Group items tagged

Finally a massively multi-core general-purpose architecture.

Well, nice, but I wonder how much cache per core will be available... With 48 cores a single memory bus becomes nothing more than one big (small? :) ) bottleneck.

Apparently they have separated L2 cache per-tile (i.e., every two processors) and a high speed bus connecting the tiles. As usual, whether it will be fast enough will depend from the specific applications (which BTW is also true for other current multi-core architectures). The nice thing is of course that porting software to this architecture will be one order of magnitude less difficult than going to Tesla/Fermi/CELL architectures. Also, this architecture will also be suitable for other tasks than floating point computations (damn engineers polluting computer science :P) and it has the potential to be more future-proof than other solutions.

This has been around for over a year. The current trend in deep learning is "deeper is better". But a consequence of this is that for a given network depth, we can only feasibly evaluate a tiny fraction of the "search space" of NN architectures. The current approach to choosing a network architecture is to iteratively add more layers/units and keeping the architecture which gives an increase in the accuracy on some held-out data set i.e. we have the following information: {NN, accuracy}. Clearly, this process can be automated by using the accuracy as a 'signal' to a learning algorithm. The novelty in this work is they use reinforcement learning with a recurrent neural network controller which is trained by a policy gradient - a gradient-based method. Previously, evolutionary algorithms would typically be used. In summary, yes, the results are impressive - BUT this was only possible because they had access to Google's resources. An evolutionary approach would probably end up with the same architecture - it would just take longer. This is part of a broader research area in deep learning called 'meta-learning' which seeks to automate all aspects of neural network training.

Btw that techxplore article was cringing to read - if interested read this article instead: https://research.googleblog.com/2017/05/using-machine-learning-to-explore.html

"The concept of the cosmic web-viewing the universe as a set of discrete galaxies held together by gravity-is deeply ingrained in cosmology. Yet, little is known about architecture of this network or its characteristics. Our research used data from 24,000 galaxies to construct multiple models of the cosmic web, offering complex blueprints for how galaxies fit together. These three interactive visualizations help us imagine the cosmic web, show us differences between the models, and give us insight into the fundamental structure of the universe."

This December, two Swiss architects and an Italian robotics engineer will, for the first time, build a tower solely by flying robots.

very nice!

VERY nice! one of the promised apps of "swarms" at last demonstrated...and i was beginning to lose hope! (pity this article goes under this "singularity" website...)

This is related to the "Verified Software" item in NewScientist's list of ideas that will change science. At the link below you'll find the text of the patents being auctioned: http://icapoceantomo.com/item-for-sale/exclusive-license-related-improved-methodology-formally-developing-control-systems :) Patent #7,627,538 ("Swarm autonomic agents with self-destruct capability") makes for quite an interesting read: "This invention relates generally to artificial intelligence and, more particularly, to architecture for collective interactions between autonomous entities." "In some embodiments, an evolvable synthetic neural system is operably coupled to one or more evolvable synthetic neural systems in a hierarchy." "In yet another aspect, an autonomous nanotechnology swarm may comprise a plurality of workers composed of self-similar autonomic components that are arranged to perform individual tasks in furtherance of a desired objective." "In still yet another aspect, a process to construct an environment to satisfy increasingly demanding external requirements may include instantiating an embryonic evolvable neural interface and evolving the embryonic evolvable neural interface towards complex complete connectivity." "In some embodiments, NBF 500 also includes genetic algorithms (GA) 504 at each interface between autonomic components. The GAs 504 may modify the intra-ENI 202 to satisfy requirements of the SALs 502 during learning, task execution or impairment of other subsystems."

NVIDIA Fermi preview...

sllllll, nerd porn ;) awesome

A control and design architecture inspired by bio-arms (i.e. fluffy elastic things)

Switching between fear and defence with just a few neurons. Reminds me of my suggestion of including emotions in control architectures. Anyone interested? TSe

12 new NIAC 1 studies - many topics familiar to us ... please have a look at those closest to your expertise to see if there is anything new/worth investigating (and in general to be knowledgeable on them since we will get questions sooner or later on them)
Principal Investigator Proposal Title Organization City, State, Zip Code
Atchison, Justin Swarm Flyby Gravimetry Johns Hopkins University Baltimore, MD 21218-2680
Boland, Eugene Mars Ecopoiesis Test Bed Techshot, Inc. Greenville, IN 47124-9515
Cash, Webster The Aragoscope: Ultra-High Resolution Optics at Low Cost University of Colorado Boulder, CO 80309-0389
Chen, Bin 3D Photocatalytic Air Processor for Dramatic Reduction of Life Support Mass & Complexity NASA ARC Moffett Field, CA 94035-0000
Hoyt, Robert WRANGLER: Capture and De-Spin of Asteroids and Space Debris Tethers Unlimited Bothel, WA 98011-8808
Matthies, Larry Titan Aerial Daughtercraft NASA JPL Pasadena, CA 91109-8001
Miller, Timothy Using the Hottest Particles in the Universe to Probe Icy Solar System Worlds John Hopkins University Laurel, MD 20723-6005
Nosanov, Jeffrey PERISCOPE: PERIapsis Subsurface Cave OPtical Explorer NASA JPL Pasadena, CA 91109-8001
Oleson, Steven Titan Submarine: Exploring the Depths of Kraken NASA GRC Cleveland, OH 44135-3127
Ono, Masahiro Comet Hitchhiker: Harvesting Kinetic Energy from Small Bodies to Enable Fast and Low-Cost Deep Space Exploration NASA JPL Pasadena, CA 91109-8001
Streetman, Brett Exploration Architecture with Quantum Inertial Gravimetry and In Situ ChipSat Sensors Draper Laboratory Cambridge, MA 02139-3539
Wiegmann, Bruce Heliopause Electrostatic Rapid Transit System (HERTS) NASA MSFC Huntsville, AL 35812-0000

Eh, the swarm flyby gravimetry is very similar to the "measuring gravitational fields" project I proposed in the brewery

love this one ... it seems to take physicist to explain to the AI crowd what they are actually doing ... Deep learning is a broad set of techniques that uses multiple layers of representation to automatically learn relevant features directly from structured data. Recently, such techniques have yielded record-breaking results on a diverse set of difficult machine learning tasks in computer vision, speech recognition, and natural language processing. Despite the enormous success of deep learning, relatively little is understood theoretically about why these techniques are so successful at feature learning and compression. Here, we show that deep learning is intimately related to one of the most important and successful techniques in theoretical physics, the renormalization group (RG). RG is an iterative coarse-graining scheme that allows for the extraction of relevant features (i.e. operators) as a physical system is examined at different length scales. We construct an exact mapping from the variational renormalization group, first introduced by Kadanoff, and deep learning architectures based on Restricted Boltzmann Machines (RBMs). We illustrate these ideas using the nearest-neighbor Ising Model in one and two-dimensions. Our results suggests that deep learning algorithms may be employing a generalized RG-like scheme to learn relevant features from data.

Swarm-based fabrication of interwoven composite tubes via a fully autonomous, cooperative system can help create architectural-scale structures in effective and efficient ways, including in remote environments.

New optics based computer architecture

one more step in the right direction for thermoelectrics ...

As we are allowed to think ...

Charles Stross... this one was interesting, but I think the best essay he has written in the past few months is this one: How low (power) can you go? -- On the subject of ubiquitous computing devices and urban architecture

Cute info-graphics on different machine learning architectures

Ideas for ESTEC 2 ...

A 7x7x3m large unfolding structure from the car trunk. "Before the design is ready to roll, the team still needs to tune up the solar components, a challenge made more difficult by the pavilion's inherent mobility, making it impossible to gauge how sunlight will fall on it."

Recharge in seconds and efficiently store power for weeks between charges. Added bonus is the cheap and abundant components needed. One of the applications they foresee is to attach such a super-capacitor to the back of solar panels to store the power and discharge this during the night

very nice indeed - is this already at a stage where we should have a closer look at it? what you think? With experience in growing carbon nanostructures, Pint's group decided to try to coat the porous silicon surface with carbon. "We had no idea what would happen," said Pint. "Typically, researchers grow graphene from silicon-carbide materials at temperatures in excess of 1400 degrees Celsius. But at lower temperatures - 600 to 700 degrees Celsius - we certainly didn't expect graphene-like material growth." When the researchers pulled the porous silicon out of the furnace, they found that it had turned from orange to purple or black. When they inspected it under a powerful scanning electron microscope they found that it looked nearly identical to the original material but it was coated by a layer of graphene a few nanometers thick. When the researchers tested the coated material they found that it had chemically stabilized the silicon surface. When they used it to make supercapacitors, they found that the graphene coating improved energy densities by over two orders of magnitude compared to those made from uncoated porous silicon and significantly better than commercial supercapacitors. Transmission electron microscope image of the surface of porous silicon coated with graphene. The coating consists of a thin layer of 5-10 layers of graphene which filled pores with diameters less than 2-3 nanometers and so did not alter the nanoscale architecture of the underlying silicon. (Cary Pint / Vanderbilt) The graphene layer acts as an atomically thin protective coating. Pint and his group argue that this approach isn't limited to graphene. "The ability to engineer surfaces with atomically thin layers of materials combined with the control achieved in designing porous materials opens opportunities for a number of different applications beyond energy storage," he said.