• Heta Desai

Rapid advancements in machine learning and artificial intelligence (AI) have sparked growing concern about the potential societal impacts of these technologies. This paper addresses one such impact: the issue of accidents in machine learning systems, which refers to unintended and harmful behaviors that may arise from poorly designed real-world AI systems. We outline five key research challenges related to accident risk, categorized based on their source: problems arising from incorrect objective functions ("avoiding side effects" and "avoiding reward hacking"), issues stemming from objective functions that are costly to evaluate ("scalable supervision"), and undesirable behaviors during the learning process ("safe exploration" and "distributional shift"). We review prior research in these areas, propose new research directions, and emphasize their relevance to advanced AI systems. Finally, we discuss the broader question of how to effectively approach the safety of future AI applications.

Practical Challenges and Research Directions in Ensuring AI Safety

"Practical Challenges and Research Directions in Ensuring AI Safety", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.4, Issue 3, page no.831-840, March-2017, Available :http://www.jetir.org/papers/JETIR1703139.pdf

"Practical Challenges and Research Directions in Ensuring AI Safety", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.4, Issue 3, page no. pp831-840, March-2017, Available at : http://www.jetir.org/papers/JETIR1703139.pdf