3 years, 6 months ago
Use of classical neural networks for image understanding is complicated, as a rule, by big dimension of vector of entry values of neural network, large number of neurons in interfacial layers and, as a result, big costs of computing resources of training and calculation of network. The shortcomings described above are to a lesser extent inherent in convolution neural networks.
The convolution neural network (English convolutional neural network, CNN) — the special architecture of artificial neural networks offered by Jan Lekun and aimed at effective image understanding is part of technologies of deep learning (English deep leaning). This technology is constructed by analogy with the principles of work of visual cerebral cortex in which the so-called simple cages reacting to direct lines under different corners and difficult cages which reaction is connected with activation of certain set of simple cages have been open. Thus, the idea of convolution neural networks consists in alternation of convolution layers (English convolution layers) and the subsampling layers (English subsampling layers, subselection layers). 
Rice 1. Architecture of convolution neural network
The key moment in understanding of convolution neural networks is the concept of the so-called "separated" scales, i.e. the part of neurons of some considered layer of neural network can use the same weight factors. The neurons using same weight integrate in cards of signs (feature maps), and each neuron of the card of signs is connected with part of neurons of the previous layer. At calculation of network it turns out that each neuron executes convolution (convolution operation) of some area of the previous layer (determined by set of the neurons connected with this neuron). The layers of neural network constructed in a described way are called as convolution layers. In addition to, convolution layers in convolution neural network there can be subsampling layers (executing functions of reduction of dimension of space of cards of signs) and full-meshed layers (output layer, as a rule, always full-meshed). All three types of layers can alternate in random order that allows to make cards of signs of cards of signs, and it in practice means ability of recognition of difficult hierarchies of signs .
What influences quality of pattern recognition when training convolution neural networks? Having puzzled with the matter, have come across Mathew Zayler's (Matthew Zeiler) article.
3 years, 11 months ago
Once, to boil water, the person had to extract fire friction force or to cut sparks by means of silicon. But we were lucky to be born in century when instead of stamping behind firewood we click button of the electric kettle. By the way, at the time of use of silicon for such manners it would be possible and to ring out on fire.
3 years, 11 months ago
The digital future, virtual reality, animation of objects, so qualitative that is short and to confuse with reality. Until recently all this seemed impossible. We stayed at home and Doom, Dune and Lemmings were cut in primitive computer games with not less primitive computer graphics, such as Mario. Yes, my young friend! There were such games in the nineties where the background has been drawn worse than ever, and the picture resolution was so a little that all characters and badly drawn background including, have been broken, to put it mildly, on without effort the pixels seen to human eye.
4 years agoProcessing.js. The library has quickly won popularity and actively developed within several years after release. Hours on Bezier curves about which on Habré wrote a few days ago, have been created by means of Processing.js.
Example of the animation created by means of the Processing language
4 years ago
The actual standard for work with the mathematical formulas in web written in the TeX format is the MathJax library. The platform of online training of Khan Academy has set of the courses devoted to mathematics in the arsenal. One of important points of mission of academy? to make quality education available to all, irrespective of geographical and financial position therefore fast work of the site on weak old computers and mobile devices is very essential to Khan Academy. For this reason developers of academy have decided to write own library for rendering of TeX to HTML which in the long term could replace completely MathJax and would work much quicker.
The KaTeX library is still happy is young (the current version? 0.1.0) and for the present it cannot be compared to MathJax on functionality. The current list of the supported functions is available in Wicky of the project. KaTeX supports less output formats in comparison with MathJax which is able to render formulas in HTML+CSS, SVG and MathML. In KATEH the output in HTML is implemented, support of MathML is planned in the future. KaTeX does not support old (is more senior 8.0) the IE version.
4 years, 2 months ago
Element <time;> represents date, time or time frame provided in machine-readable format. It can be useful to creation of the schedules, archives or other functions connected with time. WordPress uses this element in basic subject of design. One more example of use <time;> ? Reddit:
The course of life of this element was hard. It was added to the HTML5 specification in 2009. Two years later, in 2011, he was moved away and replaced with much wider <data;>.; However same year it was returned and added new opportunities. Now it can be used surely.
4 years, 2 months ago
Whether it was necessary to you, having followed the link to the interesting video to Youtube, to find out, what for the sake of several seconds where really there is something interesting, you have just spent neskolk minutes for contemplation of the absolutely useless? garbage? only because the author of video has laid out entirely the file from the video recorder or the smartphone? The quantity of video cameras promptly grows, and number of the people capable at least constants need to cut off couple of excess fragments, seemingly. And problem not only in several minutes of time killed on the Internet? after all there are also more serious cases, for example, of ten and hundreds of hours of video from surveillance cameras which sometimes should be viewed to solve crime.
Scientists from Carnegie Mellon University have developed effective algorithm of selection of the most interesting fragments of video on the basis of machine learning. The new algorithm called by them? LiveLight? considerably exceeds analogs on the speed and quality of work. LiveLight selects characteristic fragments of video and sotavlyat them? dictionary?, and then tries to predict the following frame on their basis. If it works well with sufficient accuracy rating, it means that the frame does not add almost any new information and it can be excluded. Unlike? the mechanical? the approaches reacting to any movement in frame or sharp brightness variation, color or contrast, LiveLight is rather universal? it well works and at video, removed by the motionless camera, and at amateur shooting as the shivering smartphone.
4 years, 3 months ago
Anna Makosinski, the young inventor awarded awards for the invention of thermoelectric small lamp has shown the new version of the invention: the nalobny small lamp working from heat of human body.