@inproceedings {Savin1824_2016,
	year = {2016},
	author = {Savin, Cristina and Monk, Travis and Lücke, Jörg},
	title = {Intrinsic plasticity for optimal learning of variable stimulus intensities},
	booktitle = {Int. Conf. Computational Systems Neuroscience (COSYNE)},
	abstract = {In many situations the meaning of a stimulus is the same despite fluctuations in its overall strength.  A visual
scene’s content does not depend on light intensity, or a word utterance should be recognised irrespective of its
loudness. Nonetheless, gain fluctuations are an integral part of the input statistics and they can help differentiate
between stimuli.  In the visual domain, for instance objects of the same class are likely to have similar surface
properties, resulting in a distinct distribution of light intensities. Light intensities can therefore help identify objects.
The  neural  underpinnings  of  such  computation  are  unclear.   Existing  models  discard  gain  information  by  ad
hoc preprocessing (Nessler, 2009; Keck, 2012) or by divisive normalisation (Schwarz, 2001) before learning the
input statistics from normalized data.  Overall, it is unknown how neural circuits can robustly extract statistical
regularities in their inputs when the overall intensity of stimuli is variable.  Here we develop a principled account
of unsupervised learning in the face of gain variations. We introduce a novel generative mixture model (Product-
Poisson-Gamma) that explicitly models the statistics of stimulus intensity, and we derive a biologically-plausible
neural circuit implementation for inference and learning in this model.  We find that explicitly taking into account
gain variations improves the robustness of unsupervised learning, as differences in input strength help distinguish
between classes with similar features but different gain statistics. From a biological perspective, the derived neural
circuit, in which feature-sensitive neurons are equipped with a specific form of intrinsic plasticity (IP), provides
novel insights into the interaction between Hebbian and IP during learning.  Furthermore, our results imply that
neural excitability reflects nontrivial input statistics, in particular the intensity of the features to which a neuron is
sensitive.}
}