Search

Understanding the inner behaviors of deep neural networks is an important issue, and various visualization methods have been suggested. However, recent studies showed that most existing saliency methods have low visualization accuracy, even compared to a random importance assignment. In this study, we seek an improved saliency algorithm. Focusing on class activation mapping (CAM)-based saliency methods, we discuss two problems with the existing studies. First, we introduce conservativeness, a property that prevents redundancy and deficiency in saliency map and ensures that the saliency map is on the same scale as the prediction score. We identify that existing CAM studies do not satisfy the conservativeness and derive a new CAM equation with the improved theoretical property. Second, we discuss the common practice of using bilinear upsampling as problematic. We propose Gaussian upsampling, an improved upsampling method that reflects deep neural networks' properties. Based on these two options, we propose Extended-CAM, an advanced CAM-based visualization method. In various visualization benchmarks, datasets, and architectures, our Extended-CAM presents a more accurate visualization than is obtainable with other existing methods. • Existing class activation mapping (CAM) was observed as inaccurate. • Extended-CAM, a theoretically grounded new CAM method, was developed. • For the natural upsampling of CAM, Gaussian upsampling was proposed. • Extended-CAM showed more accurate visualization than the existing methods. [ABSTRACT FROM AUTHOR]

Abstract: COUP-TFII (also known as Nr2f2), a member of the nuclear orphan receptor superfamily, is expressed in several regions of the central nervous system (CNS), including the ventral thalamus, hypothalamus, midbrain, pons, and spinal cord. To address the function of COUP-TFII in the CNS, we generated conditional COUP-TFII knockout mice using a tissue-specific NSE-Cre recombinase. Ablation of COUP-TFII in the brain resulted in malformation of the lobule VI in the cerebellum and a decrease in differentiation of cerebellar neurons and cerebellar growth. The decrease in cerebellar growth in NSECre/+/CIIF/F mice is due to reduced proliferation and increased apoptosis in granule cell precursors (GCPs). Additional studies demonstrated that insulin like growth factor 1 (IGF-1) expression was reduced in the cerebellum of NSECre/+/CIIF/F mice, thereby leading to decreased Akt1 and GSK-3β activities, and the reduced expression of mTOR. Using ChIP assays, we demonstrated that COUP-TFII was recruited to the promoter region of IGF-1 in a Sp1-dependent manner. In addition, dendritic branching of Purkinje cells was decreased in the mutant mice. Thus, our results indicate that COUP-TFII regulates growth and maturation of the mouse postnatal cerebellum through modulation of IGF-1 expression. [Copyright &y& Elsevier]