May 15th 2024.
In today's society, the concept of bias is widely acknowledged and discussed. It's something that is not only evident in human interactions, but also in artificial intelligence systems. We have seen countless examples of racial and gender biases in various fields, such as law, medicine, and finance. It's a complex issue that has sparked many discussions and debates. But what if we could eliminate bias altogether? Would that solve all our problems? According to the late Nobel laureate Daniel Kahneman, it's not that simple. In fact, he argued that there are two sources of errors in judgments: bias and noise.
As computer and information scientists, my colleagues and I have delved into this topic and found that noise plays a significant role in AI as well. But what exactly is noise? In this context, it refers to the variation in how people make judgments of the same problem or situation. And it's more pervasive than we initially thought. For instance, a study dating back to the Great Depression found that different judges gave different sentences for similar cases. This raises concerns about the fairness and consistency of our justice system. It's not just biased, but also arbitrary at times.
Noise can also be observed in other fields, such as insurance and beauty pageants. Even in AI, where we expect machines to make unbiased decisions, noise can still impact their performance. This may seem counterintuitive since machines are not affected by external factors like weather or football games. However, we must remember that AI is trained on data that may contain biases. And as a result, these biases can manifest in the AI's decisions.
Take the example of the new AI model, ChatGPT, which is measured against human performance on general intelligence problems. To determine its level of common sense, researchers compare its answers to those of humans. But what about questions with more uncertainty or disagreement? This is where noise can come into play. And unfortunately, most AI tests do not account for this noise, which can significantly impact the results.
To address this issue, my colleagues and I conducted a study and published our findings in Nature Scientific Reports. We discovered that even in the domain of common sense, noise is inevitable. Our study involved two types of experiments: one with paid workers from Amazon Mechanical Turk and the other with a smaller-scale labelling exercise. Both settings showed a nontrivial amount of noise, which can impact an AI system's performance by 4-10%.
This means that even a seemingly significant improvement in an AI system's performance may not be entirely accurate due to noise. And as we see in AI leaderboards, where the differences between rival systems are narrow, noise can make it challenging to distinguish the effects of true performance improvements from those of noise.
So what's the solution? Kahneman proposed the concept of a "noise audit" for quantifying and mitigating noise in AI systems. Just like how we audit AI for bias, we should also consider auditing for noise. It's crucial to estimate the influence of noise on an AI system's performance to ensure its accuracy and fairness.
In conclusion, while we often focus on eliminating bias, we must also consider the effects of noise in AI systems. It's a complex issue that requires further research and attention. As AI continues to advance and play a more significant role in our lives, we must address this problem and strive for more accurate and fair AI systems.
[This article has been trending online recently and has been generated with AI. Your feed is customized.]
[Generative AI is experimental.]
