Browsing by Author "Ajanki, Antti"

Filter results by typing the first few letters
Now showing 1 - 9 of 9
  • Evaluating explainable AI models for convolutional neural networks with proxy tasks
    (2019-12-16) Kanerva, Olli
     Perustieteiden korkeakoulu | Master's thesis
    The field of artificial intelligence and machine learning have been developing rapidly throughout the years. The machine learning models are growing in a way that it is harder than ever to understand why certain decisions are made. Explainable AI is a group of different methods that tries to fix this problem. This thesis starts by looking into the current state of explainable AI and examine which kind of explainable models there are and how are they tested. We will also research three different explainable AI models in detail to understand and test them with different tasks. As the research currently lacks standardized automated testing and evaluation of explainable AI models, we introduce two different proxy tasks: pattern task and Gaussian blot task, which both generate a dataset for convolutional neural networks to classify. These neural network results are fed into the explainable AI models. Additionally, we introduce an evaluation scheme that will evaluate the explainable AI by scoring each explanation. In the results of this thesis, we use the introduced proxy tasks and evaluation scheme to evaluate the three explainable AI models, that are suitable for convolutional neural networks, and discuss in which kind of situations each model seems to work best. We conclude that the proxy tasks and the evaluation scheme work as they were intended. Both tasks provide hard enough tasks for the neural network and explainable models while the evaluation scheme does provide a sensible score for each sample.
  • Expertise Seeking and Automatic Expertise Retrieval in a Software Consultancy
    (2021-05-17) Nevalainen, Niklas
     Perustieteiden korkeakoulu | Master's thesis
    Consulting internal experts is a frequent activity in many organizations but locating them becomes harder as organizations grow. This is particularly true for consultancy business that involves a rapidly changing knowledge environment and where employees frequently navigate between projects acquiring knowledge from multiple domains. Digital engineering and innovation company Futurice launched an expert search application to help their employees locate internal expertise and increase their ability to apply talent to projects more broadly. Despite initially promising results, the application did not establish its position among the general expertise seeking practices of Futurice employees. This master's thesis studies ways to make the application more valuable for users and better integrate it with current working practices. To achieve this, we conduct semi-structured interviews to understand factors controlling expert selection and the manual practices we aimed to augment. Additionally, we perform a literature review to understand how the application compares to previous automatic expertise retrieval methods. The interview results showed that task context heavily influenced the employees' source selection process. Multiple contextual factors controlled the expert selection process, and employees most often located expertise either through social networks or broadcasting the question. The literary review illustrated that approaches of varying complexity have been proposed for the expertise retrieval task, but no single clear winner exists. The majority of previous attempts have adopted an abstracted view of the retrieval task and not considered the real dynamics of expertise seeking. Finally, we propose tangible ways to incorporate the domain understanding acquired from the interviews into the expert search application.
  • Fiksaatioiden tunnistusalgoritmit katseenseurannassa
    (2010) Kanto, Jaakko
     Informaatio- ja luonnontieteiden tiedekunta | Bachelor's thesis
  • Inference of relevance for proactive information retrieval
    (2013) Ajanki, Antti
     School of Science | Doctoral dissertation (article-based)
    Search engines have become very important as the amount of digital data has grown dramatically. The most common search interfaces require one to describe an information need using a small number of search terms, but that is not feasible in all situations. Expressing a complex query as precise search terms is often difficult. In the future, better search engines can anticipate user's goals and provide relevant results automatically, without the need to specify search queries in detail. Machine learning methods are important building blocks in constructing more intelligent search engines. Methods can be trained to predict which documents are relevant for the searcher. The prediction is based on recorded feedback or observations of how the user interacts with the search engine and result documents. If the relevance can be estimated reliably, interesting documents can be retrieved and displayed automatically. This thesis studies machine learning methods for information retrieval and new kinds of applications enabled by them. The thesis introduces relevance inference methods for estimating query terms from eye movement patterns during reading and for combining relevance feedback given on multiple connected data domains, such as images and their captions. Furthermore, a novel retrieval application for accessing contextually relevant information in the real world surroundings through augmented reality data glasses is presented, and a search interface that provides browsing cues by making potentially relevant items more salient is introduced. Prototype versions of the proposed methods and applications have been implemented and tested in simulation and user studies. The tests show that these methods often help the searcher to locate the right items faster than traditional keyword search interfaces would. The experimental results demonstrate that, by developing custom machine learning methods, it is possible to infer intent from feedback and retrieve relevant material proactively. In the future, applications based on similar methods have the potential to make finding relevant information easier in many application areas.
  • Knowledge Management: Document Similarity Based Recommendation
    (2021-01-25) Khan, Omer
     Perustieteiden korkeakoulu | Master's thesis
    Futurice works on developing and designing digital services and products. Its an innovative organization which invests its expertise and experience in promoting and creating a knowledge management system powered by data and artificial intelligence to make the internal process of the organization robust and dynamic. Within this vision, a project called Exponential AI focuses on connecting internal knowledge and developing a platform for querying related and similar materials. This thesis is a part of the same project. The study was based on recommendation systems which can create a new exploratory viewpoint to investigate related, unexplored and untapped materials (documents). Interviews were conducted with various sales representatives regarding the challenges and barriers faced during the current process. It was found that the most critical need was to develop a searching mechanism for proposal creation, since it was hard to access and recycle similar material within the organization. Therefore, a literature study was conducted on different approaches that have been used in the past to structure knowledge management and apply document similarity methods to textual data. Several methods were tested including K-nearest neighbors, association ruling, clustering and word embeddings (Word2Vec, Doc2Vec) to measure text similarity. In this thesis, a recommendation system was developed. It was based on document-based similarity between proposals. The features for the model were extracted from documents and they support matching of these documents on the basis content, context, and industry domain. From the comparative analysis of different methods, we found that word embedding models performed better, in general, and in particular Word2Vec and Doc2Vec models outperformed the other techniques. Finally, the results from both methods were combined and ranked according to the date and score of similarity of the document to provide qualitative recommendations. A set of sales representatives were asked to test and provide feedback on the usefulness and results of the recommendations. According to the responses, results are promising, and the recommendation system has potential to resolve challenges like material recycling, exploration, and time management in the current sales process.
  • Kontekstin hyödyntäminen tiedonhaussa
    (2012-04-29) Koponen, Laura
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Geenisäätelyn mallinnus tilanneriippuvilla Bayes-verkoilla
    (2006) Ajanki, Antti
     Helsinki University of Technology | Master's thesis
    Solut tulevat toimeen useissa erilaisissa olosuhteissa, koska niiden toimintaa ohjaavien geenien ilmentymisaktiivisuus voi muuttua ympäristöstä tulevien signaalien tai toisten geenien tuottamien proteiinien vaikutuksen perusteella. Geenien väliset säätelysuhteet määräävät solun käyttäytymisen. Säätelyverkoston koko ja monimutkaisuus tekevät sen selvittämisestä haastavan ongelman. Todennäköisyyslaskentaan perustuvat Bayes-verkot ovat eräs yleisesti käytetty esitystapa geenien säätelysuhteiden matemaattiseen mallintamiseen. Niille on olemassa opetusalgoritmeja, jotka etsivät mitattuihin ilmentymisprofiileihin parhaiten sopivan verkon. Opitun verkon rakenne voidaan tulkita geenien säätelyverkoksi. Yleensä Bayes-verkkojen opetusmenetelmät olettavat, että kaikki havainnot on tehty samoissa olosuhteissa. Jos halutaan tutkia miten säätelyvuorovaikutukset muuttuvat olosuhteiden välillä, eräs tapa olisi opettaa erilliset verkot kuvaamaan eri olosuhteiden havaintoja ja verrata opittuja verkkoja keskenään. Silloin kunkin verkon opetukseen olisi kuitenkin käytettävissä vain osa opetusnäytteistä, mikä saattaisi johtaa ylisovittumiseen. Tämä työ esittelee verkkorakenteen ja opetusalgoritmin, joita voidaan käyttää säätelyerojen etsimiseen. Näytteen mittausolosuhde huomioidaan itsenäisenä luokkamuuttujana. Uutta työssä on tapa, jolla luokkaa käytetään määräämään solmujen jakaumien riippuvuudet. Se helpottaa opitun verkon tulkintaa. Luokkamuuttujan ansiosta kaikki riippuvuudet voidaan esittää yhdessä verkossa, jonka opetukseen voidaan käyttää kaikkia havaintoja. Esiteltävä opetusalgoritmi löytää automaattisesti ne verkon osat, joissa on eroja luokkien välillä. Työssä osoitetaan keinotekoisia opetusnäytteitä käyttäen, että ehdotettu opetusalgoritmi tuottaa paremmin oikeaa vastaavia verkkoja kuin oman verkon opettaminen erikseen joka olosuhteelle. Menetelmää sovelletaan stressaavien olosuhteiden aiheuttamien säätelyerojen etsimiseen hiivassa.
  • Ubiquitous contextual information access with proactive retrieval and augmentation
    (2009) Ajanki, Antti; Billinghurst, Mark; Kandemir, Melih; Kaski, Samuel; Koskela, Markus; Kurimo, Mikko; Laaksonen, Jorma; Puolamäki, Kai; Tossavainen, Timo
     Faculty of Information and Natural Sciences | D4 Julkaistu kehittämis- tai tutkimusraportti taikka -selvitys
    In this paper we report on a prototype platform for accessing abstract information in real-world pervasive computing environments through Augmented Reality displays. Objects, people, and the environment serve as contextual channels to more information. Adaptive models will infer from eye movement patterns and other implicit feedback signals the interests of users with respect to the environment, and results of proactive context-sensitive information retrieval are augmented onto the view of data glasses or other see-through displays. The augmented information becomes part of the context, and if it is relevant the system detects it and zooms progressively further. In this paper we describe the first use of the platform to develop a pilot application, a virtual laboratory guide, and early evaluation results.
  • Uncertainty in Recurrent Neural Network with Dropout
    (2020-08-18) Nguyen, Kha
     Perustieteiden korkeakoulu | Master's thesis
    Recurrent Neural Network is a powerful tool for processing temporal data. However, assessing prediction uncertainty from recurrent models has proven challenging. This thesis attempts to evaluate the validity of uncertainty from recurrent models using dropout. Traditional neural network focuses on optimising data likelihood; in order to obtain model and predictive uncertainty, we need to, instead, optimise model posterior. Model posterior is usually intractable, thus we employ various dropout based approach, in the form of variational Bayesian Monte Carlo, to estimate the learning objective. This technique is applied to existing recurrent neural network benchmarks MIMIC-III. The thesis shows that Monte Carlo dropout applied to recurrent neural network can give comparable performance to the current state of the art methods, and meaningful uncertainty of predictions.