An Alarming Bias Is Growing With AI

NAB

article here

More money and more capability, yes, but more bias and more men. That’s the questionable state of Artificial Intelligence and Machine Learning in 2022, according to the latest annual report from a team at Stanford.

The “2022 AI Index Report” measures and evaluates the rapid rate of AI advancement from research and development to technical performance and ethics, the economy and education, AI policy and governance, and more.

It is compiled and analyzed by a group of experts from across academia and industry at the Stanford Institute for Human-Centered Artificial Intelligence (HAI).

Here’s a short summary of key findings.

Private Investment in AI Soars

The amount of money pouring into AI is mind-boggling — and most of it is coming from private investment. Private spend more than doubled in 2021 from 2020 to around $93.5 billion, but the number of newly funded AI companies continues to drop, from 1051 in 2019 and 762 companies in 2020 to 746 companies in 2021.

As IEEE puts it, “It’s a great time to join an AI startup, but maybe not to found one yourself.”

Language Models are More Capable but More Biased

Large language models like Open-AI’s GPT-2 are setting new records on technical benchmarks, but new data shows that they are also more capable of reflecting biases from their training data.

While language systems are growing significantly more capable over time, the Stanford team conclude, “so does the potential severity of their biases.”

Their term for this bias is “toxicity.” As an example, the report notes that a 280 billion parameter model developed in 2021 shows a 29% “increase in elicited toxicity” over that of a comparable model in 2018.

A number of research groups are working on the toxic-language problem AI presents, with both new benchmarks to measure bias and detoxification programs. This chart shows the results of running the language model GPT-2 through three different detox methods. Cr: Stanford Institute for Human-Centered Artificial Intelligence

A number of research groups, including Open-AI, are working on the toxic-language problem, with both new benchmarks to measure bias and detoxification programs.

Related: The report shows that algorithmic fairness and bias has shifted from being primarily an academic pursuit to becoming “firmly embedded as a mainstream research topic with wide-ranging implications.” Researchers with industry affiliations contributed 71% more publications year over year at ethics-focused conferences, the report found, though the “wide-ranging” implications are not defined.

AI Becomes More Affordable and Higher Performing

Since 2018, the cost to train an image classification system has decreased by 63.6%, while training times have improved by 94.4%. The trend of lower training cost but faster training time appears across other machine learning task categories such as recommendation, object detection and language processing, and favors the more widespread commercial adoption of AI technologies.

A Plateau in Computer Vision?

The AI Index shows that computer vision systems are tremendously good at tasks involving static images such as object classification and facial recognition, and they’re getting better at video tasks such as classifying activities.

But there are limits: As spotted by IEEE’s analysis, computer vision systems are great at identifying things, but not so great at reasoning about what they see.

The report notes that performance improvements have become increasingly marginal in recent years, “suggesting that new techniques may need to be invented to significantly improve performance.”

AI Needs Women and People of Ethnicity

AI research is dominated by men. The report finds that the percentage of new AI and Computer Science PhDs that are female has moved only a few points over the last decade, at least in North America. Not much has changed since 2021. The data for AI and CS PhDs involving people of color tells the same story.

“The field of AI needs to do better with diversity starting long before people get to Ph.D. programs,” IEEE urges.