History of Previous Talks in Autumn 2018

Something Old, Something New, Something Borrowed, Something Blue

Date: December 17, 2018

Abstract: In this talk I will briefly describe (something old) some of our recent work with hierarchical probabilistic models that are not deep neural networks. Nevertheless, these are currently among the state of the art in classification and in topic modelling: k-dependence Bayesian networks and hierarchical topic models, respectively, and both are deep models in a different sense. On deep neural networks (something new), I will describe as a point of comparison some of the state of the art applications I am familiar with: multi-task learning and learning to learn. These build on the RNNs widely used in semi-structured learning. The old and the new are remarkably different. So what are the new capabilities deep neural networks have yielded? Do we even need the old technology? What can we do next? To complete the story, I’ll introduce some efforts to combine the two approaches (something borrowed), borrowing from earlier work in statistics. Time permitting, I’ll also describe something blue.

Speaker: Wray Buntine

Affiliation: Professor of Information Technology, Monash University

Place of Seminar: Seminar Room T6, Konemiehentie 2, Aalto University

Slides, Video

Towards Emotion AI

Date: December 10, 2018

Abstract: Developing emotionally intelligent machines with natural perceptual interfaces is one of the ultimate goals of modern computer science and artificial intelligence research. The aim is to bring emotional intelligence to various human-computer-interaction. This talk will focus on analyzing micro-expressions which are subtle, brief and involuntary facial movements that reveal the emotions in high-stake situation that people is trying to conceal his/her true feelings. The data set collection, supervised recognition and unsupervised spotting of spontaneous micro-expression are presented. In addition, separable modal learning will be discussed for multi-modal learning & separate-modal prediction scenarios.

Speaker: Guoying Zhao

Affiliation: Professor of Computer Science and Engineering, University of Oulu

Place of Seminar: Lecture Hall Exactum D122, University of Helsinki

Video

Exploring Large And Hierarchical Online Discussion Venues With Probabilistic Models

Date: November 26, 2018

Abstract: In many domains, document sets are hierarchically organized such as message forums having multiple levels of sections. Analysis of latent topics within such content is crucial for tasks like trend and user interest analysis. Nonparametric topic models are a powerful approach, but traditional Hierarchical Dirichlet Processes (HDPs) are unable to fully take into account topic sharing across deep hierarchical structure. Moreover, in addition to the underlying trends of content and the structure of the venue, another key aspect of online discussion venues is the multitude of participants; authors may participate differently at multiple levels of sections, with different interests and contributions across the hierarchy. We introduce the Tree-structured Hierarchical Dirichlet Process (THDP), allowing Dirichlet process based topic modeling over a given tree structure of arbitrary size and height, where documents can arise at all tree nodes. We further introduce the Author Tree-structured Hierarchical Dirichlet Process (ATHDP), allowing Dirichlet process based topic modeling of both text content and authors over a given tree structure of arbitrary size and height. Experiments on a hierarchical forums demonstrate better generalization performance of THDP than traditional HDPs in terms of ability to model new data and classify documents to sections, and better performance of ATHDP compared to traditional HDP based alternatives in terms of perplexity and authorship attribution accuracy. Lastly, we introduce a novel interactive system for visualizing and exploring a large hierarchical text corpus of online forum postings, based on large-scale scatter plots created by flexible nonlinear dimensionality reduction of posting contents, coupled with a coloring optimized to represent the forum hierarchy. We exploit the hierarchy to provide data-driven summaries of plot areas at multiple levels of detail, allowing the user to quickly compare both the content-based similarity of groups of posts and how near they arise in the forum hierarchy.

Speaker: Jaakko Peltonen

Affiliation: Professor of Statistics, University of Tampere

Place of Seminar: Seminar Room T5, Konemiehentie 2, Aalto University

Video

MegaSense: Scalable Air Pollution Sensing in Megacities

Date: November 19, 2018

Abstract: This talk gives an overview of the MegaSense research program that is a collaboration between Computer Science, Atmospheric Sciences and Geosciences at the University of Helsinki. The research program designs and deploys an air pollution monitoring system for realizing low-cost, near real-time and high resolution spatio-temporal air pollution maps of urban areas. MegaSense involves a novel hierarchy of multi-vendor distributed air quality sensors, in which more accurate sensors calibrate lower cost sensors. Current low-cost air quality sensors suffer from measurement drift and they have low accuracy. We address this significant open problem for dense urban areas by developing a calibration scheme based on machine learning that detects and automatically corrects drift. MegaSense integrates with the 5G cellular network and leverages mobile edge computing for sensor management and distributed pollution map creation.

Speaker: Sasu Tarkoma

Affiliation: Professor of Computer Science, University of Helsinki

Place of Seminar: Lecture Hall Exactum D122, University of Helsinki

High-dimensional Covariance Matrix Estimation With Applications in Finance and Genomic Studies

Date: November 12, 2018

Abstract: We consider the problem of estimating a high-dimensional (HD) covariance matrix when the sample size is smaller, or not much larger, than the dimensionality of the data, which could potentially be very large. We develop a regularized sample covariance matrix (RSCM) estimator that is optimal (in minimum mean squared error sense) when the data is sampled from an unspecified elliptically symmetric distribution. The proposed covariance estimator is then used in portfolio optimization problems in finance and microarray data analysis (MDA). In portfolio optimization problem we use our estimator for optimally allocating the total wealth to a large number of assets, where optimality means that the risk (i.e., variance of portfolio returns) is minimized. Microarray technology is a powerful approach for genomics research that allows monitoring the expression levels of tens of thousands of genes simultaneously. We develop a compressive regularized discriminant analysis (CRDA) method based on our covariance estimator and illustrate its effectiveness in MDA. Our analysis results on real stock market data and microarray data illustrate that the proposed approach is able to outperform the current benchmark methods.

Speaker: Esa Ollila

Affiliation: Professor of Electrical Engineering, Aalto University

Place of Seminar: Seminar Room T5, Konemiehentie 2, Aalto University

Video

Satellite remote sensing and the potential of machine learning

Date: November 5, 2018

Abstract: Satellites have become an essential technology for monitoring the Earth’s environment. The modern society has become more and more dependent on real-time observations to support security and critical functions of the society. The changing climate further emphasises the needs for reliable global observations, e.g., for understanding the carbon and water cycles. The EU’s Copernicus Remote Sensing programme and its Sentinel satellites will ensure operational observations till 2030s from six dedicated satellite missions, many of them consisting of several satellites. The free and open data policy of Copernicus programme gives potential for developing services and applications based on the data for scientific needs as well as for commercial, public and decision making purposes. The huge data volumes of the satellites call for automatic methods for data analysis and the potential of machine learning techniques needs to be further explored. In this presentation, an overview of the large the satellite programme is given with some examples and applications on using the satellite data. The aim is to motivate research on developing machine learning methods for applications using satellite observations of the Earth’s environment.

Speaker: Johanna Tamminen

Affiliation: Professor of Atmospheric Remote Sensing, Finnish Meteorological Institute

Place of Seminar: Lecture Hall Exactum D122, University of Helsinki

Slides, Video

Noise2Noise: Learning Image Restoration without Clean Data

Date: October 29, 2018

Abstract: We apply basic statistical reasoning to signal reconstruction by machine learning–learning to map corrupted observations to clean signals–with a simple and powerful conclusion: under certain common circumstances, it is possible to learn to restore signals without ever observing clean ones, at performance close or equal to training using clean exemplars. We show applications in photographic noise removal, denoising of synthetic Monte Carlo images, and reconstruction of MRI scans from undersampled inputs, all based on only observing corrupted data. These results have significant implications for ease of training high-performance image restoration models and certain inverse problem solvers.

Joint work with Jacob Munkberg, Jon Hasselgren, Timo Aila, Tero Karras, Samuli Laine, and Miika Aittala.

Speaker: Jaakko Lehtinen

Affiliation: Professor of Computer Science, Aalto University

Place of Seminar: Seminar Room T5, Konemiehentie 2, Aalto University

Slides, Video

Human-guided data exploration

Date: October 22, 2018

Abstract: The outcome of the explorative data analysis (EDA) phase is vital for successful data analysis. EDA is more effective when the user interacts with the system used to carry out the exploration. A good EDA system has three requirements: (i) it must be able to model the information already known by the user and the information learned by the user, (ii) the user must be able to formulate the objectives, and (iii) the system must be able to show the user views that are maximally informative about desired features data that are not already know for the user. Furthermore, the system should be fast if used in interactive system. We present the Human Guided Data Exploration framework which satisfies these requirements and generalizes previous research,. This framework allows the user to incorporate existing knowledge into the exploration process, focus on exploring a subset of the data, and compare different complex hypotheses concerning relations in the data. The framework utilises a computationally efficient constrained randomization scheme. To showcase the framework, we developed a free open-source tool, using which the empirical evaluation on real-world data sets was carried out. Our evaluation shows that the ability to focus on particular subsets and being able to compare hypotheses are important additions to the interactive iterative data mining process.

For references see:
Puolamäki Kang, Lijffijt, De Bie, 2016, Interactive Visual Data Exploration with Subjective Feedback. In Proc ECML PKDD, https://doi.org/10.1007/978-3-319-46227-1_14
Puolamäki, Oikarinen, Kang, Lijffijt, De Bie, 2018. Interactive Visual Data Exploration with Subjective Feedback: An Information-Theoretic Approach. In Proc ICDE 2018, https://arxiv.org/abs/1710.08167
Puolamäki, Oikarinen, Atli, Henelius, 2018. Human-guided data exploration using randomization. https://arxiv.org/abs/1805.07725

Speaker: Kai Puolamäki

Affiliation: Professor of Computer Science, University of Helsinki

Place of Seminar: Lecture Hall Exactum D122, University of Helsinki

Slides, Video

Natural Language Inference with Multilingual Supervision

Date: October 15, 2018

Abstract: Natural Language Inference (NLI) is the task of identifying inferential relationships between a premise p and a given hypothesis h. Both, p and h are expressed in natural language, typically as a pair of sentences with a specific relationship between them coming from a limited set of inferential relations (e.g. entailment, contradiction, neutral). NLI requires semantic analyses and, therefore, can be seen as a test of text understanding capabilities of a system. Modern NLI models are based on deep neural nets and either cross-sentential encoding or independent sentence embeddings. In this talk, I will present our work on sentence representation learning and its appliction to common NLI benchmarks. I will start with the introduction of a state-of-the-art supervised NLI model with hierarchical bi-LSTM architectures and, after that, discuss our research in multilingual supervision for representation learning. The latter is motivated by the use of translations as semantic mirrors and the idea of applying highly multilingual data sets in neural machine translation to learn language-independent meaning representations.

Speaker: Jörg Tiedemann

Affiliation: Professor of Language Technology, University of Helsinki

Place of Seminar: Seminar Room T6, Konemiehentie 2, Aalto University

Slides, Video

Saving Lives with Big Data: Data Analytics in Intensive Care Units

Date: October 8, 2018

Abstract: Critical care units, with their sophisticated patient technologies, generate massive amounts of heterogeneous data that include measurements, medical imagery, diagnosis transcriptions, etc. These data have the potential to improve our understanding of diseases and clinical care in general. The application of mathematical models, algorithms, and deep learning strategies to these records enables us to design data-driven clinical decision support tools for physicians and to devise population-level health policies for governmental organizations. In addition to helping in the study of rare disease conditions, the sheer magnitude of the available datasets will facilitate the application of advanced algorithmic techniques and prospective bedside decision support
tools.

Speaker: Pan Hui

Affiliation: Professor of Computer Science, University of Helsinki

Place of Seminar: Lecture Hall Exactum D122, University of Helsinki

Sample-Efficient Deep Learning

Date: October 1, 2018

Abstract: Critical care units, with their sophisticated patient technologies, generate massive amounts of heterogeneous data that include measurements, medical imagery, diagnosis transcriptions, etc. These data have the potential to improve our understanding of diseases and clinical care in general. The application of mathematical models, algorithms, and deep learning strategies to these records enables us to design data-driven clinical decision support tools for physicians and to devise population-level health policies for governmental organizations. In addition to helping in the study of rare disease conditions, the sheer magnitude of the available datasets will facilitate the application of advanced algorithmic techniques and prospective bedside
decision support tools.

Speaker: Alexander Ilin

Affiliation: Professor of Computer Science, Aalto University

Place of Seminar: Seminar Room T6, Konemiehentie 2, Aalto University

Machine Learning for Data Management – and vice versa

Date: September 24, 2018

Abstract: Research efforts in the areas of Machine Learning (ML) and Data Management (DM) have, to a large degree, run in parallel for many years. DM research focused on the efficient querying of databases, and led to standardized and widely used database management systems. ML research focused on building predictive models from data, and led to high predictive performance for some particularly difficult tasks (e.g., image recognition). As ML software is used in more and more applications, it’s worth discussing whether we can have for ML systems the kind of standardization we have for DM systems – as well as whether we can use ML to improve DM systems. In this talk, I’ll present one recent paper for each direction from SIGMOD2018 [1] and VLDB2018 [2].

[1] Kraska T, Beutel A, Chi EH, Dean J, Polyzotis N. The case for learned index structures. In Proceedings of the 2018 International Conference on Management of Data 2018 May 27 (pp. 489-504). ACM.

[2] Hasani S, Thirumuruganathan S, Asudeh A, Koudas N, Das G. Efficient construction of approximate ad-hoc ML models through materialization and reuse. Proceedings of the VLDB Endowment. 2018 Jul 1;11(11):1468-81.

Speaker: Michael Mathioudakis

Affiliation: Professor of Computer Science, University of Helsinki

Place of Seminar: Lecture Hall Exactum D122, University of Helsinki

Slides, Video

Progressive Growing of GANs for Improved Quality, Stability, and Variation

Date: September 17, 2018

Abstract: We describe a new training methodology for generative adversarial networks. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. This both speeds the training up and greatly stabilizes it, allowing us to produce images of unprecedented quality, e.g., CelebA images at 1024^2. We also propose a simple way to increase the variation in generated images, and achieve a record inception score of 8.80 in unsupervised CIFAR10. Additionally, we describe several implementation details that are important for discouraging unhealthy competition between the generator and discriminator. Finally, we suggest a new metric for evaluating GAN results, both in terms of image quality and variation. As an additional contribution, we construct a higher-quality version of the CelebA dataset.

Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen, ICLR 2018

Speaker: Tero Karras

Affiliation: Research Scientist, NVIDIA

Place of Seminar: Seminar Room T6, Konemiehentie 2, Aalto University

Slides, Video

The Power of Gaussian Processes: Magnetic Localisation and Mapping

Date: September 3, 2018

Abstract: Gaussian processes (GPs) are convenient tools for model-building and inference. This talk goes through how to encode knowledge from high-school physics into a GP model for the ambient magnetic field (observed by a smartphone compass). Small disturbances in the magnetic field are then used in simultaneous localisation and mapping (SLAM) to simultaneously build a map of the magnetic field and localise the device on it by Rao-Blackwellised particle filtering (Sequential Monte Carlo). The paper presenting this setup recently won the Best Paper Award at the International Conference on Information Fusion 2018.

Speaker: Arno Solin

Affiliation: Professor of Computer Science, Aalto University

Place of Seminar: Seminar Room T6, Konemiehentie 2, Aalto University

Slides, Video