<html> <h1> Make a neural network </h1> <h2> Shrink the neural network up to 10 times </h2> <ul> Make the neural network 50% smaller - $ 2000 <br> ... and so on <strong> up to 10 times </strong> </ul> <h2> Raise Test Accuracy to 99.5% of your neural network </h2> <ul> <strong> Increase Test set accuracy: </strong> <br> </ul> <ul> from 84% to 86% - $ 3000 <br> from 86% to 88% - $ 10000 <br> ... and so on <strong> up to 99.5% </strong> when the neural network is doing better than person. <br> </ul> <ul> tel: +79633123320, WhatsApp +79269319322 <br> Telegram @shatatoy or +79269319322 <br> https://vk.com/shatatoy </ul> <ul> Zhuikov Alexander - Tatoy (Shatatoy - @ Syrius4 - Τατόι) </ul> <ul> <br> Modeling and developing new neural networks. <br> Direction: <br> Robots, systems, speech, text, video. </ul> <ul> My technologies allow me to make <strong> neurons </strong> and derivatives, <strong> which are 5 to 10 years ahead </strong> the technologies of many companies. </ul> <p> Example </p> <ul> Recognition of voice commands. <br> <br> Summary: <em> <br> My Model1 is <strong> 11 times smaller </strong> than the one in the manual (count_params 148,645 versus 1,625,611) <strong> and significantly better in all respects </strong>. </em> </ul> <ul> Metrics for this tutorial as of 3/21/2021: <br> <a href="https://www.tensorflow.org/tutorials/audio/simple_audio" rel="nofollow"> https://www.tensorflow.org / tutorials / audio / simple_audio </a> <br> Total params: 1,625,611 <br> val_loss: 0.4848 - val_accuracy: 0.8388 <br> Test set accuracy: 84% <figure> <img src = "https://habrastorage.org/getpro/moikrug/uploads/redactor_image/10042021/images/3f23227df73938c84d94fba9a26753cc.png" data-image = "ves8qf2blynu"> </figure> <figure> <img src = "https://habrastorage.org/getpro/moikrug/uploads/redactor_image/10042021/images/e861cf87a9e5758190f3e23455ddc570.png" data-image = "0v6sa48vu1ah"> </figure> <ol> <strong> <br> </strong> </ol> <br> <strong> Model1 </strong> * collected by my robot automatically: * • count_params 148.645 • evaluate loss 0.3729 - evaluate accuracy 0.88 • Test set accuracy 84.62% <figure> <img src = "https://habrastorage.org/getpro/moikrug/uploads/redactor_image/10042021/images/0d7a4a95526f6554c3c5ce06ba89d7c0.png" data-image = "esjbpdukezjp"> </figure> <br> <strong> Model2 </strong> * more, collected by my robot automatically * • count_params: 3,003,038 • evaluate loss: 0.2823 - evaluate accuracy: 0.9087 • Test set accuracy: 89.00% <figure> <img src = "https://habrastorage.org/getpro/moikrug/uploads/redactor_image/10042021/images/5d8a56d35e01de249d867c24873fbcdd.png" data-image = "y70k49oheavg"> </figure> <figure> <img src = "https://habrastorage.org/getpro/moikrug/uploads/redactor_image/10042021/images/b7d35b6fecf284ee52205c4ee9a26cba.png" data-image = "gcyqw18htk7r"> </figure> <br> loss - less is better <br> accuracy, val_accuracy - more is better. <br> Test set accuracy - checking on data that the machine did not see, more is better. <br> count_params is the number of parameters of the neural network. less is much better. <ul> The same 6400 training sound files / 6400/800 <br> The algorithm is universal, from 2020/04 and now it is different. Suitable for any neuron. This is a simple audio example, quick. This is not aerobatics. <br> I have a lot of algorithms. very different. More often algorithms are not for collecting, but for creating special neurons, extended, universal, etc. <br> Creation of models, kernels, and systems on custom neural networks. <br> Can be used in the: - oil and gas industry - banking industry - search engines - medicine - air and rail freight - games - bots - further development of universal or 'strong' AI - and others. The scope is very broad <br> AI - the bias of its own research: - cognitive science, - evolution of information transfer and retention. - A new generation of decision-making tools. - medicine, - rejuvenation, - life extension, - health, <br> current from 06/03/22021. 1. Creation of evolutionary, genetic, universal algorithms. 2. Autonomy of neural networks, including development, evolution. 3. Neural networks which are not static, changeable, mutating, etc. <br> New development theme 08.11.2021 - Acceleration of Civilization. - Some areas of Acceleration: a. New generation and evolution of the text, speech. b. A revolution in the transmission and preservation of information, manuals. c. Ways of laying out and remembering information. d. (private) e. (private) f. New Wikipedia. - Other areas of Acceleration (private). <br> Neural networks are not static, changing, mutating, etc. - Data manipulation, autodetection of information, neural network self-learning, training without a teacher, encoders, use of trained neural networks and encoders, GPT, RuGPT, Berta, ruBerta. TF Hub, huggingface. - AI Institute and Association Development. <br> <strong> Understanding technology or practiced, among others: </strong> • TensorFlow and Keras • Preparation of datasets, data normalization • Custom Embedding, including invented copyright technologies • Freeze weights and layers. • Saving - restoring the model and weights. • Transfer Learning. • Pretrained networks. • Custom layers, models, custom activation functions and metrics. • Avtotyunig, AutoML. • Regularization and Dropout - fighting overfitting. • Batch Normalization. • Masking. • Manipulations with complex LRs. • use in TF of 16-bit and 32-bit types (float16, float32) for acceleration or precision. • TensorBoard. <br> and more. • Classification, regression, generation, etc. <br> Understanding Technology • RNN, GRU, LSTM, • Convolutional NN, • Encoding-Decoding, Variational Autoencoders (VAE), • Generation - GAN, • Seq2seq and Attention, • Transformer BERT, GPT-3, • ELMo and more. <br> Clustering: • K-means • Affinity Propagation • Mean Shift • Spectral clustering • Hierarchical clustering • DBSCAN • OPTICS and more. </ul> <ul> Hundreds of discoveries, ideas, proprietary custom technologies and directions of research, researches: • Development of the system of the Institute of AI itself and searches, works, the latest production (intelligence). • Development of research systems (special Institute) • Developer of speed education. • An exponent for AI and x intelligence ... where without it. • Universal, Strong AI (both from real to real, and Futurology (future, ideas without criticism). • Learning without a teacher, Unsupervised learning. • Training with partial involvement of a teacher, Semi-supervised learning. • Custom Embedding, including invented proprietary technologies. • Technologies and inventions of data supply for neurons, VKon the fly (video, text, sound, speech, documentation). • Training on unlabeled data. • AI self-creation. • Maximum automatism and self-optimization. • "Snowball" - endless self-improvement, training. • An expanded picture of the object, a complex neural network. • Complex sequence, pipeline, chain for a neural network. • Modernized and invented regression and prediction systems many times (for example, a robot on the Moscow and NY stock exchanges, crypt - everything is in the past). <br> </ul> <ul> Favorite direction - development, new research and development, large-scale progress, civilization, the latest industries, the development of medicine and science on a large scale. <br> Systems for AI, thinking, solutions, inventions, production of the latest technologies and products. <br> Health and life extension, support of the mind, memory and all processes in the body in excellent condition, breakthrough technologies and products. <br> I develop for experiments and experiments, for collecting and testing AI blocks. <br> I develop systems to support AI teams, help in creating and conditions for them, advisor on working with AI startups, constructor. </ul> <ul> Computer architecture, computer hardware - advanced practice. My first computers long before Pentium1 and obsessed with motherboards, caches, chips, screws, memory and peripherals. <br> </ul> <ul> • Pandas, Numpy. • digital video, sound, photo - experience in automatic, batch editing, slicing, converting, etc. - data for the neuron. • Database. • Experience in predicting rates of currencies, stocks, cryptocurrencies. Correlations. • Successful market maker - programmer (in% - super-profit, robots and neural networks for risk-free trading). <br> Creation / Programming: • Google Colaboratory, Jupyter Notebook and PC code editors • sites, shops, robots, services, Internet project management. • php, lua, qlua, mySQL, PhpMyAdmin, json, xml, api as a client, Vk api - (more than 3 years of experience), did everything himself, and also attracted masters, organizer, idea, project leader, own projects. • php OOP, python OOP, WebSocket, JS - just a little bit about that. • I will master any language and framework - quickly. • Created robots-market makers on the Moscow Exchange (Quik) and on crypto-exchanges - more than 3 years of experience. This is the last work, the result is profit, smart, they count 1,000,000 combinations in minutes on i5, 3 levels of cache, protection against hairpins, from external API glitches. • charts, informers and calculators - cache, protection from hairpins, from overload, from robots, keeping access to hidden search engine robots, handling loss of connection with an external API, using a php library, using js and whatever. • Participated in project management (creation of robots, websites for the company, design). • Bootstrap3, Bitrix, 1C - it was a long time, experience less than a year. • Experience of working with a client (trust, personal disposition, I can earn good money, appreciates, repeatedly). • statistics, Google Analytics, Yandex.Metrica, LiveInternet, Wordstat. • Search engine optimization sites, the site has been profitable for 10 years without support. Average was 2000 unique per day, max 20,000. • Yandex.Direct, Google AdWords and Target. • Yandex Advertising Network, Adsense. • CRM (wrote CRM to myself in php + Vkontakt API). • Advertising banners. <br> Internet - shop • experience in an elite boutique: new website, photo, video, net, all IT, image maker, designer, tailor, fashion. • OpenCart. Auto-filling, loading-unloading-parsing of goods from other sites, synchronization with 1C • any CMS. </ul> <ul> Radio engineering, radio electronics, equipment - soldered, assembled according to diagrams, repaired. <br> Car enthusiast, obsessed with the iron of the car, assembled - disassembled the car. <br> Photo-video filming, graphics, video: <br> Portrait, backlit, flash, depth of field, no shadows, interchangeable lenses, movable platform and tripod. • CorelDRAW, Photoshop. • 3D editors. • Video editors. • Design and Web Design. </ul> <ul> <strong> Sound, voice acting and music. </strong> • Cubase. Fruits. Cakewalk Sonar, Waves, Sound Forge, MIDI and VST plugins - Great experience. • effect processors, sound processing, samples and synthesizers, mixing. <br> Systems, OS, office • installed / configured Windows, office networks and the Internet. Great experience. • Computer hardware, repair, office network setup. • Linux, Ubuntu, Debian, Apache / Nginx / php / http-server. • Works on FTP. • English - 20 years of reading scientific, complex documentation, technical - very different, with a very wide coverage. </ul> tel: +79633123320, WhatsApp + 79269319322 Telegram @shatatoy or + 79269319322 or https://t.me/shatatoy vk - shatatoy or https://vk.com/shatatoy https://www.facebook.com/alex.tatoi Zhuykov Alexandr - Tatoy (Shatatoy - @ Syrius4 - Τατόι) </ul></html>