Gesture recognition on GPU

Problem prepoznavanja gesti jako je interestantan problem zbog prisutnosti vremenske dimenzije. Postoji više načina kako se problem može riješiti, a u ovom radu istražen je način rješavanja korištenjem slojevite potpuno povezane unaprijedne neuronske mreže. Da bi to bilo moguće, potrebno je transformirati problem tako da ne ovisi o vremenskoj dimenziji. Zbog toga je istraženo više načina uzorkovanja gdje se eliminira vremenska dimenzija. Znanje o problemu implicitno je pohranjeno u težinama između umjetnih neurona u umjetnoj neuronskoj mreži. Kako bi neuronska mreža što bolje obavljala svoj posao, potrebno je što bolje podesiti te težine i taj se postupak zove učenje neuronske mreže. Razmotrena su dva algoritma učenja neuronske mreže, algoritam propagiranja greške unatrag i algoritam rprop. Kao i kod čovjeka, puno je više vremena potrebno utrošiti na učenje nego na prepoznavanje. To je povod za istraživanje načina ubrzavanja učenja i jedan od tih načina jest paralelizacija. Svojom arhitekturom, grafički procesor predstavlja idealnoga kandidata za paralelizaciju učenja neuronskih mreža. U sklopu ovoga rada istražena je paralelizacija na razini mreže, tj. jedna radna stavka predstavlja jedan primjerak neuronske mreže za predočeni uzorak. Problem prepoznavanja gesti još je neistraženo područje koje danas postaje sve zanimljivije sve većom popularnošću pametnih telefona i ekrana osjetljivih na dodir. Postoji još puno segmenata koje je potrebno istražiti i provjeriti poput drugih algoritama učenja ili samih arhitektura neuronskih mreža.The problem of gesture recognition is a very interesting problem because of the temporal dimension that exists and has to be handled somehow. There are several ways in which the problem can be solved and in this paper way that uses fully connected feed-forward neural network has been explored. To make this possible, it is necessary to transform the problem so it does not depend on the temporal dimension. Therefore, a few ways of sampling have been explored in order to eliminate the temporal dimension. Knowledge about the problem is implicitly stored in the weights between the neurons in the neural network. In order for neural network to better perform its job, it is necessary to better adjust the weights and that process is called learning. Two learning algorithms have been implemented, backpropagation algorithm and rprop algorithm. As in humans, much more time is spent on learning than on recognition itself. Because of that, methods that speed up learning had to be explored and one of those method is parallelization. With its architecture, GPU is an ideal candidate for parallel neural network learning. As part of this study, parallel processing at the network level have been explored, ie. one work item presents one instance of neural network. The problem of gesture recognition is still uncharted territory which is becoming more and more interesting, especially with the growing popularity of smart phones and touch screens. There are still a lot of things that need to be explored, like other learning algorithms or other architectures of neural networks