HNHacker News
TopNewBestAskShowJobs

deepnet

1,376 karma · joined May 12, 2015

Reinforcement Learning with Deep Convolutional Neural Networks for Agent and Robot Control.

Machine Intelligence, functional programming, photogrammetry, and Machine creativity.

Currently : Transferring learning between modalities in robotic neural nets.

[deepnet] [\@] jmhz [\.] net

submissionscomments
deepnet··on Tim Berners-Lee approves Web DRM, but W3C members have two weeks to appeal
“Under the spreading chestnut tree I sold you and you sold me"

Orwell, 1984 - A very sad day, a very poor decision:

"Today, the W3C announced that it would publish its DRM standard with no protections and no compromises at all, stating that W3C Director Tim Berners-Lee had concluded that the objections raised "had already been addressed" or that they were "overruled."

"EFF understood that the W3C had members who wanted to make DRM, so we suggested a compromise: a covenant, modeled on the existing W3C member-agreement, that would require members to make a binding promise only to use the law to attack people who infringed copyright, and to leave people alone if they bypassed DRM for legal reasons, like making W3C-standardized video more accessible for people with disabilities.

This was a very popular idea. It was endorsed by Unesco, by the Internet Archive, by the creator of the W3C's existing membership agreement, by hundreds of top security researchers, by the competition expert who coined the term "Net Neutrality", and by hundreds of human rights organizations and activists from the global south. The Open Source Initiative amended its definition of "open standard" so that DRM standards could only qualify as a "open" if they protected legitimate activity.

Now, it's fair to say that the W3C's DRM advocates didn't like the idea. After a perfunctory discussion process (during which some progress was made), they walked away from the negotiations, and the W3C decided to allow the standardization work to continue despite their unwillingness to compromise."[1]

https://www.eff.org/deeplinks/2017/07/amid-unprecedented-con...

deepnet··on New studies of ancient concrete could teach us to do as the Romans did
I live in a limecrete building from 1857, in the Scottish Borders.

It is similar to a tower block in minature, a girder frame with cast walls, only 3 stories high - but the rest of the village is all 2 story stone cottages. Innovative technology for 1857 !

But being limecrete it has set into limestone over the last 150 years - a very long lasting technology.

The only downside is it suffers from the 1960s portland-cement based harling (pebble dash) which keeps the moisture close to the stone - hopefully we can change this to a lime render and limewash as it should be.

deepnet··on Commander Keen Level Maps
Jason Scott has archived Commander Keen, playable in Browser via DosBox

https://archive.org/details/msdos_Commander_Keen_1_-_Maroone...

https://archive.org/details/msdos_Commander_Keen_2_-_The_Ear...

https://archive.org/details/msdos_Commander_Keen_4_-_Secret_...

https://archive.org/details/msdos__K6DEMO__shareware

https://archive.org/details/msdos_Commander_Keen_-_Keen_Drea...

Commander Keen created iD software.

Also noteworthy for the use of John Carmack's adaptive tile refresh to smooth scroll on early 90s PCs.

deepnet··on Flawed reporting about WhatsApp
> This is precisely why there was minimal reaction to the Snowden revelations - what revelations?

Or perhaps the 1984 Orwellian explanation - fear of surveillance curtails dissent.

E.g. perhaps there was minimal reaction because while people had always joked about the Govt. reading their messages, when Snowden revealed it was utterly true, and way worse than anyone had supposed, they became afraid to voice dissent about surveillance.

They were right to fear this as Snowden's stated reason for whistleblowing was not the existence of mass surveillance but because the mass surveillance was not for national security but its primary goal was to protect the program of mass surveillance.

Snowden revealed this: that the primary mission of the spying was to keep the spying secret and those who opposed the spying faced the greatest extra-legal retaliatory attacks. [1]

As the recent cataclysmic cyber attacks have shown national defense was not a priority for those 3 letter agencies developing weaponised exploits.

The weaponised exploits they screwed us all over by letting fall into the wrong hands, e.g. Wannacry an American state sponsored cyber-weapon funded by US tax-payer monies that wrecked the UK's National Health Service and now threatens international shipping - oopsie.

[1] Snowden outlines his reasons for turning whistleblower in this conversation with Chomsky & Greenwald https://www.youtube.com/watch?v=IOksJKfapVM

deepnet··on Observational Learning by Reinforcement Learning
Copying behaviour without divining intent can lead to problems such as 'cargo-culting'.

Imagine observing a man shaking his leg, first one then the other, then his whole body convulses and twitches - is he dancing ?

Absent the knowledge that a wasp has flown up his trousers leg.

Copying without comprehension may lead to getting stung !

Inverse Reinforcement Learning [1] to reverse engineer goals will be needed especially for embodied AI in Partially Observed Enviroments, i.e. the real world (as opposed to simulations).

Berkeley's CS294-112 [2] Deep Reinforcement Learning for Robotics provides good coverage of methods of mirroring, DAGGer, Deep-Q, iLQR, and IRL.

[1] https://people.eecs.berkeley.edu/~pabbeel/cs287-fa12/slides/...

[2] https://www.youtube.com/playlist?list=PLkFD6_40KJIwTmSbCv9OV...

deepnet··on Wikileaks reveals CIA’s ‘Brutal Kangaroo’ toolkit for hacking air-gapped network
There seems to be no evidence that public money is being spent on national cyber-defense, i.e. counter measures - at least there have been no links suggesting the NSA, CIA or GCHQ are working on patches.

This leaves private industry, e.g. Microsoft et al, solely responsible for defending citizens.

As there is no disclosure by the spies, Microsoft cannot patch without knowledge of the vulnerabilities - at least until there is a leak or potential leak that promts a reluctant disclosure.

This leaves one wondering is tax money being spent solely to make us less safe ?

Could Wannacries NHS (UK) devastation have been averted if academic security researchers had responsibly disclosed the potential for this vulnerability ? Surely the UK Health Minister should have been more properly briefed on the dangers to patients that his decision not to patch Windows could have ?

If these agencies come crying for more tax dollars we should seriously enquire if they have any plans to make us safer rather than continuing to backdoor their own citizens.

Surely laws being passed in the UK and USA to curtail security research the danger to the public can only increase.

Without whistleblowers and journalists such as Wikileaks we would far more at risk from cyber-attacks !

I say Wikileaks as we know journalists are being targeted by Nation State actors to hack the identities of their sources. and Wikileaks is a fine exemplar of decent Op-Sec.

deepnet··on Euro MPs back end-to-end encryption for all citizens
It is impossible to conclude this given Brexit was far from the only issue - ending austerity and saving the NHS were big priorities too.

This doesn't work so well under the UK's First Past The Post system.

If you wanted to get the Tories out then feeling one was throwing away a vote for the Lib-Dems in a seat were they are too far behind is one argument for proportional representation.

Also the Lib-Dems lost a lot of trust with the coalition with the Tories, many left the Lib Dems as they felt betrayed by this.

deepnet··on Euro MPs back end-to-end encryption for all citizens
Quite the opposite actually.

The UK is a mixed economy, we have both Capitalism and Socialism and we like it - our safety-net welfare state and the NHS save lives everyday.

EU membership is a shield for former soviet satellite states to join the West and escape Russian influence.

The main leave "argument" was £350,000,000 was being given to the EU and could save the beloved NHS - this turned out to be a downright lie.

Many of the other popular arguments were equally false and fearmongery about out of control EUrocrats legislating on the curvature of bannanas and attacks on the British sausage and that farmers would be better off outside the common market - all very jingoistic.

It has emerged that there was a lot of carefully crafted individually targeted Facebook adverts that played on voters fears derived from personal data held outside the UK (and outside UK data laws), sadly these went undocumented so we may never know the extent of the falsehoods or their level of influence.

The UK electoral commission has expressed concern over this type of campaigning as no-one knows what is being promised or how much was spent on it.

There has been an effort to try to document these dark ads for the recent election where they seem to have been much less effective.

Many Brits feel conned as the referendum was pitched as being only advisory but is now being taken as iron-clad and a 4% majority of those who offered an opinion is very little mandate to enact such major constitutional change.

As the realities for science funding, farming subsidies have kicked polls are reporting many Brexiters have changed their tune. Goldman Sax's relocation to Europe is a bellweather for the realities for the financial industry which is close enough to Tory hearts that one hopes they'll snap out of their dreams of Empire and Commonwealth.

Hopefully the staggering level of incompetence so far demonstrated by David Davis' negotiating team will be the rope that hangs them and we'll get a second real referendum.

The alternative Red White and Blue Hard Brexit promised by Teresa May is not good for anyone in the UK unless they are shorting the pound.

deepnet··on Hard Questions
Yes but not quite.

I suggested that the consumer is not forced to use Facebook's user-facing services but would have a choice.

> all those groups stayed on Facebook anyway...

I don't fully accept your assertion that all these groups stayed on Facebook, many exist elsewhere.

You may well be right about some or most but I have not seen this surveyed anywhere. Do you have a citation ?

deepnet··on Open Source Datasets
Dear Deepmind, as you have retired AlphaGo please open source the dataset of Go games used to train it.
deepnet··on Hard Questions
Facebook is unprecendentedly widely used, global in reach.

It has become for many an essential service.

Facebook ( in many countries ) is a monopoly, unassailable competitively due to the size of its network.

In days of yore, such a huge monopoly that is a utility would be broken up like Ma Bell in the 1980s or like rail, water and electricity in the UK were nationalised.

Consumer data protections are currently woefully insufficient.

I propose that such vast walled gardens should be curtailed for the public good.

Thus Facebook should be just as viewable without signup, data exportable and deletable in whole or in part.

Facebook should provide an API so competing and extending services can be freely built on it, restoring competition and innovation and niche utilities it brings.

Consumers should be free to use whatever services and programs they choose on their own data whether stored on Facebooks servers or downloaded and held privately.

Huge transnational data monopolies should not be walled gardens.

deepnet··on Learning to Cooperate, Compete, and Communicate
Peter Abeel ( OP link author with Igor Mordatch ) explains his groups work [1] as a guest lecturer for Berkeley's cs294-112 Deep Reinforcement Robotics.

[1] https://www.youtube.com/watch?v=f4gKhK8Q6mY&list=PLkFD6_40KJ...

This great talk starts with his work at OpenAI on neural net safety and adverserial images, then the OP research paper Emergence of Grounded Compositional Language in Multi-Agent Populations[2] concluding with his work ( with Andrew Ng ) reinforcement learning helicopter flight and stunt controllers from human pilots.

The OP multi-agents divide labour and apparent collaborative plans appear. That the goal seeking agents split up, appear to dance to and from from their goal, distracting the predators from their kin is to all appearance coordinated and clever.

Dawkin's Selfish Gene espouses alturism as an inevitability of genetic relatedness, the individual sacrifices but the genes persist in siblings.

In this work alturism emerges purely memetically.

The Nash equilibrium of cooperation jumps the local minima of selfishness in this prisoners dillemma.

Multi agent enviroments are difficult to learn with many false minima for the learners.

This work hints that the loose coupling of language, rather than direct sharing of memories or genes, is noisy enough to find more global solutions that appear complex or 'plan-like'.

Maybe these AI's should be considered as planners, yet in a bottom-up immediate-heuristic emergent way.

[2] https://arxiv.org/abs/1703.04908

deepnet··on Accelerating Neural Networks with Binary Arithmetic
Binarisation is not for backpropagation which requires a gradient (binary is a cliff) for the calculus backprop to work - this gradient is provided by the slope of the sigmoid.

The OP is only suggesting binarisation for the feedforward use of nets, not the learning and training phase.

from the OP : " Real valued gradients are required for SGD to work. The weights are stored in real valued accumulators and are binarized in each iteration for forward propagation and gradient computations. "

deepnet··on So, Where Are All Those Robots?
I agree, paper is absent from many offices, yet very many business still use some paper.

As the amount of businesses and transactions has increased by several orders of magnitude the overall effect is that more paper is being used than ever before.

deepnet··on So, Where Are All Those Robots?
Where too is the once promised paperless office ?

IMHO the robo jobpocalypse is as overhyped as the paperpocalypse was in the early 1980's

deepnet··on Theresa May says the internet must now be regulated after London terror attack
Deliberately weaking public security is entirely the wrong approach.

Banning encrytion makes the British public more vulnerable not less.

And whitelisting all posts and uploads quells the utility of criticism, opinion and opposition.

It seems the government don't trust the population.

Teresa May is entirely wrong in her approach and one suspects utterly ill advised.

deepnet··on First draft of the new Artificial Intelligence and Games textbook available now
Togelius is pure mental (in a good OOTB way) his evo-devo HyperNEAT approach to problems intractable to random-walk SGD learning learning leaps over snares and pitfalls, bypasses cul-de-sacs and avoids oubliettes -traversing the whole search space by a sort of teleportation - a fast global search.

Backprop can then refine the most efficient net architectures. Curiously some evolve structures akin to LSTM units.

deepnet··on Ask HN: Books you wish you had read earlier?
"From Bacteria to Bach the evolution of minds" by Daniel Dennet.

Should be called How Minds Evolve as Heirarchies of Darwinian Turing Machines ( analagously to Deep Neural Nets (Dennet cites Geoff Hinton and Edinburgh's Andy Clarke).

"working computer models have been developed that can do a good job identifying handwritten—scribbled, really—digits, involving a cascade of layers in which the higher layers make Bayesian predictions about what the next layer down in the system will “see” next; when the predictions prove false, they then generate error signals in response that lead to Bayesian revisions, which are then fed back down toward the input again and again, until the system settles on an identification (Hinton 2007). Practice makes perfect, and over time these systems get better and better at the job, the same way we do—only better" p.178 [1]

"Hierarchical, Bayesian predictive coding is a method for generating affordances galore: we expect solid objects to have backs that will come into view as we walk around them; we expect doors to open, stairs to afford climbing, and cups to hold liquid. These and all manner of other anticipations fall out of a network that doesn’t sit passively waiting to be informed but constantly makes probabilistic guesses about what it is about to receive in the way of input from the level below it, based on what it has just received, and then treating feedback about the errors in its guesses as the chief source of new information, as a way to adjust its prior expectations for the next round of guessing."

Which echoes Richard Gregory's concept of vision (or perception) as a hypothesis continually tested against input.

This is Paradigm shifting; weltanschauung shattering stuff. Dennet very clearly lays out a methodology for how all aspects of minds can evolve using heirarchical compositions of wetware robots or :

"Si, abbiamo un anima. Ma é fatta di tanti piccoli robot! (Yes, we have a soul, but it’s made of lots of tiny robots!)" p.24 [1]

[1] https://www.penguin.co.uk/books/253900/from-bacteria-to-bach...

deepnet··on 3D printing for mathematical visualisation [pdf]
George Hart has .stl files of his mathematical forms on his home page, including a diagonally sliced Menger Sponge with its surprisingly star shaped holes.

http://georgehart.com/rp/rp.html

deepnet··on Breaking Glass in Infinite Dimensions
Analogously Neural Network learning, i.e. learning modeled as a phase transition between disorder and order:

“... studies the connection between the highly non-convex loss function of a simple model of the fully-connected feed-forward neural network and the Hamiltonian of the spherical spin-glass model”

from "The Loss Surfaces of Multilayer Networks" Choromanska, Henaff, Mathieu, Arous, & LeCun [1]

[1] https://arxiv.org/abs/1412.0233

Also the physics of collapsing sand piles.

deepnet··on Europeana Collections, a Portal of 48M Free Artworks, Books, Videos, Artifacts
The search allows for filtering by licenses.

Not all content is reusable, for instance the Art Nouveau Posters [1] has :

  Free Reuse      62   cc-0, pub domain, cc-by 
  Limited Reuse  142   cc-by-nc-nd, cc-by-bc-sa
  No Reuse        64   in copyright
[1] http://www.europeana.eu/portal/en/collections/art?q=posters+...
deepnet··on 22-Year-Old Lidar Whiz Claims Breakthrough
What is the output of LIDAR ? A point cloud or stripes ?

What are the specs of the output of a LIDAR such a Velodyne or this one ?

deepnet··on Android in Cloud – An Open Source Android cloud emulator
Magic! Thanks.
deepnet··on Android in Cloud – An Open Source Android cloud emulator
To register it seems one must be a member of an organisation like a University.
deepnet··on Roboschool: open-source software for robot simulation
"However, a policy that succeeds in simulation often doesn't work when deployed on a real robot. Nevertheless, often the overall gist of what the policy does in simulation remains valid in the real world. In this paper we investigate such settings, where the sequence of states traversed in simulation remains reasonable for the real world, even if the details of the controls are not, as could be the case when the key differences lie in detailed friction, contact, mass and geometry properties"

from Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model

https://arxiv.org/abs/1610.03518

by Christiano, Shah, Mordatch, Schneider, Blackwell, Tobin, Abbeel, & Zaremba

deepnet··on Mark Zuckerberg should try living in the real world
Hilariously along these lines is "When Seinfeld met Zuckerberg".

The world weary Jerry is counterpoint to Mark's fresh faced idealism.

The bit about staff prepping Jerry not to mention Marc's 'broken arm' had me in stitches.

https://www.youtube.com/watch?v=0mkR3BkuKCI

Seinfeld : "You and I showing up would ruin everything he likes about that community."..."Your goal was to take the social experience and put it online; my goal was to take the actual human social experience and get away from that."

Zuckerberg : "Get away ?"

Seinfeld : "I would like to eliminate the entire social experience."

deepnet··on Roger Penrose on Why Consciousness Does Not Compute
"Perhaps the best way of seeing the reality, indeed the ubiquity in Nature, of reasons is to reflect on the different meanings of “why.”

The English word is equivocal, and the main ambiguity is marked by a familiar pair of substitute phrases: what for? and how come?”

“Why are you handing me your camera?” asks what are you doing this for?

“Why does ice float?” asks how come: what it is about the way ice forms that makes it lower density than liquid water?"

P48, From Bacteria to Bach, The Evolution of Minds by Dan Dennet

deepnet··on Roger Penrose on Why Consciousness Does Not Compute
Dennet is concerned with the origin of minds on Earth and how this is explicable within our current evolutionary framework.

He does not seek to explain the origins of the universe but argues that this is not necessary to understand how minds could come about.

deepnet··on Roger Penrose on Why Consciousness Does Not Compute
For the completely opposite view, listen to Daniel Dennet on The Life Scientific [1].

Dennet argues that combining Darwin's strange inversion of reason (complexity from bottom up iterative refinement) with Turing's Universal Machine provides a way of understanding how we are machines, built of mahcines, built of machines, etc. and it is the heirarchy that allows the complexity of minds to emerge.

That heirarchical iterative schemes are unexpectedly powerful is well mirrored by the recent successes of deep neural nets, and Dennet cites Hinton.

It's worth a listen and summarises his new book From Bacteria to Bach.

[1] http://www.bbc.co.uk/programmes/b08kv3y4

deepnet··on Some Android apps are using ultrasonic beacons to track users
Some sort of aftermarket high bandpass filter on the microphone will be necessary.
← PreviousPage 7 of 23Next →