Lookahead+RAdam vs. Adam on some RL benchmarks

"The Ranger optimizer combines two very new developments (RAdam + Lookahead) into a single optimizer for deep learning. As proof of it’s efficacy, our team used the Ranger optimizer in recently capturing 12 leaderboard records on the FastAI global leaderboards (details here).

Lookahead, one half of the Ranger optimizer, was introduced in a new paper in part by the famed deep learning researcher Geoffrey Hinton (“LookAhead optimizer: k steps forward, 1 step back” July 2019). Lookahead was inspired by the recent advances in the understanding of neural network loss surfaces and presents a whole new way of stabilizing deep learning training and speed of convergence. Building on the breakthrough in variance management for deep learning achieved by RAdam (Rectified Adam), I find that combining RAdam plus LookAhead together (Ranger) produces a dynamic dream team and an even better optimizer than RAdam alone."

light on details

"DeepMind's version of reinforcement learning that uses "temporal value transport" to send a signal from reward backward, to shape actions, does better than alternative forms of neural networks. Here, the "TVT" program is compared to "Long-short-term memory," or LSTM, neural networks, with and without memory, and a basic reconstructive memory agent."

posted my comment on Tildes with links:

This research studied the charity of a foundation called GiveDirectly that specializes in cash transfers, and studying their effects. One of the authors is its co-founder.

I recently stumbled across this foundation in this GiveWell blog post of their 2019 top charities. I found this to be by far the most compelling on the list, both for their direct impact, and their support of research like this, which can go a long way to dispelling myths, and supporting more evidence-based charity as well as policies in developed countries like universal basic income.

They also have a page called GDLive in beta, which lists unedited survey responses to questions like personal impact, what could be improved, what the money was spent on, etc. If I find myself catastrophising my own problems, I find going there helps put my own problems into perspective, reminding me how privileged I am to have won the global birth lottery to have been born to hard-working immigrants that had the education and means to move to Canada (noticing the "invisible things that make my life easier" which "explains why most of us aren’t nearly as grateful as we ought to be" Freakonomics: Why Is My Life So Hard?). It's shocking to see how much of a life-altering difference a cash donation can make to these people, in amounts that often seem paltry to me.

in-depth explanation of the efficiency of increasing developed country public spending on clean energy research and development, which through technology spillover can benefit developing countries, not punish them, unlike cruder methods like carbon limits, criticized as disproportionately hurting developing countries that are most reliant on fossil fuel for economic growth

why do gender-segregated sports leagues exist?

good bundle backup tips

cultural genocide through dystopian digital surveillance and Orwellian re-education

"One document explicitly states that the purpose of the pervasive digital surveillance is “to prevent problems before they happen” -- in other words, to calculate who might rebel and detain them before they have a chance."

"Sauytbay called the detention center a “concentration camp...much more horrifying than prison,” with rape, brainwashing and torture in a “black room” were people screamed. She and another former prisoner, Zaomure Duwati, also told the ICIJ detainees were given medication that made them listless and obedient, and every move was surveilled."

"The documents make clear that many of those detained have not actually done anything. One document explicitly states that the purpose of the pervasive digital surveillance is “to prevent problems before they happen” -- in other words, to calculate who might rebel and detain them before they have a chance."

an interesting guide. who said the perfect media server doesn't exist?

overview and benchmarks

if you are building a 10TB ZFS proof of concept NAS with 1GbE networking, the Intel Optane Memory 32GB M.2 drive is an enormous upgrade over SATA and SAS devices in the sub $100 category if you can live with the lower endurance and reliability ratings.

As we have tested the device and had both it and the M.2 version in our lab for some time, the question is whether it is worthwhile. Since 1PB written would be about the maximum we plan for on low-end storage server equipped with only 10GbE, the Optane 900p/ 905p we see as a better value. They are faster and higher capacity with a lower cost per GB.

from October 2018:
"If you already have an SSD as your main drive and have your games and large files on your mechanical drive, it is still worth getting a second SSD to use as a cache for the mechanical drive. This small investment to get either a 32GB or a 64GB SSD and use it as your hard drive’s cache is going to be worth your money for the performance gains.

In fact, since SSDs are becoming cheaper and cheaper, there should be no reason for you not to get a 64GB SSD to use as a cache memory for your HDD."

“Authority” in this case doesn’t get much better than Matthew Ahrens, one of the cofounders of ZFS at Sun Microsystems and current ZFS developer at Delphix. In the comments to one of my filesystem articles on Ars Technica, Matthew said “There’s nothing special about ZFS that requires/encourages the use of ECC RAM more so than any other filesystem.”

So, Are Enterprise Drives Worth The Cost?
From a pure reliability perspective, the data we have says the answer is clear: No.

Enterprise drives do have one advantage: longer warranties. That’s a benefit only if the higher price you pay for the longer warranty is less than what you expect to spend on replacing the drive.

This leads to an obvious conclusion: If you’re OK with buying the replacements yourself after the warranty is up, then buy the cheaper consumer drives.