AI has plenty of weaknesses - I've talked about some before, and I'll continue to talk about them in the future, but two specific weaknesses were brought to my attention this week.
AI Portraits - Won't Steal Your Data, But Might Steal Your Soul Dorian Gray-Style
I assume most of you have seen the FaceApp trend - people age-ifying their photos and unwittingly giving the rights to their photos to a shadowy Russian tech company. You've also likely seen AI paintings selling for ridiculous money.
But have you seen their lovechild AI Portraits - a more wholesome experiment run by the MIT-IBM Watson AI Lab. AI Portraits uses approximately 45,000 different Renaissance-esque 15th-century portraits and General Adversarial Networks to translate your selfie into an artistic masterpiece. It's novel because instead of simply drawing over your face it's generating new features and creating an entirely new version of your face.
Mauro Martino via YouTube
It's impressive because it determines the best style for your portrait based on your features, your background, and more.
However, it's not without "flaw". The choice of 15th-century portraiture creates a couple of clear biases. At the time, portraits of smiling or laughing individuals were rare, so your smile will likely not transfer. As well, there's a clear bias towards anglo-saxonification.
via AI Portraits
My son got excited while playing with the app and sent several of his coworkers, friends, and family through the app. If you look at the bottom right, you'll see my lovely wife Jen's portrait.
Most of you have seen my wife and know that she is Indonesian, something that is very much removed from the translation.
You can also see it in Beyonce's AI portrait.
via MyModernMet
All photos are immediately deleted from their servers after creating your image, so your privacy is safe (this time!)
All biases can be considered quirks of this current iteration of the program - which I do earnestly believe is interesting.
Later, you can imagine an AI choosing between various different styles of art based on a cornicopia of factors - or off human selection - but you have to walk before you can run, and this is a fun way to get people excited about AI.
Computer Answering Systems - No, The Answer Isn't 42
“Yes…Life, the Universe, and Everything. There is an answer. But I’ll have to think about it...the program will take me seven-and-a-half million years to run.” - Deep Thought, Hitchhiker's Guide To The Galaxy
Think of the global excitement when IBM's Watson first beat Ken Jennings in Jeopardy ... it's widely considered one of the holy grails of AI research to create a machine that truly understands the nuances of language and human thought. Yet, if you've talked to Alexa recently, you know there's a long way to go.
Today's question answering systems are basically glorified document retrieval systems. They scan text for related words and send you the most relevant options. Researchers at the University of Maryland recently figured out how to easily create questions that stump AI (without being paradoxical, impossible to answer, requiring empathy etc.) in order to enhance those systems.
A system that understands those questions will be a massive step toward a real understanding and processing of language.
So what's the secret to these "impossible" questions?
From the University of Maryland article:
The questions revealed six different language phenomena that consistently stump computers. These six phenomena fall into two categories. In the first category are linguistic phenomena: paraphrasing (such as saying “leap from a precipice” instead of “jump from a cliff”), distracting language or unexpected contexts (such as a reference to a political figure appearing in a clue about something unrelated to politics). The second category includes reasoning skills: clues that require logic and calculation, mental triangulation of elements in a question, or putting together multiple steps to form a conclusion [...]
For example, if the author writes “What composer's Variations on a Theme by Haydn was inspired by Karl Ferdinand Pohl?” and the system correctly answers “Johannes Brahms,” the interface highlights the words “Ferdinand Pohl” to show that this phrase led it to the answer. Using that information, the author can edit the question to make it more difficult for the computer without altering the question’s meaning. In this example, the author replaced the name of the man who inspired Brahms, “Karl Ferdinand Pohl,” with a description of his job, “the archivist of the Vienna Musikverein,” and the computer was unable to answer correctly. However, expert human quiz game players could still easily answer the edited question correctly.
The main change is increasing the complexity of the questions by nestling another question. In the above example, the second question forces the AI not only to decide the composer inspired by Karl Ferdinand Pohl, but also to decipher who is inspiring (hint: It's Karl Ferdinand Pohl).
AI isn't great yet at mental triangulation; at putting together multiple steps to form a conclusion. While AI is great at brute force applications - we're still coding the elegance.
via Gaping Void
