Tool for explainable face biometrics, neural networks open-sourced by TruEra
TruEra has made its tool for explainability in machine learning models based on neural networks, like many biometric systems, available as open source software, according to a company announcement.
The new TruLens provides a uniform API for explaining models built with Tensorflow, Pytorch and Keras, with a uniform abstraction layer. The library provides a coherent and consistent approach to explaining deep neural networks, the company says, based on public research. TruLens also natively supports internal explanations, such as what visual concepts a facial recognition model is drawing on to identify people from images.
The tool is inspired in significant part by the paper ‘Influence-Directed Explanations for Deep Convolutional Networks’ by the creators of Carnegie Mellon University’s library, TruEra says.
Use cases for TruLens include explanations for computer vision models like object recognition and face biometrics, natural language processing like identifying malicious speech or smart assistants, forecasting and personalized recommendations.
“Image recognition and text recognition machine learning models are both highly in demand and have a lot of consumer wariness about them, due to highly publicized stories about error or possible misuse,” says Shayak Sen, co-founder and CTO, TruEra. “The recent European Commission regulations specifically listed cautions around machine learning models and how they deal with personal data or images. So there is a huge need for explainability for these types of models, to ensure that they are effective, but also compliant and easily explained to a concerned society. We feel strongly about the ethical use of AI, and wanted to make TruLens freely available to the world to help ensure responsible adoption of AI for uses like image recognition.”
TruLens is the product of eight years of explainability research conducted at both Carnegie Mellon University and TruEra, and is available now.
Explainability has been recognized as a necessary feature to increase the trustworthiness of artificial intelligence in general, and biometrics in particular.