Geometric Morphometric Data Augmentation using Generative Computational Learning Algorithms

Abstract

The fossil record is notorious for being incomplete and distorted, frequently conditioning the type of knowledge that can be extracted from it. In many cases, this often leads to issues when performing complex statistical analyses, such as classification tasks, predictive modelling, and variance analyses, such as those used in Geometric Morphometrics. Here different Generative Adversarial Network architectures are experimented with, testing the effects of sample size and domain dimensionality on model performance. For model evaluation, robust statistical methods were used. Each of the algorithms were observed to produce realistic data. Generative Adversarial Networks using different loss functions produced multidimensional synthetic data significantly equivalent to the original training data. Conditional Generative Adversarial Networks were not as successful. The methods proposed are likely to reduce the impact of sample size and bias on a number of statistical learning applications. While Generative Adversarial Networks are not the solution to all sample-size related issues, combined with other pre-processing steps these limitations may be overcome. This presents a valuable means of augmenting geometric morphometric datasets for greater predictive visualization.

Developments in data science solutions for carnivore tooth pit classification

“…Firstly, prior augmentation of each dataset provided both algorithms with enough information to learn from, obtaining above average accuracy when used to classify the original samples. While the present datasets are unable to reach the 100% accuracy reported originally using SVMs 20 , this is likely due to the use of bootstrapping in the original study 41 . Here, more robust data augmentation techniques produced completely new synthetic data from which to learn from, providing a more general overview of the target domain.…”

“…Here we have shown how a number of different data science tools can be employed for GM analyses. From one perspective, unsupervised computational learning approaches were able to produce highly realistic augmented datasets, using both neural network based approaches 38 – 41 , as well as Bayesian Inference Engines 42 – 45 . While the use of Graphics Processing Units (GPUs) are likely to speed up GAN performance, MCMC can still be considered the fastest approach to modelling these datasets with exceptional synthetic-data quality.…”

“…Appendix 4 ). Moreover, to ensure the present study does not fall into the trap of over-generalising the Gaussian nature of the population, the precise definitions of our target probability distributions were robustly defined 41 , 46 , 47 .…”

“…Data augmentation is a very popular technique in computer vision, nevertheless, not all of these algorithms are readily applicable to numeric data of this type 48 . Here augmentation has been shown to not only improve the accuracy of most models (Tables 2 , S8 - S12 & S23 - S28 , Supplementary Appendix 7 ), but also improve the generalization capabilities of both SVMs and NSVMs 41 , 49 . This is mostly seen through the decrease in loss values across taxa (Tables S8 - S12 & S23 - S28 ), thus supporting observations made by Courtenay and González-Aguilera 41 when applied to other GM datasets of palaeoanthropological and primatological origin.…”

“…Here augmentation has been shown to not only improve the accuracy of most models (Tables 2 , S8 - S12 & S23 - S28 , Supplementary Appendix 7 ), but also improve the generalization capabilities of both SVMs and NSVMs 41 , 49 . This is mostly seen through the decrease in loss values across taxa (Tables S8 - S12 & S23 - S28 ), thus supporting observations made by Courtenay and González-Aguilera 41 when applied to other GM datasets of palaeoanthropological and primatological origin. Similarly, the impact dataset imbalance has on algorithm performance is clearly evident, as seen through great drops in precision, recall and F values (Supplementary Appendix 6 , Tables S8 - S12 & S23 - S28 ).…”