SciTech #ScienceSunday Digest - 43/2016.
SciTech #ScienceSunday Digest - 43/2016.
Permalink here: http://www.scitechdigest.net/2016/10/reservoir-computing-analogue.html
Reservoir computing, Analogue neuromorphic chips, Network neuroscience, Eggs from skin cells, Conversational speech recognition, Fast FPGA chips, Diamond anvils, Smart 3D printers, Big data automation, Full colour epaper.
1. Reservoir Computing
Reservoir computing aims to perform useful computations on different bulk materials by exploiting the basic properties of physical systems; the basic idea is to input or stimulate the material and measure the output or change in state, which counts as a calculation - string these together and you can perform computations. A recent reservoir computer was built out of a bucket or water, while observations suggest the brain functions like a reservoir computer https://theconversation.com/theres-a-way-to-turn-almost-any-object-into-a-computer-and-it-could-cause-shockwaves-in-ai-62235. In related news artificial reservoir computing algorithms combined with backpropagation algorithms results in an analogue computer with some superior learning algorithms.
2. Analogue & Neuromorphic Chips
Analogue computing is picking up steam as it promises to provide a significant range of benefits to neuromorphic computing architectures http://spectrum.ieee.org/automaton/robotics/artificial-intelligence/analog-and-neuromorphic-chips-will-rule-robotic-age. Recent examples include analogue circuits that better see and hear while consuming a fraction of the power, computing systems that are much more resistant to noise, and deep neural networks using analogue approaches that use 100 times less energy to run. Watch this space; we haven’t seen anything yet.
3. Network Neuroscience & Control
Network control theory is seeking to go beyond communications, gene regulatory networks, and other systems to being applied to control the brain via network neuroscience https://www.technologyreview.com/s/602695/how-network-neuroscience-is-creating-a-new-era-of-mind-control/. This starts with simple manipulations that inject energy into one part of the network to alter activity in another part, for example the deep brain stimulation techniques employed with Parkinson’s disease patients. Combined with connectome data it is becoming apparent that the brain employs a range of different control strategies, each of which is a target for directed control to, for example, shift the brain into desirable patterns of activity.
4. Converting Skin Cells into Eggs into Adult Animals
For the first time skin cells from mice have been reprogrammed in a petri dish to form viable egg cells that have then been fertilised to produce healthy animals that proceeded to successfully birth a second generation of mice http://www.nature.com/news/mouse-eggs-made-from-skin-cells-in-a-dish-1.20817. One limitation is the need for an ovary-like support of cells isolated from ovaries to be present with the cells being transformed; the group are hoping to identify and create artificial reagents that would finally obviate the need for this. Combined with earlier techniques for producing sperm cells from skin cells the complete reproductive technology stack is coming together to allow a wide range of flexible reproductive (and industrial) strategies to be employed.
5. Human Level Conversational Speech Recognition
Microsoft demonstrated neural network speech recognition software that can transcribe conversational speech at human levels of proficiency and performance http://blogs.microsoft.com/next/2016/10/18/historic-achievement-microsoft-researchers-reach-human-parity-conversational-speech-recognition/#sm.0015hf9cv11bqfiqqwu1iixt8xitf. The system produces an error rate of 5.9% that is the same or less than professionals who transcribe conversational speech. This is expected to feed directly into Cortana and other software products. This advance was announced as another Microsoft Research group won first prize in an image segmentation competition for delineating specific objects in images.
6. Fast FPGA Optimisation Chips
Fujitsu demonstrated a new FPGA flexible circuit architecture for solving combinatorial optimisation problems that can perform these computations 10,000 times faster than a conventional computer http://www.fujitsu.com/global/about/resources/news/press-releases/2016/1020-02.html. The chips are made with conventional semiconductors, include features that allow the optimisation computations to escape local minimums, and running software processes called simulated annealing. The parallelisation incorporated into the design allows a pathway from the demonstration problems with 1,024 bits to 100,000 or more bits by 2018 that will start to allow practical implementation.
7. Latest Diamond Anvils
New microanvils made of diamond and adorned with a nanocrystalline diamond pillar measuring 30 micrometers wide and 15 micrometers tall has achieved some of the highest experimental pressures ever http://www.uab.edu/news/innovation/item/7702-working-under-pressure-diamond-micro-anvils-made-by-uab-will-produce-immense-pressures-to-make-new-materials. The anvils reached 264 gigapascals, about 75% of the pressure found at the centre of the Earth, and the ultimate goal of the group is to improve the anvils to reach 1 terapascal of pressure, or 147 million pounds per square inch - the pressure at the centre of Saturn. The nanocrystalline diamond anvils showed no sign of deformation and survived the immense pressures; applications include new materials analysis and development.
8. Intelligent 3D Printers
Ai Build is retrofitting robotic arms with 3D printers and AI algorithms to create machines that can see, create, and learn from mistakes http://www.digitaltrends.com/cool-tech/ai-build-wants-to-change-the-way-we-build-the-future/. The key innovation was attaching cameras to the arms with machine vision algorithms to analyse structures as they were being printed, establishing a feedback loop between the physical and virtual environments. The arm detected defects and compensated for them in later layers, and was able to print much larger complex structures about twice as quickly as a result. Partnerships with NVIDIA and robot manufacturer KUKA helped the startup make progress in this area.
9. Further Automating Big Data Analysis
A powerful automated big data analysis system has progressed since last year, moving beyond mere automated feature set selection to include automatic data presentation and specification of problems, and to perform in days what normally took months http://news.mit.edu/2016/automating-big-data-analysis-1021. The approach employs a new software language called Trane, works well with time series data, and significantly speeds up the process of finding what questions and problems are worth asking of the data. In tests on previous time-consuming work performed by humans the system reproduced every question that the researchers had asked of particular data sets and proposed hundreds of others that had not been considered.
10. Full Colour Flexible Electronic Paper
A new micrometer thin polymer material provides the basis for full colour flexible electronic paper capable of providing the full range of colours that a conventional LED display is capable of while needing significantly less energy than a Kindle tablet http://www.chalmers.se/en/departments/chem/news/Pages/Bendable-electronic-paper-shows-full-colour-scale-.aspx. This is a proof of concept with core pixels having been built, but a final product will require significant scale-up to large high pixel density screens and also drastically reduce the amount of gold used in manufacturing the films.
SciTech Tip Jar: http://www.scitechdigest.net/p/donate.html
Permalink here: http://www.scitechdigest.net/2016/10/reservoir-computing-analogue.html
Reservoir computing, Analogue neuromorphic chips, Network neuroscience, Eggs from skin cells, Conversational speech recognition, Fast FPGA chips, Diamond anvils, Smart 3D printers, Big data automation, Full colour epaper.
1. Reservoir Computing
Reservoir computing aims to perform useful computations on different bulk materials by exploiting the basic properties of physical systems; the basic idea is to input or stimulate the material and measure the output or change in state, which counts as a calculation - string these together and you can perform computations. A recent reservoir computer was built out of a bucket or water, while observations suggest the brain functions like a reservoir computer https://theconversation.com/theres-a-way-to-turn-almost-any-object-into-a-computer-and-it-could-cause-shockwaves-in-ai-62235. In related news artificial reservoir computing algorithms combined with backpropagation algorithms results in an analogue computer with some superior learning algorithms.
2. Analogue & Neuromorphic Chips
Analogue computing is picking up steam as it promises to provide a significant range of benefits to neuromorphic computing architectures http://spectrum.ieee.org/automaton/robotics/artificial-intelligence/analog-and-neuromorphic-chips-will-rule-robotic-age. Recent examples include analogue circuits that better see and hear while consuming a fraction of the power, computing systems that are much more resistant to noise, and deep neural networks using analogue approaches that use 100 times less energy to run. Watch this space; we haven’t seen anything yet.
3. Network Neuroscience & Control
Network control theory is seeking to go beyond communications, gene regulatory networks, and other systems to being applied to control the brain via network neuroscience https://www.technologyreview.com/s/602695/how-network-neuroscience-is-creating-a-new-era-of-mind-control/. This starts with simple manipulations that inject energy into one part of the network to alter activity in another part, for example the deep brain stimulation techniques employed with Parkinson’s disease patients. Combined with connectome data it is becoming apparent that the brain employs a range of different control strategies, each of which is a target for directed control to, for example, shift the brain into desirable patterns of activity.
4. Converting Skin Cells into Eggs into Adult Animals
For the first time skin cells from mice have been reprogrammed in a petri dish to form viable egg cells that have then been fertilised to produce healthy animals that proceeded to successfully birth a second generation of mice http://www.nature.com/news/mouse-eggs-made-from-skin-cells-in-a-dish-1.20817. One limitation is the need for an ovary-like support of cells isolated from ovaries to be present with the cells being transformed; the group are hoping to identify and create artificial reagents that would finally obviate the need for this. Combined with earlier techniques for producing sperm cells from skin cells the complete reproductive technology stack is coming together to allow a wide range of flexible reproductive (and industrial) strategies to be employed.
5. Human Level Conversational Speech Recognition
Microsoft demonstrated neural network speech recognition software that can transcribe conversational speech at human levels of proficiency and performance http://blogs.microsoft.com/next/2016/10/18/historic-achievement-microsoft-researchers-reach-human-parity-conversational-speech-recognition/#sm.0015hf9cv11bqfiqqwu1iixt8xitf. The system produces an error rate of 5.9% that is the same or less than professionals who transcribe conversational speech. This is expected to feed directly into Cortana and other software products. This advance was announced as another Microsoft Research group won first prize in an image segmentation competition for delineating specific objects in images.
6. Fast FPGA Optimisation Chips
Fujitsu demonstrated a new FPGA flexible circuit architecture for solving combinatorial optimisation problems that can perform these computations 10,000 times faster than a conventional computer http://www.fujitsu.com/global/about/resources/news/press-releases/2016/1020-02.html. The chips are made with conventional semiconductors, include features that allow the optimisation computations to escape local minimums, and running software processes called simulated annealing. The parallelisation incorporated into the design allows a pathway from the demonstration problems with 1,024 bits to 100,000 or more bits by 2018 that will start to allow practical implementation.
7. Latest Diamond Anvils
New microanvils made of diamond and adorned with a nanocrystalline diamond pillar measuring 30 micrometers wide and 15 micrometers tall has achieved some of the highest experimental pressures ever http://www.uab.edu/news/innovation/item/7702-working-under-pressure-diamond-micro-anvils-made-by-uab-will-produce-immense-pressures-to-make-new-materials. The anvils reached 264 gigapascals, about 75% of the pressure found at the centre of the Earth, and the ultimate goal of the group is to improve the anvils to reach 1 terapascal of pressure, or 147 million pounds per square inch - the pressure at the centre of Saturn. The nanocrystalline diamond anvils showed no sign of deformation and survived the immense pressures; applications include new materials analysis and development.
8. Intelligent 3D Printers
Ai Build is retrofitting robotic arms with 3D printers and AI algorithms to create machines that can see, create, and learn from mistakes http://www.digitaltrends.com/cool-tech/ai-build-wants-to-change-the-way-we-build-the-future/. The key innovation was attaching cameras to the arms with machine vision algorithms to analyse structures as they were being printed, establishing a feedback loop between the physical and virtual environments. The arm detected defects and compensated for them in later layers, and was able to print much larger complex structures about twice as quickly as a result. Partnerships with NVIDIA and robot manufacturer KUKA helped the startup make progress in this area.
9. Further Automating Big Data Analysis
A powerful automated big data analysis system has progressed since last year, moving beyond mere automated feature set selection to include automatic data presentation and specification of problems, and to perform in days what normally took months http://news.mit.edu/2016/automating-big-data-analysis-1021. The approach employs a new software language called Trane, works well with time series data, and significantly speeds up the process of finding what questions and problems are worth asking of the data. In tests on previous time-consuming work performed by humans the system reproduced every question that the researchers had asked of particular data sets and proposed hundreds of others that had not been considered.
10. Full Colour Flexible Electronic Paper
A new micrometer thin polymer material provides the basis for full colour flexible electronic paper capable of providing the full range of colours that a conventional LED display is capable of while needing significantly less energy than a Kindle tablet http://www.chalmers.se/en/departments/chem/news/Pages/Bendable-electronic-paper-shows-full-colour-scale-.aspx. This is a proof of concept with core pixels having been built, but a final product will require significant scale-up to large high pixel density screens and also drastically reduce the amount of gold used in manufacturing the films.
SciTech Tip Jar: http://www.scitechdigest.net/p/donate.html
Comments
Post a Comment