Using web technologies for designing and implementing multi-layer neural networks

Ovaj rad obrađuje relativno novo područje izrade i prikaza višeslojnih neuronskih mreža korištenjem web tehnologija i radi usporedbu između uvježbavanja neuronskih mreža pomoć web tehnologija i trenutno najpopularnijih rješenja. U prvom dijelu rada objašnjava se što su neuronske mreže, kako je došlo do njihovog razvoja, koji su neki od popularnih tipova učenja i po čemu se razlikuju i kako radi jedna popularna arhitektura neuronske mreže ‘plitkog učenja’ i jedna ‘dubokog učenja’, obje koje se mogu stvoriti, uvježbavati i koristiti u projektu ovog rada. Posebno se obrađuje algoritam učenja koji je, uz male razlike, korišten za obje arhitekture, u svrhu razumijevanja složenosti procesa učenja preko kojega će se kasnije uspoređivati performanse različitih implementacija sustava za rad s neuronskim mrežama. U sljedećem dijelu prikazuje se dio projekta ovog rada, tj. aplikacije koja je izrađena isključivo u web tehnologijama kao dokaz da su se web tehnologije dovoljno razvile za omogućavanje razvoja takvih aplikacija. Kako je ‘duboko učenje’ postalo jako popularno u području umjetne inteligencije pa tako i u svijetu, kao primjer da web tehnologije mogu podržati i tu vrstu odabrana je LSTM (Long Short-term Memory) arhitektura kao jedne od popularnih u području obrade prirodnog jezika. Nakon toga performanse se uspoređuju s TensorFlow-om, trenutno najpopularnijim rješenjem za rad s neuronskim mrežama i koji uvježbavanje mreže izvodi u strojnom jeziku. U zadnjem dijelu prikazuje se drugi dio projekta koji se sastoji od web aplikacije koja se u pozadini koristi TensorFlow-om za brže i efikasnije uvježbavanje i korištenje neuronskih mreža. Aplikacija se temelji na modernim praksama razvoja web aplikacija i napravljena je na način koji osigurava jednostavno nadograđivanje.This thesis explores the relatively new field of constructing and displaying artificial neural networks using web technologies and compares the performance of neural networks trained with web technologies and with the currently most popular machine learning framework. In the first part of this thesis neural networks are explained, what they are, how they evolved, what some of the more popular learning types are and what their differences are and how a shallow learning and a deep learning neural network operate, both of which can be created, trained and used in the practical part of the thesis. The learning algorithm is particularly explored, which is the one that the both supported neural network architectures use, with slight difference, all to demonstrate the complexities included in the training process which is later used to compare the performances of two different implementations of neural network applications. The next part of the thesis shows the web application that was created using purely web technologies as a proof that the technologies have evolved sufficiently to allow creating such applications. Because deep learning became very popular in the field of artificial intelligence and is widely used in the world, as an example that the web application can support that kind of neural networks, the LSTM (Long Short-term Memory) architecture is selected. The LSTM network is popular in the field of natural language processing. After that the performance of the web application is compared with TensorFlow, the currently most popular neural network framework which performs the training by compiling the code and then running it. In the last part of this thesis explores the second part of the project which uses TensorFlow for faster and more efficient training and usage of neural networks. The application is based on current web development standards and is created to be easily extendable