Description of original award (Fiscal Year 2017, $44,348)
This research will address a need of the forensic sciences using recent developments in machine learning - deep learning. Deep learning algorithms are expected to perform so called end-to-end learning, which means it does not involve any human crafting or human engineering during the process of feature learning.
By learning a better representation of the data, deep learning algorithms are able to automatically generate features from the data and achieve better performance on forensic problems such as verification.
This project will evaluate several architectures for the forensic comparison task. Beginning with human extracted features for handwriting and machine extracted features for footwear impressions and then proceed to raw images as input. If successful with raw images, it would circumvent the need for human and machine effort in feature extraction and would achieve the goal of end-to-end learning.
Exploration of both conventional neural network and convolutional neural network (CNN)architectures. One design consists of separate networks for the two inputs with shared weights, with the results combined at later stages. Another design would interlace the two inputs so that minor feature differences can be learned, while it does not lose information through a distance/kernel function. The usage of all kinds of regularization methods such as weight-decay, dropout, data augmentation to curb over-fitting and improve performance of learning.
Data analysis techniques include preprocessing the raw image data, augmenting training data and evaluation of performance of the trained model using loss functions such as cross entropy.
For this part, the major product of this research will be publications to major forensic science journals.