Three Human-Centered Deep Learning Design Principles

2019-12-17

Over the past decade, many companies have begun to rely on AI algorithms to assist in making business decisions, from delivering logistics, route planning, risk detection to financial fraud detection and image recognition. We are now experiencing the end of the second wave of AI and entering the third wave, known as perceptual AI. In the next wave, a specific subset of AI called deep learning will play a more critical role.Like other forms of AI, deep learning will self-adjust and learn by generating outputs using datasets and comparing them with empirical facts. As companies adopt deep learning, leadership must ensure the accuracy and precision of artificial neural networks, as tuned networks can impact business decisions and potentially harm customers, products, and services.


As we enter the next phase, a key question for businesses will be how to embrace deep learning to drive better business decisions while avoiding biases and potential adverse outcomes. Through collaboration with numerous clients across multiple industries, we have identified patterns that can help companies reduce error rates when implementing deep learning initiatives.Our experience working with businesses in these early stages of AI adoption has helped us establish design principles for human-centered deep learning ethics, with a focus on adjusting the data of the network. Human-centered design approaches can address both short-term issues (well-trained AI networks producing false solutions) and long-term concerns, such as machines potentially replacing humans in making business decisions.When we talk about human-centered design, what we expect is to bring benefits to all individuals and teams, rather than at the expense of sacrificing a few individuals' interests. Our human-centered design principles support the goal of providing and informing data to give people more opportunities in their work. Based on our experience, businesses need to adhere to three key principles as the pillars of any AI implementation:


Transparency


Make the high-level implementation details of your AI project available to all relevant stakeholders as much as possible. In the case of this deep learning initiative, people should understand what deep learning is, how it works (including how algorithms are adjusted using datasets), and how deep learning affects their work. When there is a possibility of exposing intellectual property or other sensitive information, companies may want to include an external stakeholder group. Remember that if certain datasets contain sensitive information or privacy concerns, they may need to be protected from disclosure.


Interpretability


Employees and external stakeholders in a company, including potential customers, should be able to understand how any deep learning system makes its contextual decisions. The focus here is not on explaining how the machine arrives at its conclusions (deep learning is often not explainable at that level), but more on explaining the methods used to adjust the involved algorithms, which datasets are used, and the conclusions human decision-makers draw on how to use the algorithms.


Reversibility


Companies must also have the ability to reverse the content of deep learning. Treating it as the ability to unlearn certain knowledge or data helps prevent unnecessary biases in datasets. Reversibility is a concept that must be designed into AI work and often requires cross-functional expertise and support.


Companies can adopt three approaches to put these principles into practice in their AI projects. These approaches aim to reduce the introduction of poorly tuned AI systems and the risks of incorrect or biased decisions during piloting and implementation.

RECENT AI NEWS

RECENT AI TOOLS