lawrence Berkeley National Laboratory have built an AI trained on old scientific papers in materials science and made discoveries that humans missed.

Using a natural language processing (NLP) algorithm called Word2vec, the AI combed through 3.3 million abstracts on materials science and analyzed relationships between a vocabulary of about 500,000 words. From the article:

Using just the words found in scientific abstracts, the algorithm was able to understand concepts such as the periodic table and the chemical structure of molecules. The algorithm linked words that were found close together, creating vectors of related words that helped define concepts. In some cases, words were linked to thermoelectric concepts but had never been written about as thermoelectric in any abstract they surveyed.

In other words, using word prediction algorithms, the researchers discovered links between words that suggested scientific discoveries about certain properties of materials that haven’t yet been discovered.

One thing that AI is really good at is discovering patterns that can be missed by the human eye and brain—in other words, sorting through complicated, or “noisy” data. This ability has been leveraged to discover mites in bee colonies (see the last section of this blog post and this article), diagnose skin cancers (lots of articles on this one), and recognize traffic signs (this is actually an AI classification benchmark) and line markers on the roads (pretty much any new car on the road nowadays).

In this case, we are learning that we can apply AI to a bunch of noisy data—what could be noisier than language processing—and draw conclusions based on word proximity. We already have an example of this kind of technology: your phone already has a language predictability feature when you send texts.

How else can this technology be used?

The semiconductor engineer might suggest that we can analyze successful designs and make suggestions for new ways of designing chips or boards or systems, thus creating intelligent system design. The academic writer might suggest analyzing classic writers—say, Shakespeare—and determine, for example, whether he was the one who actually wrote all those plays and sonnets. The musician might suggest that we analyze all of Beethoven’s symphonies and spit out what his 10^th symphony could have sounded like. The scientist might suggest that we can make discoveries by analyzing cancer research and come up with possibilities of new treatments. The artist might suggest that we can analyze works of art and determine whether they are originals. The parent says maybe AI can analyze a toddler’s temper tantrum and determine why he is falling apart. (And then I realize that toddlers’ behaviors are just too noisy and there’s no rhyme or reason to that scenario.)

Seriously, though—this seems to be a huge step forward in making the world more “intelligent”. Humans are noisy creatures with messy lives and crowded data all vying for our attention. This AI is not to replace humans, but to be used as a tool in catching things we flawed humans might have missed. We might as well find tools to help quiet the maelstrom.

—Meera