Get up to speed on GANs in just two posts
Plus DIY stock photos with AI and no PhDs, and a very nice state of AI report
|Robert Osazuwa Ness||Jun 30, 2019|
AltDeep is a newsletter focused on microtrend-spotting in data and decision science, machine learning, and AI. It is authored by Robert Osazuwa Ness, a Ph.D. machine learning engineer at an AI startup and adjunct professor at Northeastern University.
Thanks for sticking with me. Over the next few weeks I’m experimenting with making the content more digestible without sacrificing content.
Ear to the Ground
Curated posts aimed at spotting microtrends in data science, machine learning, and AI.
Get up to speed on GANs in just two posts
Read together, these two posts bring you up to speed quickly.
A useful state of AI report
This is a good report, tying together items report here and elsewhere. It goes beyond the big research milestones to talk about broader trends in industry and across borders. These items jumped out at me.
Compensation of senior engineers is approaching $1 million, while at the same time there are $1.47/hour data labeling jobs
China’s publications’ impact and average citation rate is growing
US, China, the UK, Germany, and Canada accounted for employment of 72% of authors
AI workers Singapore, Canada, the UK, and Switzerland are more transient. Less AI workers are attracted to Japan and the US (immigration difficulties?), but when they come they tend to stay.
Full report at Stateof.ai
A reminder of the staggering costs of deep learning’s cutting edge
XLNet is a natural language model that beats a recent state-of-the-art model called BERT on NLP tasks. According to the paper "We train XLNet-Large on 512 TPU v3 chips for 500K steps with an Adam optimizer, linear learning rate decay and a batch size of 2048, which takes about 2.5 days". Using 512 TPU v3 chips for two and a half days at $8 a TPU costs $245,000. If one were not using renewable energy (which Google Brain does), this would put about 4.9 metric tons of CO2 into the atmosphere.
I believe pushing the frontier of performances is important. But imagine a startup deciding whether or not to commit to training such a model when in the end it still might not work.
AI Long Tail
Pre-hype AI microtrends
East European design SaaS shakes up stock photo business with AI
I've written recently that the exciting thing about generative machine learning models such as GANs is the ability to synthesize new data, such as images. This synthesis seems odd in a world inundated with images and other content until one remembers that it is the job of designers and content creators to create original content (or at least novel remixes of old material). They've been doing this job since long before machine learning was a thing, so the question now is can generative machine learning make designers more productive by automating some of the more laborious and commoditized (as in all the basic design gigs on Fiverr) elements of the design workflow.
My friend Glynnis put me on to an app called Photo Creator 2.0 by Icons8. The tool lets you build your own stock photos. They provide a database of images of human and animal models, objects, and rendered scenes, and an editor for composing these elements together, which can export vector images. According to their promotional collateral, the "AI" is used for two purposes within the tool:
Swapping the faces or expressions on human models
Background removal, the kind of thing one does with layer masks in Photoshop. According to Icons8 "Before, we masked 10 photos a day. Now, we process 250 photos an hour."
Icons8 provides freemium SaaS for designers, and has historically focused on icons, photos, and music. Perusing their company Linkedin page, it is clear their team has been experimenting with neural network architectures for image segmentation and generation. However, none of the members of their team with public profiles have a machine learning-specific job title. I suspect their back-end developers are behind the ML efforts. Further, the team is based mostly in Russia and Ukraine. I see this as evidence that one can build an AI-powered tool of this type without a team of prohibitively expensive Bay Area PhDs.
Photo Creator 2.0: Make Free Custom Photos with AI Power — Icons8 post on Medium
Data Sciencing for Fun and Profit
Data-Sciencing for Fun & Profit examines emerging trends on the web that could be exploited for profit using quantitative techniques.
Creating new scripts with StyleGAN
Style-transfer is the use of machine learning to apply a style learned from some training data (e.g. the look and feel of Picasso’s blue period) to a new item not in that training data (e.g. your selfiee). This post describes using style-transfer for making new scripts (scripts as in Latin, Arabic, and Chinese characters).
Creating new scripts with StyleGAN — Towards Data Science