A Judgmental Network

Previously on this blog, we've explored how humans have inadvertently passed some of our less-than-desirable tendencies on to artificial intelligence--discrimination. Today, however, let's talk about a more positive result of teaching computers to be critical: judging books by their covers.

Two scientists at Kyushu University in Japan, Brian Kenji Iwana and Seiichi Uchida, have developed a method to do precisely that. They've trained a deep neural network to scan the covers of books and determine their genre.

Now, just what is a deep neural network? According to Wikipedia, it is "an artificial neural network (ANN) with multiple hidden layers of units between the input and output layers." In English major's layman's terms, it is a set of algorithms that, when used together, are able to perform similarly to a human brain. They specialize in pattern recognition.

The two computer scientists trained their deep neural network to "read" book covers using--get this--Amazon. They downloaded exactly 137,788 book covers from the website along with the genre of each volume. The book covers they used belonged to one of 20 genres; if a book was listed under multiple genres, they just used the first.

Uchida and Iwana then used 80% of the data set to train the neural network to choose the book's correct genre just by looking at its cover. The remaining 20% was split up; the two used one 10% to validate the model and test the network and the other 10% to see how well it categorizes covers it's never seen before.

The network has had varying degrees of success. The algorithm often confuses children's books with graphic novels, and frequently mixes up biographies with books on historical eras. The algorithm lists the correct genre within its first three tries 40% of the time and correctly guessed it first 20%. While that's not exactly perfect, it is significantly better than chance.

According to Iwana and Uchida, “This shows that classification of book cover designs is possible, although a very difficult task.” 

Can't argue with that. My grandma once plucked Fifty Shades of Grey from the shelves expected to dive into a courtroom drama.

That brings me to one of the downsides of this algorithm: It's interesting and might perhaps be useful one day, sure, but there hasn't been a real study into how well humans can guess a book's genre by glancing at its cover before. Does our human experience give us a leg up to accomplishing the task? Are we, on average, superior book cover interpreters? I can't be sure.

Either way, this whole concept of deep neural networks is definitely fascinating. Who knows what patterns computer brains will be able to interpret once the programming and technology become even more sophisticated?

 Sources

• "Deep Neural Network Learns to Judge Books by Their Covers"--from MIT Technology Review
• Picture--from MIT Technology Review
• "Deep learning"--from Wikipedia
• "Introduction to Deep Neural Networks"--from DeepLearning4J.org