Release date: 2016-09-19
During the 3rd Shanghai International Science and Technology Products Expo, Xijing Technology R&D team of China's brain-like artificial intelligence startup company realized the first-ever "On-Chip Learning" in the world through technology demonstration.
It is understood that in the field of machine learning, "on-chip learning" is a big breakthrough in the current "On-Line Learning", which means that machine learning can be done directly on the hardware side. Through the continuous "self-learning" and "self-improvement" at the hardware level, high efficiency and low power consumption are realized. In the future, whether in mobile smart terminals, home smart terminals or big data dedicated servers, "on-chip learning" will show its unique and powerful advantages.
"On-chip" independent learning style classification
Abstract painting, Baroque, flower, interior, portrait, minimalism, pop art, surrealism... a variety of paintings and styles, dazzling.
For ordinary people, even if you infiltrate the art for many years, it is not easy to separate the painting genre at a glance. Can the machine do all this?
Since 2014, when Sergey Karayev and others of the California Institute of Technology have collected wiki-paintings, many teams around the world have joined the rapid identification and classification of machines to test the results of machine learning.
Recently, Xijing Technology's Neuromorphic Laboratory (WNR) has achieved “on-chip learning†for the first time on a global scale, and tested this technology directly on the Wiki-paintings collection. The results show that the machine can automatically complete the classification of more than 1000 pictures in just 1 second, and the correct rate is close to 100%.
At the exhibition site, Xijing Technology officially demonstrated the “on-chip learning†process in the WIKI system. After clicking on a few styles of painting and painting on the hardware side of the film, we hope to find the selected style and painting from the 10,000 oil paintings correctly and quickly.
During the period, the chip can interrupt learning at any time to test the learning effect. As a result, we found that as the progress of chip learning progresses, the correct rate of oil painting recognition and classification will gradually increase until after a complete learning of 8 to 10 seconds, the recognition accuracy rate is close to 100%.
"On-chip learning" has been slow in research
The purpose of machine learning is to make machines have the ability to learn, recognize, and understand things similar to humans.
From a global perspective, machine learning currently uses "online learning", which is a sequential, sequential learning process that continually modifies models and optimizes them. Many companies have tasted the sweetness of “online learning†and have made breakthroughs in artificial intelligence in the fields of vision, sound and big data.
At the same time, the researchers have not stopped the study of "on-chip learning", because as long as this technology is implemented, learning and testing can be done directly on the hardware, so that the development of the basic algorithm is iterative, and the product upgrade has a shorter cycle. And higher efficiency and consume less energy.
In fact, since the 1990s, a lot of research has focused on how to embed algorithms into the theoretical study of hardware, but at that time it was only for very simple waveforms, signal processing.
After 2010, due to the continuous improvement of the research and development costs of maintaining Moore's Law hardware, it stimulated the research of specialized chips in the academic industry, and began to have a high-level language interface for configuring hardware.
However, "on-chip learning" is still slow in research in the harsh hardware environment, and there is no demonstration of the actual problem, which has been stuck in the theoretical research stage.
At this stage, deploying machine algorithms in hardware is more "online learning", that is, using large-scale high-performance computing platforms for model training, and then deploying the trained models directly in hardware.
Four unique advantages of “on-chip learningâ€
According to reports, the "on-chip learning" realized by Xijing Technology, for example, is like a student who is always following a "personal education" and learning the training results directly on the chip. The biggest advantage is the real realization. "Self-learning, self-real-time improvement" in the case of "no network".
"Online learning" is like a student going to the "school" class regularly, going home to do homework to test the learning effect, and then instilling the optimized model in the hardware, each new learning needs to be re-networked, clouded, etc. Transfer and migrate.
Secondly, because "on-chip learning" can achieve localized learning, it helps the machine to greatly improve efficiency and speed up computing. In the case of relatively harsh or limited network environments, "on-chip learning" chips consume less bandwidth and traffic, significantly reducing the cost of communication for cloud servers and taking less time. After all, "online learning" requires a higher quality network environment, which causes delays in data transmission.
Google's AlphaGo is a typical example of "online learning." In this year's game with Li Shishi, AlphaGo's server is connected to the game room in Seoul, South Korea through the Google Cloud service in the Midwest. The Google headquarters team must ensure that AlphaGo is connected to Google's servers. smoothly.
In addition, the "online learning" algorithm is stored in the background server or the cloud, and generates high power consumption when the algorithm is updated or used. If the algorithm is localized, the power consumption can be greatly saved.
It is reported that the power consumption of China's data center in 2015 is 100 billion kWh, and the annual electricity consumption exceeds 1.5% of the total electricity consumption of the whole society. Some large companies can even reach billions of dollars in data center energy costs.
If "on-chip learning" is used in the data center, the data can be processed locally, and the power consumption cost of the server operation will be greatly reduced, which will greatly reduce the company's electricity bills, and more importantly, improve big data mining. The quality and efficiency of video stream processing.
Especially worth mentioning is that "on-chip learning" makes terminal handheld devices more "smart" possible, and the intelligence level of mobile phones, glasses and drones can reach a new level, achieving depth in the "no network" state. Learning, but also making minicomputer miniaturization possible.
Take gene sequencing as an example. Because it involves a lot of calculations, it relies on supercomputer processing. For us, relying on on-chip learning can save a lot of sorting time, the hospital can be used as a miniaturized computing terminal to perform genetic sequencing in the hospital in advance.
For example, the walking video of Boston-powered humanoid robots that many people have seen, the stability of robot walking is shocking, thanks in large part to its powerful algorithms. In order to meet the energy consumption of this algorithm, robots are learning. Not only do you have to carry a bulky battery, but you have to plug in a data cable to connect to a large processor. And "on-chip learning" directly learning on the hardware side can help the robot to be more dexterous and rely on "intellectual intelligence" to continuously learn and improve its ability.
In terms of home smart terminals, intelligent routing and smart office servers can implement localization of algorithms because of the “on-chip learning†rather than the “online learning†in the past, which ensures the security and privacy of data.
Thanks to the addition of “on-chip learningâ€, the future of both artificial data such as the deep mining of big data and the monitoring and analysis of video streaming in smart cities will become convenient and intelligent, and this will undoubtedly accelerate. The application process of artificial intelligence.
Source: People's Network
Rapid test kits are designed for use where a preliminary screening test result is required and are especially useful in resource-limited settings.
• High quality
• Quick (10 minutes to 2 hours) and easy to perform .
• Tests based on agglutination, immuno-chromatographic, immuno-dot, and/or immuno-filtration techniques.
• Are suitable for individual or a limited number of samples, and require little or no additional equipment, which make them more economical than ELISAs in fully equipped laboratories.
• Possibility to store at room temperature for an extended period of time.
• Same-day results provide timely treatment possibility.
[Test Principle]
Urobilinogen:Urobilinogen with diazonium salt produce red violet dyes in strong acid medium.
Bilirubin:The direct bilirubin with dichlorobenzene diazonium produce azo dyes in acid medium.
Ketone:The acetoacetic acid and sodium nitroprusside cause reaction in alkaline medium,which produces red violet compound.
Blood:Hemoglobin acts as peroxidase. It can cause peroxidase release neo-ecotypes oxide (O).(O) oxidizes the indicator and make the color change subsequently.
Protein:It is based on a specific pH indicator attracted by cation on protein molecule,the indicator further ionized and make the color change.This phenomenon is called protein-error-of -indicator principle.
Nitrite:Nitrite and aromatic amine are diazotized to form a diazonium compound.The diazonium compound reacts with tetrahydrobenzo(h)quinolin 3-phenol produce the red azo dye.
Leukocyte:Pyrrole amino acid ester produce free phenol under the hydrolysis of esterase in neutrophile granulocyte,the free phenol couple with phenyl diazonium salt produce purple azo dyes.
Glucose:The glucose oxidized by glucose oxidase catalyzes the formation of glucuronic acid and peroxide hydrogen.Peroxide hydrogen releases neo-ecotypes oxide(O)under the function of peroxidase.(O)oxidizes tetramethyl benzidine, which make the color change.
Specific Gravity:Methyl vinyl ether and maleic acid copolymer is weak acid (-COOH) exchanger,the M+ cation (the major is Na+) in urine reacts with the exchanger and release hydrogen ion,hydrogen ion reacts with indicator produce color change.
pH:The method of pH indicator is applied.
Ascorbic Acid:Ascorbic acid deoxidizes the 2.6-dichlorphenolindophenol dye into colorless in alkaline medium.
Microalbumin:Sulfone phthalein dye has high sensitivity to microalbumin by the method of protein error.
Creatinine:Creatinine with 3,5-Dinitrobenoic acid produce violet compound in alkaline medium.
Calcium:Calcium ion reacts with methyl bromothymol blue produce color change in alkaline medium.
Rapid Test Kit,Drug Abuse Test Kit,Rapid Diagnostic kit,Fast Test Kit
Changchun ZYF science and technology CO.,LTD , https://www.zyf-medical.com