Profile cover photo
Profile photo
Wayne Radinsky
18,571 followers -
Software Design Engineer
Software Design Engineer

18,571 followers
About
Posts

Post is pinned.Post has attachment
Add a comment...

Post has attachment
OpenAI is releasing Spinning Up in Deep RL, "an educational resource designed to let anyone learn to become a skilled practitioner in deep reinforcement learning. Spinning Up consists of crystal-clear examples of RL code, educational exercises, documentation, and tutorials."

"Spinning Up in Deep RL consists of the following core components: A short introduction to RL terminology, kinds of algorithms, and basic theory. An essay about how to grow into an RL research role. A curated list of important papers organized by topic. A well-documented code repo of short, standalone implementations of: Vanilla Policy Gradient (VPG), Trust Region Policy Optimization (TRPO), Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), Twin Delayed DDPG (TD3), and Soft Actor-Critic (SAC). And a few exercises to serve as warm-ups."
Spinning Up in Deep RL
Spinning Up in Deep RL
blog.openai.com
Add a comment...

Post has attachment
"Jason Antic's recent DeOldify deep learning project not only colorizes images but also restores them, with stunning results."

"The reason why black and white to color (and other deep learning applications) didn't work as well as they could was because there was still a human involved in hand-coding a key step: Evaluating whether or not the generated image 'looked good.' A.K.A the 'loss function.'
The most immediately obvious (and wrong!) way to evaluate if the neural network creates a good image is by comparing pixels directly and penalizing according to how different they are. This just encourages the neural network be very conservative in its predictions: Green for grass/trees (that's easy!), blue for skies (easy!), skin for...skin (easy!)....and then brownish for everything else it just doesn't know for sure. Brown is numerically a very average color so it's a good bet if your network is encouraged to just minimize deviations from expected pixel values."

"In contrast, GANs effectively replace those hand coded loss function with a network -- the critic/discriminator -- that learns all this stuff for you, and learns it well."
Add a comment...

Post has attachment
Yoshua Benjio get asked about some AI ethics. AI race between different countries? Collaboration between countries? Just a few AI companies dominating the field of AI? Military uses of AI? New AI research? Causality?
Add a comment...

Post has attachment
The kilogram is being redefined. Yay! Finally!

The official date of the switchover is May 20, 2019. Mark your calendars! We should all have a big celebration on that day!

On that day, ALL SI units will (finally!) be defined by fundamental constants of nature (yay!):

- The transition frequency of caesium-133 (which defines the unit of time, the second, and the unit of frequency, the hertz, and from those, the unit of radioactive decays per unit time, the becquerel)
- the speed of light in a vacuum (which defines the unit of length, the meter, and from that the unit of absorbed dose of ionizing radiation, the sievert),
- the Planck constant (which now defines the unit of mass, the gram/kilogram, and from that we get the unit of force, the newton, the unit of pressure, the pascal, the unit of energy, the joule, and the unit of power, the watt),
- the charge of the electron (which defines the unit of charge, the coulomb, and from it, units of current and voltage, the ampere and the volt, as well as the unit of electrical capacitance, the farad, the unit of electrical resistance, the ohm, the unit of electrical conductance, the siemens, the unit of magnetic flux, the weber, the unit of magnetic field strength, the telsa, and the unit of electrical inductance, the henry),
- the Boltzman constant (which defines the unit of temperature, the kelvin, from which we get... just other units of temperature, like the degree Celsius),
- the Avogadro constant (which defines the unit of chemical quantity, the mole, and from that the unit of catalytic activity, the katal), and
- the luminous efficacy of monochromatic radiation of 555 nm light (which defines the unit of brightness of light, the candela, which in turn gives us the unit of luminous flux, the lumen, and the unit of illuminance, the lux)


The official definitions will be:

− the unperturbed ground state hyperfine transition frequency of the caesium 133 atom will be defined as 9,192,631,770 hertz,
− the speed of light in vacuum will be defined as 299,792,458 meters per second,
− the Planck constant will be defined as 6.62607015 x 10^−34 joule-seconds,
− the charge of the electron will be defined as 1.602176634 x 10^−19 coulombs,
− the Boltzmann constant will be defined as 1.380649 x 10^−23 joules per degree kelvin,
− the Avogadro constant will be defined as 6.02214076 x 10^23 per mole, and
− the luminous efficacy of monochromatic radiation at 555 nanometers (540 x 10^12 Hz) will be defined as 683 lumens per watt.
Add a comment...

Post has attachment
"The University of San Francisco has struggled to hire enough professors to teach AI. Being in the Bay Area, the combination of expensive home prices and the strong draw from industry is hard for practitioners to ignore when looking for work. The University of San Francisco gets 700 to 800 applications a year to the school's machine learning master's program, and takes less than 10%. They could potentially accept more students or expand offerings, but they don't have the faculty to do it."

"Some of the biggest names in AI maintain a heavy load of teaching and industry work, or cycle between the two. Geoffrey Hinton joined Google and splits his time there with the University of Toronto. Yann LeCun divides his time with New York University and Facebook. Andrew Ng balanced being a professor at Stanford University and leading Baidu's AI group before leaving the latter to start his own fund and other projects."

"To create highly trained AI researchers, it's not just teaching students. It is also advising and helping research get published. That type of work is much different than the applied AI most workplaces are interested in."
Add a comment...

Post has attachment
"In early 2018, Microsoft announced the first machine translation system to translate sentences of news articles from Chinese to English with the same quality and accuracy as a person. Today, Microsoft Translator brings some of these research advances into production as it releases new translation systems for Chinese and German from and to English, increasing translation quality and expanding application scenarios."

"Based on human evaluations using industry standard test sets, we are seeing up to 11% improvement in overall quality. The improvement is compared to our previous production API."

"These improvements are a result of the research systems that achieved human parity translation on a commonly used set of news stories earlier this year, and which are now adapted to the production API. These include next generation neural net architectures and dual learning" (whatever that is).
Add a comment...

Post has attachment
In case you're wondering what neutrinos are (from the video yesterday on supersymmetry), Fraser Cain explains in this video, "The search for neutrinos".
Add a comment...

Post has attachment
Pixyz is a library for developing deep generative models in PyTorch. "Recently, many papers on deep generation models have been published. However, it is likely to be difficult to reproduce them with codes since there is a gap between mathematical formulas presented in these papers and actual implementation of them. Our objective is to create a new library which enables us to fill this gap and easy to implement these models. With Pixyz, you can implement even more complicated models just as if writing these formulas."
masa-su/pixyz
masa-su/pixyz
github.com
Add a comment...

Post has attachment
Google has announced that it's absorbing DeepMind Health, a part of its London-based AI lab DeepMind."

"DeepMind's founders said it was a 'major milestone' for the company that would help turn its Streams app -- which it developed to help the UK's National Health Service (NHS) -- into 'an AI-powered assistant for nurses and doctors' that combines 'the best algorithms with intuitive design.' Currently, the Streams app is being piloted in the UK as a way to help health care practitioners manage patients."

"DeepMind says its Streams team will remain in London and that it's committed to carrying out ongoing work with the NHS. These include a number of ambitious research projects, such as using AI to spot eye disease in routine scans."
Add a comment...
Wait while more posts are being loaded