Machine-learning models more powerful, toxic than ever
US-China research collaborations and startup investments on the up, too
AI systems are becoming increasingly larger and complex, but despite the technology's progress, they exhibit higher levels of toxic behaviors, according to the latest AI Index Report.
The 2022 report, the development of which was led by Stanford University's Institute for Human-Centered AI, analyses how machine learning affects research and development, economies, and policy-making across the world.
One trend highlighted by the study, now in its fifth year, that will probably remain in vogue for the foreseeable future, is that neural networks are getting bigger. There are now language models with hundreds of billions and even trillions of parameters trained on terabytes of text scraped from the internet.
These massive systems can complete all sorts of tasks, from generating content to helping developers code, to a degree. They have become more powerful and potentially more harmful.
"A 280 billion parameter model developed in 2021 shows a 29 percent increase in elicited toxicity over a 117 million parameter model considered the state of the art as of 2018 – this increase in toxicity is accompanied by a broad and significant increase in capabilities," the report stated.
It is now more important than ever to understand the shortcomings of these systems as they're commercialized and deployed in the real world, the authors warned.
The current boom in deep learning is often traced back to 2012 when academics at the University of Toronto won an image-classification competition using a convolutional neural network [PDF].
Architectures in computer vision and other areas in AI have evolved since, and training and running a model is becoming more affordable. In fact, the cost of training an image classification system has decreased by 63.6 percent, while training times for AI systems have improved by 94.4 percent, we're told.
As these costs come down, the technology can be adopted on a wider scale, and there's demand for it. The total investment into private AI companies in 2021 around the world was $93.5bn, more than double the amount compared to 2020. Interestingly, the number of newly funded startups, however, has continued to drop. There were 1,051 companies in 2019, 762 companies in 2020, and 746 companies in 2021.
Ray Perrault, co-chair for the AI Index's steering committee, told The Register the decline "could be a interpreted as a sign of the growing maturity of the AI sector, with larger investments being required as products containing AI are being delivered on a larger scale."
"The investment need not be for development of AI algorithms. It could also cover large-scale data collection, cost of training machine learning models, and investment in hardware designed to be integrated with AI software, as would be the case with self-driving cars. However, there could be underlying challenges in collecting this data—as AI has become more prominent, a broader set of companies may call themselves 'AI companies'," he said.
- Research finds data poisoning can't defeat facial recognition
- Cerebras brings wafer-size AI chips to medical data analysis
- Ukraine uses Clearview AI facial-recognition technology
- Microsoft, OpenAI method could make training large neural networks cheaper
Commercialization of AI is heavily driven by state-of-the-art research. China is the world leader in this area, at least in terms of number of papers published, with the highest number of academic journal publications, leading the US, in second place, by 63.2 percent.
There are, as you can imagine, all sorts of reasons why paper counts aren't a direct measure of actual research being done, though it's an interesting stat, nonetheless.
The geopolitical beef between China and the United States right now doesn't seem to have had a negative impact on R&D. Cross-country collaborations between both countries in AI is at an all-time high, increasing five times since 2021, we're told.
"Researchers want to work with the best scholars in their fields, and as a major source of AI talent, naturally there will be a large number of Chinese researchers that American researchers will want to collaborate with," Michael Sellitto, deputy director of Stanford HAI, told The Register.
"Additionally, US companies such as AWS and Microsoft have research teams in China, as do Chinese firms such as Baidu in the US, which may contribute to the number of binationally-authored papers. It is also worth noting that, geopolitics and 'arms race' fears aside, most academic research is intended to be published openly - and even AI researchers working in industry demand the right to publish, as it is important to their careers.
"So, in large part, knowledge produced would be made available to Chinese researchers, even if there were no US-China collaborations involved in a particular project."
"Many may also consider academic engagement as an opportunity to increase mutual understanding and decrease tensions, much as the United States and Soviet Union promoted academic and cultural exchanges during the Cold War. That said, anecdotally, government-to-government tensions and the Department of Justice China Initiative does seem to have cast a pall over collaborations involving Chinese scholars and institutions for some researchers, particularly those of Chinese descent," he added.
Another sign that machine learning research is maturing is the decreasing costs of mechanical arms in robotics research. According to the AI Index, the median price of robotic arms decreased by 46.2 percent in the past five years — from $42,000 per arm in 2017 to $22,600 in 2021. "Robotics research has become more accessible and affordable," we're told.
Finally, as AI is practically deployed around the world, countries are beginning to regulate the technology. The AI Index report revealed the number of bills containing the words "artificial intelligence" passed as legislation grew from just one case in 2016 to 18 in 2021, with the US, UK, and Spain approving most of them.
"The concern with ethical issues in AI systems is growing, as demonstrated in our report. Exploration of where the issues of concern are to be found is relatively recent, but we are not yet at the stage where the identification of the source of the concerns has led to the development of robust methods for avoiding them in the future. It is too early to say that these issues will not be fixed," Perrault concluded. ®