Computational neuroscience; learning; memory; neural networks; reinforcement learning; working memory; biophysics; sequence learning; spiking neuron models; synaptic plasticity models; hippocampus
Khajeh-Alijani Azadeh, Urbanzcik Robert, Senn Walter (2015), Scale-Free Navigational Planning by Neuronal Traveling Waves, in PLoS ONE
, 10(7), 1-15.
Johanni Brea Walter Senn & Jean-Pascal Pfister (2013), Matching Recall and Storage in Sequence Learning with Spiking Neural Networks, in The Journal of Neuroscience
, 33(23), 9565-9575.
J. Brea W. Senn and J.-P. Pfister (2012), Sequence learning with hidden units in spiking neural networks, in Advances in Neural Information Processing Systems NIPS
, MIT Press, http://papers.nips.cc.
Johannes Friedrich & Walter Senn (2012), Spike-based Decision Learning of Nash Equilibria in Two-Player Games, in PLoS Computational Biology
, 8(9), 1-12.
Torsten Lüdge Robert Urbanczik Walter Senn, Modulation of orientation-selective Neurons by motion: when additive, when multiplicative?, in Frontiers in Computational Neuroscience
, 8(67), 1-12.
At any moment in time, various types of sensory information impinge on the brain, where it is integrated with internally generated activity. How these signal streams interact on a neuronal level in space and time, and how the brain is able to keep sensory and expectation signals apart, represents a largely unsolved problem. The proposal explores through mathematical modeling and computer simulations how such binding can be accomplished via dendritic gain modulation and time multiplexing, without loosing spatial and temporal feature identities. We investigate how sensory signals of different qualities can modulate neuronal firing, and how past, present and future events can be represented in the same neurons at specific time slots relative to a global background oscillation.The project splits into two subprojects, one for a PhD student and one for a postdoctoral student.(A) Dendritic pointers. The PhD project builds on our earlier work on task-dependent gain modulation of pyramidal neurons in the primary visual cortex (V1). We consider moving visual stimuli and hypothesize that these are analyzed based on a series of static images obtained through stroboscopic snap-shots. Motion information is integrated in V1 and in higher cortical areas and fed back to modulate the static image analysis. Unlike a global attentional signal, however, this modulation can specifically highlight the spots of the actual and expected motion. The motion signal is assumed to transiently modulate the gain of V1 pyramidal neurons through dendritic calcium spikes. We build a network model of V1 which is able to reconstruct edge orientations in moving images based on motion integration, although the motion signal itself only highlights edge location. We also formulate a mathematical framework in which we derive the type of interactions between motion and orientation selective V1 neurons which best serves contour enhancement. The predictions are planned to be tested based on rat in vivo experiments performed by M. Larkum.(B) Time multiplexing. The postdoc project will explore temporal binding mechanisms in the context of sequence learning and predicting distal future events. To bridge the temporal gap between past, present and future states of a sequence we consider the binding of these states in individual theta cycles (with period length of roughly 150ms). While a new stimulus drives the corresponding neuronal assembly at a certain phase within the theta cycle, past events are represented at earlier and expected future events at later phases. This leads to a time-compressed representation of sequences by nesting several gamma cycles (of period length 20ms) into one theta cycle, as has been observed in hippocampal place field recordings during navigation. We will extend these ideas and apply them to the prediction of non-spatial and non-Markovian stimulus sequences where the choice of a next stimulus may depend on a combination of previous stimuli. To prevent a temporal low-pass filtering of the stimulus information we suggest to multiply represent the same sequence across different areas, each endowed with its own phase shift. This allows us to predict close and distal events with the same precision. Predicting future events instead of future reward moreover supports a fast re-evaluation of putative decisions in a quickly changing context. We will evaluate the model based on in vivo recordings from delayed match-to-sample experiments planned by G. Rainer.