Unsupervised Learning for Tracking and Classification of Sequences
2024 (English)Independent thesis Basic level (degree of Bachelor), 10 credits / 15 HE credits
Student thesis
Abstract [en]
Everyday use of artificial intelligence in the form of large language models(LLMs) has become a noticeable part of life for many. This emphasises the importance ofeffective machine learning models working in an unsupervised learning setup. This paperaims to examine such a model, the recently proposed DANSE model. Specifically, the modelwas adapted for the classification of noisy sequences and compared to an RNN (recurrentneural network) with the same purpose. The sequences analysed were two different three-dimensional trajectories, the Lorenz and Chen attractors. To adapt the proposed model, theDANSE model was trained once for each of the attractors. Furthermore, a maximum-likelihood function was applied, and a prediction was made. The RNN model wasimplemented for classification as a baseline. The results were compared by calculatingaccuracy, recall, precision, and F1 score from a confusion matrix. The results showed that theRNN classifier was more effective for the attractors with more noise. However, with lessnoise, the DANSE model showed better and more consistent results.
Abstract [sv]
Vardaglig användning av artificiell intelligens i form av storaspråkmodeller (LLMs) blivit en noterbar del av livet for många. Detta betonar vikten aveffektiva maskininlärningsmodeller som arbetar i en oövervakad inlärningsmiljö. Dennaartikel syftar till att undersöka en sådan modell, den nyligen föreslagna DANSE-modellen.Specifikt anpassades modellen for klassificering av brusiga sekvenser och jämfördes med ettRNN (Recurrent Neural Network) med samma syfte. De analyserade sekvenserna var tvåolika tredimensionella banor, Lorenz och Chen attraktorer. For att anpassa den föreslagnamodellen tränades DANSE-modellen en gång för varje attraktor. Dessutom tillämpades enmax-likelihood-funktion och en förutsägelse gjordes. RNN-modellen implementerades förklassificering som baslinje. Resultaten jämfördes genom att beräkna noggrannhet, känslighet,precision och F1-poäng från en förvirringsmatris. Resultaten visade att RNN-klassificerarenvar effektivare for attraktorerna med mer brus. Dock visade DANSE-modellen bättre och merkonsekventa resultat med mindre brus.
Place, publisher, year, edition, pages
2024. , p. 645-652
Series
TRITA-EECS-EX ; 2024:193
National Category
Electrical Engineering, Electronic Engineering, Information Engineering
Identifiers
URN: urn:nbn:se:kth:diva-359428OAI: oai:DiVA.org:kth-359428DiVA, id: diva2:1933515
Supervisors
Examiners
Projects
Kandidatexamensarbete Elektroteknik EECS 20242025-01-312025-01-31