The key to understanding vision is to acquire insight into the sensory coding of indi- vidual neurons. To this end, major advances were done over the past 50 years in fitting models to neural data to identify the mapping from sensory space to neural responses. Especially the advance of DNNs in neuroscience allowed for model fits with excellent predictive power. However, such advanced neural models are complex, and their poor in- terpretability has so far hindered deeper understanding of the principles of visual coding. To address this issue, a recent study proposed a method which identifies the stimulus that activates the neuron the most. However, the sensory coding of highly non-linear neurons, which are abundant already at the earliest stages of visual processing, is too complex for a single stimulus to sufficiently characterize it. A more robust way to char- acterize this coding is through identifying the input sub-space within which the neuron is activated identically - i.e. finding invariances of the neuron's sensory representation. In this thesis, a novel approach for finding such invariant stimuli is proposed. The proposed technique is based on a generator neural network, which maps Gaussian noise from latent space to a stimulus set which equally activates a given neuron. The method demonstrated the...
Identifer | oai:union.ndltd.org:nusl.cz/oai:invenio.nusl.cz:448118 |
Date | January 2021 |
Creators | Kovács, Peter |
Contributors | Antolík, Ján, Šikudová, Elena |
Source Sets | Czech ETDs |
Language | English |
Detected Language | English |
Type | info:eu-repo/semantics/masterThesis |
Rights | info:eu-repo/semantics/restrictedAccess |
Page generated in 0.0017 seconds