The computers are learning faster than you think
An A.I. created a carol after studying a photograph of a Christmas tree. Unfortunately the song is creepy as hell.
↩︎ posted Dec. 9, 2016
While some see this new prediction tech as like a new pipe tech that could improve all pipes, no matter their size, it is actually more like a tech only useful on very large pipes. Just as it would be a waste to force a pipe tech only useful for big pipes onto all pipes, it can be a waste to push advanced prediction tech onto typical prediction tasks.
Machine learning is all the rage, but most firms really just want the cachet of leading-edge tech; they could get their predictions with a simple regression and good clean data.
↩︎ Overcoming Bias, posted Dec. 6, 2016
Prediction is cheap, unlike ethics
The semiconductor lowered the cost of arithmetic, which the Harvard Business Review cites to predict that the economics of machine learning mean we’ll start using predictions on everything: “The first effect of machine intelligence will be to lower the cost of goods and services that rely on prediction.” The next step will be that we’ll start using prediction as an input even in industries where it has typically had little reason to be applied.
That shift is already underway as big tech companies change their business models to incorporate AI at every level. Google, Facebook, Twitter, Amazon, and Microsoft are all investing heavily; they’ve even worked together to establish best practices in AI ethics, a partnership from which Apple was notably marked absent.
↩︎ posted Nov. 29, 2016
Google Translate invented its own shortcut
Google announced that their translation team discovered something interesting: after they turned over translation to a machine learning-enabled AI, it figured out its own way of translating.
After being taught to translate between English and Korean, then between Japanese and English, the AI invented its own fourth reference language that may draw upon deep-learning insights into a shared structure between the three languages. “This “interlingua” seems to exist as a deeper level of representation that sees similarities between a sentence or word in all three languages,” says TechCrunch.
Full explanation of the phenomenon from the Google team here.
↩︎ posted Nov. 29, 2016
Artificial intelligence wrote the dialogue for the eerie short film Sunspring—part of a wave of generative programs being employed across creative endeavors.
↩︎ posted Nov. 29, 2016
Korean? Easy. Archie comics? Not so fast.
Machine learning treats a dataset as a closed system. In the case of comic books, too much happens in between and outside the frame for the machine to understand.
For now, only humans can achieve “closure,” a measure of literary comprehension that computer scientists test in humans and AI by asking for a prediction of text in successive panels.
↩︎ posted Nov. 29, 2016
Spotted on a recent jaunt to the Whitney Museum: this video, autoencoded from Blade Runner using neural networks.
↩︎ posted Nov. 29, 2016
“Power is not a great place for a good time.”
A New Zealand scientist, apparently fed up with the blindly financial reasons why academic conferences often accept papers of dubious quality, wrote a paper entirely reliant on Autocorrect’s AI. The paper was accepted.
↩︎ posted Nov. 29, 2016
[The bot] lives under the assumption that nothing will be novel, as if out of faith. It fields sentences by comparing them with those it knows, understanding phrasings using algorithms somewhat like Markov chains. Then, it assembles a response according to poetic constraints, rules and templates, or selects the best one from a list.
As in poetry, limitations give rise to creativity.
↩︎ Real Life, posted Nov. 29, 2016