Over the last few years, the progress in Artificial Intelligence (AI) has been driven to a considerable extent by deep learning (DL) and neural networks. The application areas have been explosive, leading to a cult status for the discipline. Just to cite a few examples reflected in recent headlines:
- Driverless cars are on the road without any fallback human driver onboard.
- AI-based medical diagnostic software is performing better than certified doctors, which is now recognized and published in leading medical journals.
- Digital conversational agents are becoming more and more pervasive in all spheres of life.
- AI-enabled agents are beating world champions in various complex games.
While AI has made significant technological advances, there are certain inherent issues that must be systematically addressed as part of the transformation from research to real life. These issues include:
Explainability: As AI is increasingly becoming part of our real-life applications, decisions made by AI-enabled systems need to be explainable. For example, why was a loan applicant rejected? On what basis was a patient diagnosed positive for a malignant disease? The ability to provide justifiable and reliable evidence-based decisions would increase the trust of AI users.
Fragility: AI/DL systems in various settings have occasionally been reported to be functioning in surprising and undesirable manners when put in a different but related environment. For example, very minor changes in the input data, even one as minor as one pixel in an image, can lead to very different output which may have disastrous implications for real-life applications. The heavy reliance on fine-tuning of the AI parameters brings into question the applicability of theoretical work in real-life applications.
Research Practices: While the sharing of code and implementation details has recently increased, the ability to reproduce experiments from research papers has been challenging. This is often due to exaggerated claims of empirical evaluation from papers and the increasing complexity of models. In addition, publication biases towards positive results have not been serving the community as a whole.
Some of the above-mentioned challenges such as explainability and fragility are unique to the AI field, while other challenges such as research practices are common across other scientific disciplines as well. Some of these challenges have been encountered early on in the evolution of psychology and medicine. Research communities in those disciplines have come up with structured methods to address these issues. For instance, the field of psychology created the “Registered Reports'' process.
A Registered Report is a form of scientific publication in which methods and proposed analyses are pre-registered and peer-reviewed prior to research being conducted (referred to as Stage 1). Research protocols specified in Stage 1 are provisionally accepted for publication before data collection and experimentation begins. Once the study is completed, the author will submit the completed article including results and discussion sections (referred to as Stage 2). Provided the study adhered to the agreed upon protocols, the article is then accepted for publication. This can help avoid reviewers' bias towards experimental improvements without giving adequate weight for novelty and correctness.
Controlled research in phased clinical trials has helped mitigate the risk of “unknown unknowns'' in the pharma industry. Similarly, the role of a federated approving authority such as the Food and Drug Administration (FDA) in the United States may be relevant for AI use when it has the potential to impact human safety.
We believe that the AI community needs to leverage many of these interdisciplinary learnings as it continues to evolve from being a cloistered academic research discipline to a major presence in human activities. Our paper published in 2018 International Joint Conference on Artificial Intelligence discusses the various challenges that AI faces in its evolution from research to real life and describe how they can be addressed.
For more information on artificial intelligence and additional technology and innovations, visit us here.
About the AuthorMore Content by Sandya Mannarswamy