Abstract
The receptive field is one of the most common ways to characterize the function of a neuron. Yet, receptive fields may contain further, elaborate substructure, as found, for example, in retinal ganglion cells (RGCs) [1, 2]. A method for detecting this substructure without detailed measurements of the anatomical and physiological cell properties is currently lacking.
Here, we propose a novel method, called sample-based semi non-negative matrix tri-factorization (ssNM3F), for identifying substructure in receptive fields. This method is derived from semi non-negative matrix factorization [3] and sample-based non-negative matrix tri-factorization (sNM3F) [4]. Like sNM3F, our method decomposes data into non-negative space-by-time components and activation coefficients for those components. Unlike sNM3F, however, activation coefficients can also be negative.
We present a proof of convergence for ssNM3F. We then demonstrate the technique on data that were recorded extracellularly from salamander RGCs. The data consist of the spike-triggered stimulus ensembles of individual cells, obtained under stimulation with spatio-temporal white noise. We find that ssNM3F can decompose the spike-triggered ensemble into factorized spatial and temporal components and associated activation coefficients. Variance-accounted-for (VAF) increases with increasing number of spatial and temporal components. We also find that the relaxation of the non-negativity constraint for the activation coefficients is a crucial prerequisite for increasing the VAF with increasing number of components. The number of components is selected by minimizing the Akaike information criterion. In our dataset, we find that, for many cells, the optimal number of spatial components is greater than unity.
These results suggest that ssNM3F can successfully detect substructure in receptive fields of RGCs and that semi-sNM3F can thereby provide a more detailed characterization of the spike-triggered ensemble.
Here, we propose a novel method, called sample-based semi non-negative matrix tri-factorization (ssNM3F), for identifying substructure in receptive fields. This method is derived from semi non-negative matrix factorization [3] and sample-based non-negative matrix tri-factorization (sNM3F) [4]. Like sNM3F, our method decomposes data into non-negative space-by-time components and activation coefficients for those components. Unlike sNM3F, however, activation coefficients can also be negative.
We present a proof of convergence for ssNM3F. We then demonstrate the technique on data that were recorded extracellularly from salamander RGCs. The data consist of the spike-triggered stimulus ensembles of individual cells, obtained under stimulation with spatio-temporal white noise. We find that ssNM3F can decompose the spike-triggered ensemble into factorized spatial and temporal components and associated activation coefficients. Variance-accounted-for (VAF) increases with increasing number of spatial and temporal components. We also find that the relaxation of the non-negativity constraint for the activation coefficients is a crucial prerequisite for increasing the VAF with increasing number of components. The number of components is selected by minimizing the Akaike information criterion. In our dataset, we find that, for many cells, the optimal number of spatial components is greater than unity.
These results suggest that ssNM3F can successfully detect substructure in receptive fields of RGCs and that semi-sNM3F can thereby provide a more detailed characterization of the spike-triggered ensemble.
| Original language | English |
|---|---|
| Number of pages | 1 |
| DOIs | |
| Publication status | Published - 5 Sep 2014 |
| Event | Bernstein Conference 2014 - Duration: 3 Sep 2014 → 5 Sep 2014 |
Conference
| Conference | Bernstein Conference 2014 |
|---|---|
| Period | 3/09/14 → 5/09/14 |