Mathematical modelling of neuronal learning

Research focus

Rather than investing in more powerful muscles, better sensors, or body plans that could take them into the air, our ancestors grew big expensive brains. To be useful at all, these brains have to configure themselves at the beginning of (and throughout) their lifetimes. Our research vision rests upon our conviction that the success story of big brains (which did, eventually, take us to the skies and the moon) is tightly linked to the co-evolution of learning algorithms that allow these incredibly adaptable machines to find the right set of parameters within the vast space of possible settings that control their function and behavior.

Our goal is to uncover these learning algorithms and to understand how they are implemented by cellular processes. Specifically, we study learning in spiking neural networks whose units communicate, as do the neurons in our brains, via all-or-none action potentials. By integrating different levels of abstraction within a given computational perspective, our approach bridges between cellular biophysics, synaptic plasticity, neural systems function and machine learning.

Our current projects center around the recently developed tempotron family of spiking neuronal network models and cover a broad range of topics including mathematical analyzes of information processing in spiking neuronal networks, spike-based learning in single and multi-layer neuronal networks, sensory spike data analysis, temporal processing with short term synaptic dynamics, as well as applied development of visual and speech processing systems.

Discovering features in continuous data streams. (A) Schematic of the input activity of a tempotron model neuron (blue circle on the right). Each line depicts the spiking activity (black vertical ticks) of one input afferent. Colored rectangles (top) represent the occurrences of different features that are represented by stereotyped segments of spiking activity within a stream of random background activity. (B) Membrane voltage traces of the model neuron before training (top trace) and after being trained to fire one spike in response to the blue feature (second trace), five spikes in response to the blue feature (third trace), one spike in response to the blue, light blue, green, yellow, and red features (fourth trace) and one spike in response to the blue, two in response to the light blue, three in response to the green, four in response to the yellow, and five in response to the red feature. During training the neuron was only instructed whether it responded with fewer or more spikes than the correct number of output spikes. No information about the presence of features, such as their number or timing, was provided.

Publications (Selection)

Gütig R (2016). Spiking neurons can discover predictive features by aggregate-label learning. Science, 351, 1041. doi: 10.1126/science.aab4113.

Gütig R, Sompolinsky H (2009). Time-warp-invariant neuronal processing. PLoS Biology, 7, e1000141. doi: 10.1371/journal.pbio.1000141.

Gütig R, Sompolinsky H. (2006). The tempotron: a neuron that learns spike timing-based decisions. Nature Neuroscience, 9, 420-428. doi: 10.1038/nn1643.

Gütig R, Aharonov R, Rotter S, Sompolinsky H (2003). Learning input correlations through non-linear temporally asymmetric Hebbian plasticity. Journal of Neuroscience, 23, 3697-3714. doi: 10.1523/JNEUROSCI.23-09-03697.2003.