Using Active Learning to Expand Training Data for Implicit Discourse Relation Recognition

Yang Xu, Yu Hong, Huibin Ruan, Jianmin Yao, Min Zhang, Guodong Zhou

Abstract: Author Summary The ever growing size of neural populations simultaneously recorded in electrophysiological experiments calls for urgent analytical progress in understanding how to compactly describe all sensory information present both in the spatial and temporal structure of single-trial neural population activity. Here we show the power of analytical methods, termed space-by-time tensor factorizations, which detect groups of simultaneously coactive neurons and the temporal profiles of their coactivation. By validation on simulated data and on retinal recordings, we show that the tensor decomposition performs competitively compared to other techniques both in terms of data robustness and ability to find informative patterns across diverse stimuli. We show that this method can determine the spatial and temporal resolution of neural population codes, and find which spatial or temporal components of neural responses carry information not available in other aspects of the population code. When applied to experimental data, the method demonstrates the importance of first-spike latencies in retinal population coding of visual images, particularly for decoding fine spatial details of natural images from population activity. This work shows that this methodology can improve our knowledge of population coding by allowing the discovery of informative spatial and temporal firing patterns in populations of simultaneously recorded neurons.