Eye on AI - August 28th, 2020
Welcome to Aigora's "Eye on AI" series, where we round up exciting news at the intersection of consumer science and artificial intelligence!
This week, we’ll be discussing the good, the rad, and the ugly of language AI systems, including an AI-inspired ancient language translation program, new smart labels that are helping the seeing impaired break down reading barriers, and the promises and letdowns of GPT-3, a powerful new language generating system.
Google’s New Translation Program Expedites Hieroglyphic Translation
We begin with the rad: Google’s Fabricius Project, an interactive experiment released this year, which uses machine learning to help translate ancient languages and open new avenues for research, has upended the tedious hieroglyphic translation process by making it easier and more fun to translate ancient languages.
“The journey began with The Hieroglyphics Initiative, a Ubisoft research project that was launched at the British Museum in September 2017 to coincide with the release of Assassin’s Creed Origins,” reads Google’s Arts & Culture website. “Working with Google and development agency Psycle Interactive, the project sought to identify whether machine learning could transform the process of collating, cataloguing and understanding the written language of the Pharaohs.”
The project uses three methods for translation: extraction, classification, and translation – in other words, it takes hieroglyphic script and sequences from source images to create workable facsimiles, trains a neural network to correctly identify over 1000 hieroglyphs, then matches those sequences and blocks of text to available dictionaries and published translations – and was the cumulative effort of numerous research research efforts and teams. For the time being, the project is focused solely on hieroglyphics (and is fun to use, with Google providing games and explanation videos in conjunction with the program). But researchers expect the same method will soon be used in the translation of numerous ancient languages.
Language Generating AI GPT-3 Produces High-Brow Gibberish
Let’s follow the rad with the bad: MIT Tech Review contributors and AI language experts Gary Marcus and Ernest Davis describe how OpenAI’s language-generating system, GPT-3, is falling drastically short of its initial hype.
“After researchers have spent millions of dollars of computer time on training, devoted a staff of 31 to the challenge, and produced breathtaking amounts of carbon emissions from electricity, GPT’s fundamental flaws remain,” write Marcus and Davis. “Its performance is unreliable, causal understanding is shaky, and incoherence is a constant companion. GPT-2 had problems with biological, physical, psychological, and social reasoning, and a general tendency toward incoherence and non sequiturs. GPT-3 does, too.”
When it was first introduced, many researchers and columnists described GPT-3’s AI as being “shockingly good,” and claimed it could compose legible, intelligent and thought-provoking text on almost any subject. The AI nailed the legible and thought-provoking parts. When it comes to actual intelligence, the jury is still out. In most tests, the AI fell well short of producing anything that might be described as intellectual or scholarly. However, it was adept at forming sentences that sounded intellectual – in other words, it produced sentences that amounted to high-brow gibberish.
Compounding frustrations, Marcus and Davis point out that OpenAI’s GPT-3 actually isn’t all that open, which might be a breach of scientific ethics, and a distortion of the goals of the associated nonprofit. As to whether the AI has many redeeming qualities, Marcus and Davis don’t offer many, and went so far as to call GPT-3 a “fluent spouter of bullshit” before adding, “even with 175 billion parameters and 450 gigabytes of input data, it’s not a reliable interpreter of the world.” Ouch.
New Smart Label App Helps the Seeing Impaired Regain Independence
As for the good, Google strikes again (rad + good = rood?) with Lookout, a new on-device smart label app which uses phone cameras to help the seeing-impaired identify material and regain some of the independence they lost with their sight (think IDing money value, reading cardboard box labeling, mail, etc.).
“When the user aims their smartphone camera at the product, Lookout identifies it and speaks aloud the brand name and product size. To accomplish this, Lookout includes a supermarket product detection and recognition model with an on-device product index, along with MediaPipe object tracking and an optical character recognition model. The resulting architecture is efficient enough to run in real-time entirely on-device.”
The advantage of Lookout is that it’s completely on-camera, and doesn’t require the memory space other label IDing apps require. It’s low memory, high result. VERY interesting tech that’s worth checking out. For a detailed look, I recommend the Google blog “On-demand Supermarket Product Recognition.”
That's it for now. If you'd like to receive email updates from Aigora, including weekly video recaps of our blog activity, click on the button below to join our email list. Thanks for stopping by!