Brain & Mind Computational Breakfast
NB. The December edition will be held on Dec 17.
Next seminar: Tuesday, 12 November
Topics and speakers:
New Approaches for Analyzing Multivariate and Infinite Dimensional Data
Pauliina Ilmonen, Assistant Professor, Statistics, Department of Mathematics and Systems Analysis, Aalto University
In modern data analysis, we often have to deal with very complicated observations. Dimensions can be very large and contain complicated dependency structures. Observation may be e.g. images, bits, colors, or functions. Classical statistical methods that rely on normality and i.i.d. observations can give faulty results or are at least inefficient for analyzing modern data. We approach this problem in three ways:
- We develop new methods for modelling multivariate extreme values. We provide new estimators for multivariate tail indices and multivariate extreme quantiles.
- We consider general location-scatter-models and invariant coordinate selection – methods that do not rely on normality assumptions and can be applied for analyzing multivariate and even tensor-valued observations.
- We develop new nonparametric methods for analyzing infinite dimensional – functional – observations. We consider outlier detection and different types of classification and clustering problems.
Attentive Futures: Cognition under Control
Juha Salmitaival, PhD, Academy Research Fellow, Department of Neuroscience and Biomedical Engineering, Aalto University School of Science
Senior Research Fellow, Turku Institute for Advanced Studies, University of Turku
Clinical psychologist (neuropsychology), ProNeuron, Espoo
It is estimated that most of the information ever produced by mankind is just a few years old. This change in the amount of available, and also potentially relevant information has been incredibly fast, while our brain and its very limited ability to handle new information has gradually developed during evolution under very different circumstances. The limitations in attention and executive skills are exacerbated in various clinical conditions (ADHD, autism, burn-out, depression, dementias) with rapidly increasing prevalence rates. Current health care relies on effortful, expensive, subjective, and inaccurate methods that cannot be provided for everybody in the need. I will introduce some new clinically and neuroscientifically motivated health tech solutions (online testing, mobile apps, VR games) that aim to advance the assessment and treatment of attention deficits in multiple ways. These systems provide massive amounts of reliable cognitive testing data, but unlike in other fields in computational psychiatry (symptoms data, digital phenotyping), neuroimaging or neurogenetics, there are only few researchers developing advanced computational methods for cognitive testing purposes. Cognitive testing data might be even particularly for data mining purposes because it is reliable and objectively measured (unlike symptoms data), and directly measures actual behavior (unlike neuroimaging and genetics data). Could we build together the road for individualized treatments supporting attentive futures?
Dates and Speakers after November
17 December (note the date exception!)
Matti Hämäläinen, Timo Roine
NB! The talks begin at 9:30. Breakfast is served on a first-come-first-serve basis at 9:15.
Using domain knowledge in machine learning models: a deep dive into linear regression
Marijn van Vliet, Aalto University
Linear machine learning models are a powerful tool that can learn a data
transformation by being exposed to examples of input with the desired output.
But what if we don't have enough training data? I'm going to talk about how to
help our model by giving it access to domain information. In order to do this,
we must take a deep dive into how linear regression works.
Computational limits of clinical neuroscience
Prof. Tuukka Raij, Psychiatry, AoF clinical research fellow, HUS and Aalto NBE
I will present two computational challenges faced in the attempts to develop clinical tools for (personalized) neuropsychiatry. A classical challenge is poor replicability of neuroimaging studies that is partially related to the problem of multiple comparisons. Increasing availability of prior data in databases may allow using Bayesian prior images to weight correction for multiple comparisons, but to my knowledge, such methods wait to be established. A more recent challenge relates to predictive models, relying largely on machine learning, and hoped to help to predict outcome and to select optimal treatment at the individual level. It has been suggested that error of cross-validated prediction accuracy follows a binomial law, resulting in about +/- 20 % error bars in a typical-size brain-imaging study on 30 subjects (i.e. prediction accuracies below 70 % may equal tossing a coin). In addition to sample size, several factors contribute to the error that is difficult to estimate in a single study. Discussion between neuroscientists and computational scientist is needed to develop methods to estimate and minimize error in correction for multiple comparisons and in predictive models.
Decoding Neural Drive for New Generation of Neurorehabilitation Systems
Prof. Ivan Vujaklija, Dept. of Electrical Engineering & Automation
Abstract: The generation of movement is related to a combination of discrete events (action potentials) generated in the brain, spinal cord, nerves, and muscles. These events, discharged in the various parts of the neuromuscular system, constitute the neural code for movements. Recording and interpretation of this code provides further means for analyzing the human motor system. Currently we are only partially able to detect and process the available information content and this prevents us from establishing widely applicable and robust man-machine interface systems. Our work on developing new computational methods/models for extracting functionally significant information on human movement will be presented. The focus of these developments is on providing the link between the cellular mechanisms and the behavior of the whole motor system in order to establish clinically viable and effective neurorehabilitation technologies.
Behavioral Patterns through the Lens of Smartphones: the Case of Patients with Mental Disorders
Talayeh Aledavood, Postdoctoral Fellow, University of Helsinki / Aalto University, Dept. of Computer Science
Abstract: Mental disorders are one of the main causes of disability world-wide. In recent years with ubiquity of smartphones and other wearables an unprecedented opportunity is created for researchers to collect data from humans. This data allow us to study behavioral patterns of people in situ and continuously. Using these devices and collecting data from patients with different mental disorders, we can quantity their behavior and also better understand these disorders. This can potentially lead to finding new biomarkers of these disorders. For this purpose we use various machine learning and statistical learning methods. This can eventually help us with providing in-time interventions and treatments for the patients. This talk will give an overview of the type of data that can be gathered analyzed and research questions that can be addressed using these methods.
Speaker: Prof. Aki Vehtari
Affiliation: Dept. of Computer Science, Aalto University
Bayesian Probabilistic Programming
I present how Bayesian probabilistic programming makes modeling and data
analysis easier. Probabilistic programming makes it easy to go beyond classic
statistical tests or pre-defined models, enabling to focus on the actual
question you want to answer and often leading to new questions. I discuss some
modeling ideas and challenges in context of brain research.
Speaker: Prof. Iiro Jääskeläinen
Affiliation: Dept. of Neuroscience and Biomedical Engineering
Family ethnic-cultural background shapes brain activity and associations elicited during listening to an audiobook
We studied whether the shared family ethnic-cultural background increases similarity in how the brain processes an audiobook. 48 healthy subjects, half with majority and half with the minority ethnic-cultural family background, listened to 71-min audiobook during functional magnetic resonance imaging (fMRI) of brain hemodynamic activity at 20-times faster-than-conventional temporal sampling rate. There were protagonists of both ethnic cultural backgrounds in the drama-genre audiobook alternating between portrayals of social relationships and city scenery amidst changing seasons. Subsequently, the audiobook was replayed to subjects in segments, and they were to list words best describing what had been on their minds as they initially heard the segment during fMRI. Significant within-group enhancements in cosine similarity of the word lists within a semantic vector space based on a large internet-text corpus suggested that the shared family ethniccultural background increased similarity in the meanings elicited by the audiobook. Further, there were significant within-group increases in inter-subject correlation (ISC) of brain hemodynamic activity in the left superior temporal and Heschl’s gyri, bilateral middle temporal gyri, lateral occipital cortex, and precuneus. The lateral-temporal ISC effects suggest that the family ethnic-cultural background increased similarity in processing of the audiobook at the level of individual words and sentences, ISC effects in visual cortical areas suggests increased within-group similarity in visual imagery elicited by the audiobook and, finally, ISC effects in precuneus suggest that the family ethnic-cultural background enhanced similarity in how narrative-level information was processed by the brain. In a separate session, subjects' brain activity was recorded also during MEG/EEG during listening to the audiobook, these data are currently being analyzed.
Computational Rationality: Convergence of AI, robotics, neurosciences, and cognitive science
This talk surveys recent progress in computational rationality, cognitive modelling that is based on the idea that cognitive behaviours are generated by behavioral policies that are optimally adapted to the processing limits of a cognitive architecture (Gershman, Horvitz, & Tenenbaum, 2015; Griffiths, Lieder, & Goodman, 2015; Howes et al., 2009; Richard L. Lewis, Howes, & Singh, 2014). In contrast to cognitive architectures such as ACT-R, which encourage hand-coding of behavioral policies, computational rationality assumes that these policies emerge from the limitations of the specified cognitive architecture. As a framework for modelling cognition, computational rationality has been heavily influenced by rational analysis, a method for explaining behavior in terms of utility.
Brain science needs computational science: the past, the present and the future in Otaniemi
In this first talk of the series I will tell about my vision to strengthen multidisciplinary collaborations in Otaniemi between neuroscientists and computational scientists. I will provide an overview of the successful history of brain science in Otaniemi which has involved multidisciplinary collaborations between brain scientists, computer scientists and physicists since the very beginning. But how can we take things forward? I will propose some solutions for increasing fruitful multidisciplinary collaborations with neuroscientists at Aalto and conclude with some open unsolved questions to stimulate future discussions.
Speaker: Enrico Glerean
Affiliation: Dept. of Neuroscience and Biomedical Engineering, Aalto University