21 President Oral Presentation Pick: Using Deep Neural Networks to determine birth weight data quality for genetic evaluations in beef cattle. (30th November 2020)

• Record Type: Journal Article
• Authors: Ribeiro, Andre; Golden, Bruce L; Spangler, Matthew L
• Is Part Of: Journal of animal science. Volume 98(2020)Supplement 4
• Page Start: 7
• Page End: 7
• Publication Date: 2020-11-30
• Subjects: Beef Cattle -- Deep Neural Network -- Genetic Prediction; Livestock -- Periodicals; Livestock; Electronic journals; Periodicals; 636.005
• DOI: 10.1093/jas/skaa278.013
• Languages: English
• ISSNs: 0021-8812

Abstract

Abstract: The objective of this work was to evaluate the use of deep neural networks (DNN) for classifying contemporary groups based on the method used to generate birth weight (BWT) phenotypes. Contemporary groups (CG; n = 120, 000) ranging between 10 and 500 animals were simulated assuming 12 data collection and CG formation scenarios that could impact CG phenotypic variance, including weights recorded with a digital scale (REAL), hoof tape (TAPE), and those that were fabricated (FAB). The performance of 6 activation functions (AF; ReLu, sigmoid, exponential, ReLu6, Softmax, Softplus) were evaluated. Four hidden layers were used with 7 different scenarios relative to the number of neurons. The training procedure was implemented in Python 3 with TensorFlow 1.14 and the ADAM optimization. Simulated CG were divided into training (80%) and testing (20%). The correlation between the observed and predicted CG types, averaged across 10 replicates, was used to assess accuracy and the correlations of predictions between replicates were used to measure the consistency of the model. In general, accuracy across AF and numbers of neurons were similar, with mean correlations ranging between 0.91 and 0.99. The AF ReLu, Sigmoid, Exponential and ReLu6 had the greatest consistency between the replicates, with an average correlation greater than 0.90. Independent of the number of neurons used, the sigmoid function produced the highest accuracy (0.99) and consistency (0.96). The DNN was retrained using 10-fold the number of CG and mimicking the CG size distribution observed in real data obtained from the American Hereford Association (n = 46, 177 CG). In the real data, the lowest phenotypic variance was for FAB CG (2.98 kg 2 ), REAL CG had the largest (18.33 kg 2 ) and TAPE CG was intermediate (8.64 kg 2 ). Results suggest that a well-trained DNN can be effectively used to classify data based on quality metrics prior to inclusion in routine genetic evaluation.