Neuroinformatics Aapo Hyvärinen, still Academy Research Fellow for a while Post-docs: Patrik Hoyer and Jarmo Hurri + possibly international post-docs PhD students Ilmari Kurki, Urs Köster, Jukka Perkiö, (Shohei Shimizu) Interdisciplinary and interdepartmental
Trinity of information science Computer Genome Brain
Key graph 2.009: Neuroinformatics Natural Images Brain measurements Statistical modelling
Key graph 2.009: Neuroinformatics Natural Images - Models of biological vision - Methods for image processing image retrieval - Natural language / text Brain measurements Statistical modelling
Key graph 2.009: Neuroinformatics Natural Images - Models of biological vision - Methods for image processing image retrieval - Natural language / text Brain measurements Statistical modelling - Latent variable models - Non-normalized models - Nonlinear multilayer models - Computational / statistical efficiency
Key graph 2.009: Neuroinformatics Natural Images - Models of biological vision - Methods for image processing image retrieval - Natural language / text Brain measurements - Brain imaging / behavioural - General / clinical - Temporal / spatial behaviour Statistical modelling - Latent variable models - Non-normalized models - Nonlinear multilayer models - Computational / statistical efficiency
Key graph 2.009: Neuroinformatics Natural Images - Models of biological vision - Methods for image processing image retrieval - Natural language / text search search Brain measurements - Brain imaging / behavioural - General / clinical - Temporal / spatial behaviour bioinfo Statistical modelling - Latent variable models - Non-normalized models - Nonlinear multilayer models HIIT BRU - Computational / statistical efficiency
End of mission
Statistical / data-driven neuroscience and intelligence Key assumptions: Brain is a Bayesian statistical analyzer Exploratory data analysis reveals unexpected and useful things in neuroscience Our emphasis: really understanding the brain As opposed to amateurish neural network research
Our approach Framework: The brain is a statistical analyzer: Sensory systems are adapted to natural stimuli Background: Extensive research on multivariate statistics, independent component analysis Goal: New data analysis methods To model brain function To get inspiration from brain function To be applied in neuroscientific data
Linear models of natural images = S 1 + S 2 +... + S N What are the best features (receptive fields) for natural images?
Principal component analysis of natural images
Independent component analysis Linear mixtures of source signals: can we find the original ones?
Independent component analysis
How does ICA work? Sums of independent component are more Gaussian: Maximize non-gaussianity cf. PCA: maximize variance Basis vectors can be estimated (but not in PCA!) Computationally demanding Many algorithms, e.g. FastICA (Hyvärinen, 1999)
Independent component analysis of natural images First model to show how features encoded by simple cells are statistically optimal features!
Real components are not independent Typical form of dependency of components: Correlations of squares We need better models that reflect the remaining dependencies
Example: Topographic ICA We are also able to model the spatial organization by looking at the dependencies of the components that ICA cannot cancel
Applications: neuroscience Functional understanding Quantitative understanding Towards predictive theory to guide experiments Repair of the senses: artificial retina, cochlea Repair of motor control for paralyzed
Applications: data analysis Brain is a great statistical analyzer, we can learn from it Self-organizing map, independent component analysis, multi-layer perceptrons etc. Conversely, sensory input is a very complex dataset whose analysis is informative
Applications: image processing Models of visual perception important to image processing Denoising Similarity measure (image queries) Compression Biologically-inspired computer vision Image databases: human similarity metric
Applications: sci-fi Direct input of signals to brain enables new senses Input a text document in a second Feel distance, position
Applications: sci-fi Direct input of signals to brain enables new senses Input a text document in a second Feel distance, position Direct output from the brain Control of a robot avatar Many new kinds of user interfaces Shared sensations
Conclusion Computational neuroscience is cool
Conclusion Computational neuroscience is cool One of the most important scientific challenges
Conclusion Computational neuroscience is cool One of the most important scientific challenges Multidisciplinary: e.g. Multivariate data analysis + visual neuroscience
Conclusion Computational neuroscience is cool One of the most important scientific challenges Multidisciplinary: e.g. Multivariate data analysis + visual neuroscience Everybody benefits from cross-fertilization
Conclusion Computational neuroscience is cool One of the most important scientific challenges Multidisciplinary: e.g. Multivariate data analysis + visual neuroscience Everybody benefits from cross-fertilization Good source of inspiration for statistical theory