PAPIs 2018

Using AI for testing software is an emerging field in software engineering. So, how to do it effectively? Yet a big problem to be solved. In this presentation, we will describe how Convolutional Neural Networks (CNN) and Deep Reinforcement Learning can be used for this new endeavor by outlining the challenges, mistakes, and workarounds that we faced to be successful in using AI models to build systems that can really learn from the software they are testing. We will discuss some lessons learned when using pre-trained CNN models, Image Detection APIs and CNN's built from scratch for this purpose.

ML is now a commercial and industrial technology, and while many successful algorithms exist there are still areas that require new developments. For instance, one issue with many ML methods is that they generate black-box models. Genetic Programming (GP) generates symbolic models an expressions, that can be used in many different domains. However, GP is not widely used by ML practitioners , it is still mostly an academic tool, but this is changing. This talk will present a short overview of how GP can be used to solve ML tasks, intended as a starting point for applied researchers and developers.

Despite breakthroughs in statistical performance, the widespread adoption of algorithmic decision making has also led to a rise in “black box” machine learning with unintended negative consequences. In response, attention towards methods that improve the interpretability and understanding of machine learning has also increased. These methods are useful not only for explaining how decisions are made, but also for improving models and ultimately gaining trust in adopting machine learning systems. While using interpretable machine learning methods has innate advantages, the explicit requirement of interpretability has only recently been formalized as part of the General Data Protection Regulation, implemented by the EU as of 25 May 2018. This regulation, which protects the privacy and usage of data of EU citizens, specifically outlines a “right to explanation” in regards to algorithmic decision making. This talk explores the definition of interpretability in machine learning, the trade-offs with complexity and performance, and surveys the major methods used to interpret and explain machine learning models in the context of GDPR compliance.

What are the policy and societal implications of the unprecedented capability for automated, real time identification and tracking of individuals? What tools exist or could exist for registering and enforcing user choice? What should public policy be around government and private use of biometric data? We demonstrate the technical feasibility of facial recognition adversarial attacks, describe our FOIA request about federal use of facial recognition at airports and borders and invite community discussion and technical contributions to our open sourced prototype.

The development and deployment of machine learning models is fraught with complexity exceeding that which is typically found in traditional software. Operating in a clinical environment introduces further difficulties that must be resolved to produce a successful product. In this talk, we will discuss the challenges of applying machine learning to medical imaging and highlight potential solutions to these problems.

The key to building systems that are integrated into people’s lives is trust. If you don’t have the right amount of trust, you open the system up to disuse and misuse. We will discuss the building blocks of AI from a product/design perspective, what trust is, how trust is gained, and maybe more importantly lost, and techniques you can use day-to-day to build trusted AI products. We will reference real examples from academia, industry, and my work at Philosophie.

Find the presentation here:

https://goo.gl/B2TQt6