My speciality is learning representations (sensory representations, abstractions, concepts, etc., akin to cortical learning). This is usually considered unsupervised learning but I believe that purely unsupervised learning is insufficient. The problem is that as the level of abstraction increases, the potential number of representations increases rapidly (combinatorial explosion of the model space). Therefore the learning process needs to be guided (modulated, biased) by the behavioural needs of the agent. This is similar to semi-supervised learning.
The method I have been developing is related to invariant-feature learning (e.g., slow-feature analysis by Wiskott) and independent component analysis (e.g., image features by Olshausen and Field or Hyvärinen et al.) but I'm trying to specifically take into account what kind of representations a behaving agent needs. This includes, for instance, visual representations that are useful in motor coordination. I've been working on a model where bottom-up processing is modulated by top-down (and temporal) context. Context carries information about behaviourally significant distinctions and teaches the bottom-up processing to extract features which are relevant to these distinctions.
Ref: H. Valpola. Behaviourally meaningful representations from normalisation and context-guided denoising. AI Lab technical report, University of Zurich, 2004.
One of the key ideas is that a behaving agent has, by design, task-specific modules which use the representations provided by a "neocortex". There are many tasks (such as prediction of reward, computation of retinal optic slip, computation of orientation of stimuli, prediction of movements, etc.) and consequently several different representations (coordinate systems) are needed. It is the task of specialised cortical areas to provide the representations needed by various task-specific modules.
In order for a cortical area to know which information is needed, the task-specific module provides inputs for the corresponding cortical area. The cortical area represents the information and associates it with other sources of information (implements "coordinate transforms" from one modality to another, e.g., from visual to motor, auditory or reward representation). Often the coordinate transformations are complex and highly nonlinear, requiring multistage processing. This is implemented by a hierarchy of cortical areas. The goal of the higher-level representations is to connect different lower-level representations and/or different time instants (being good at predicting).
One way of visualising the problem is that one needs to find the path (transformation) from one modality to another (the outputs of task-specific modules can be considered as "internal modalities"). Abstract concepts are important hubs connecting different modalities and time instances. Unsupervised learning based on general principles such as feature invariances can make good guesses about useful abstractions but since the path between different modalities can be long, there is a danger of missing each other (picture yourself digging two tunnels from opposite sides of a mountain). I'm using context to guide the learning process. Context can consist of top-down inputs from higher areas, lateral inputs from neighbouring areas (e.g., neighbouring modalities or neighbouring spatial locations within a single modality) and delayed inputs (providing ques about what is invariant or useful for prediction).
This section is still very much under construction but here is a skeleton:
Attention included in the model. Task-adapted sensory processing at different timescales (I'm working on learning but at perceptual timescale the analogous phenomenon is attention). Gustavo Deco's neurodynamical model of emergent attention is particularly interesting because the model structure is very similar (hiearchy of cortical areas, top-down modulation).
Planning etc. The representation should demonstrate internal dynamics in order to be able to run a "virtual reality simulation". This "higher-level cognition" could be controlled by the same mechanisms as immediate motor behaviour (cerebellum, basal ganglia).
Many of these topics have been discussed in my recent paper.
Once invariant features have been extracted, it is possible to represent objects as combinations of features. Sparse codes offer many benefits and it should be relatively straight-forward to combine invariant feature learning with sparse object representations. In fact, sparse codes may emerge automatically if long-range excitatory connections are combined with relatively strong local inhibition (as in Deco's neurodynamical model of attention). Expectation-guided learning should then develop sparse feature representations.
Multiple simultaneous objects (or even figure ground separation) probably require mechanisms such as synchronisation of the features belonging to individual (sensorimotor) objects.
I started my scientific career in 1993, after less than one year of studies, as a research assistant of Prof. Teuvo Kohonen who was then the head of the Laboratory of Computer and Information Sciences (CIS) at Helsinki University of Technology (HUT). Academician Kohonen is a pioneer in the field of unsupervised learning and neural networks. He has gained international reputation for his work on self-organising maps (SOM). My main task was to assist in the research and development of adaptive-subspace SOM which learns invariant-feature filters in an unsupervised manner. The model was inspired by the properties of so-called complex cells in the visual cortex.
I have always been interested in the brain--the very reason why I started my studies at HUT was to study neural networks at Kohonen's lab--and during my master's degree studies, I attended neuroscience courses given by the Brain Research Unit at Low Temperature Laboratory at HUT. The subject of my master's thesis, which I finished in 1996, was sparse coding, a representation scheme which is inspired by the neural coding in the brain.
I continued my doctoral studies at Neural Networks Research Centre (NNRC), the research branch of CIS which has been elected as a Center of Excellence by the Finnish Academy in 1995. For my minor I chose neurophysiology taught by Laboratory of Neurobiology at University of Helsinki to complement the studies in NNRC which cover learning from theoretical and engineering point of view. I aimed at an education which covers biological, theoretical and engineering aspects of learning and intelligence.
After finishing my doctoral thesis about a Bayesian approach to unsupervised learning of a nonlinear extension of factor analysis based on neural networks in 2000, I continued to supervise younger researchers of the Bayes Group which was established around the Bayesian research I had introduced to the lab. I also participated in the EU project BLISS which focused on independent component analysis, a research topic which had became one of the strong areas in CIS when Academy Prof. Erkki Oja took over the laboratory following Kohonen's retirement.
Although my education had covered reinforcement learning, I had no practical experience about motor learning. After consulting the robotics researchers in Automation Technology Laboratory at HUT, I decided to complement my education by post-doctoral training in Artificial Intelligence Laboratory (AI Lab) at University of Zurich. AI Lab, headed by Prof. Rolf Pfeifer, is one of the leading laboratories in the field of embodied cognitive sciences and it has good contacts with several robotics and neuroscience groups in Europe, North America and Japan.
Starting from September 2003, after the BLISS project had finished, I have been working on an EU project called ADAPT, first in Zurich and now in LCE. The goal in this project, which finishes in October 2005, is to study the learning of sensorimotor exploration strategies and emergence of multimodal representations in human infants and humanoid robots (only the torso of the humanoid robot has been implemented as locomotion is not studied and it is not important for infants during the first months of postnatal development). My responsibility in the project is the neural control architecture, the artificial brain of the robot.
Before going to Zurich, I had already agreed that I will continue the computational neuroscience research via robotics in Laboratory of Computational Engineering (LCE) at HUT. I started in LCE on November 2004. Like CIS, LCE hosts a Center of Excellence, Research Centre for Computational Science and Engineering. Its main strength is multidisciplinary research which applies mathematical modelling skills to physical, engineering and biological research.
The three focus areas of LCE are material, information and cognitive science. Particularly the research on adaptive machine vision for autonomous robots, led by Prof. Jouko Lampinen, psychophysics research on attention and perception and neurophysiology research on neural basis of auditory and visual speech perception, led by Academy Prof. Mikko Sams, are relevant to my research.
LCE has good resources in terms of equipment. Access to the 306-channel MEG and the high-resolution 3-tesla fMRI offer excellent opportunities for noninvasive imaging of the functional brain. The laboratory also has an autonomous robot build by the Intelligent Machines and Special Robotics Institute at HUT.