Dimitris Poulopoulos
There has been a lot of hype around artificial intelligence (AI) and its subfields (machine learning, deep learning, etc.) for quite some time now. Yet, we may be on the verge of another AI winter — a period of reduced funding and interest in artificial intelligence research — despite the significant advances in algorithms and infrastructure, even with the vast amount of information and data that we have at our disposal.
We may be on the verge of another AI winter, despite the great advances in algorithms and infrastructure, even with the vast amount of information and data that we have at our disposal.
However, AI itself is not accountable for this letdown. This time, technology is entering the trough of disillusionment because we do not know how to benefit from it. We do not have a robust methodology for transforming breakthrough research ideas into practical applications yet.
Yes, many algorithms have “graduated”, and consumers are using them without even noticing it, but still, putting a model to production remains arduous. Coming from an academic background, the problem is clear to me: it is time to start developing solutions, not only prototypes.
The Red Flag
The coronavirus outbreak caught us off guard. Technology, and particularly the field of AI, fell short of the expectations of investors and policymakers, not because we do not invest enough but because we do not invest effectively. Don’t just fund research; fund solutions says Ilan Gur in his MIT Tech Review article “How the US Lost its Way On Innovation”.
Indeed, although we could use AI to help us prevent, diagnose and maybe treat a disease like COVID-19, almost two years after the outbreak in China we only talk about ideas and develop prototypes of what solutions might look like.
Moreover, people that seek to capitalize on the public’s panic found startups with questionable methodologies and products. Hundreds of AI solutions are proposed for the pandemic, but few are proven, warns Elise Reuter in her article for MedCity News.
We are about to lose this big opportunity to showcase the power and worth of AI and that would not only be a shame but could make investors more skeptical in the future.
We should undoubtedly be extra careful when it comes to machine learning in the health sector; however, we are not closer to deploying an assistive machine learning solution now than we were one year ago. We are about to lose this great opportunity to showcase the power and worth of AI, and that would not only be a shame but could make investors more skeptical in the future.
The Open Issues
Now that we have the problem defined and have witnessed our reflexes respond poorly during the pandemic let us see what we could do about it.
Focus on the future
Our research agenda is stuck in research priorities and approaches that were meaningful during the past decade. The world moves at an unprecedented pace. For example, little has been done for problems like climate change. What should we do with nuclear waste? How to make sustainable forms of energy meet today’s demands? How should we prepare for the coming storm, mitigate its impact and adapt our economy?
At the same time, we should research methodologies that would allow us to bridge the gap between research and the industry faster, find the sweet spot in the intersection of AI and infrastructure to turn ideas into practical applications faster.
There are projects addressing these issues today (e.g., Kubeflow or AWS Sagemaker). Still, we should move more quickly, bring more voices to the conversation and focus our attention on such endeavors.
These are problems that we cannot solve by slightly adjusting the research budgets in a spreadsheet. We should delete the file and start from scratch. We should change our mindset and design a new course of action.
Offer different incentives
Breakthroughs in the field of Artificial Intelligence and Machine Learning mostly come from researchers whose careers depend on the number of publications they author and conference presentations they make. Usually, to pass a paper, researchers try to better the result of some benchmark by just a few decimal points.
This way of conducting research rarely drives innovation. If we expect meaningful and applicable results, we should not care that much if the new Imagenet state-of-the-art model surpasses the old one by 0.5 percent. Don’t get me wrong that may be an impressive result and require a lot of thinking. But we should also consider how to bring it into production. How to directly leverage a new good idea. An example of this would be OpenAI’s new API, for which you can read more in the story below.
In the end, if we want different and more tangible outcomes, we should offer different incentives.
Change the research playbook
I would argue that startups and University spin-offs are one of the leading forces that drive innovation and offer practical solutions in AI today. Thus, maybe it is time to start funding those ventures as well, not only as part of a research framework but also as individual entities as well.
This way, we could bring people from the industry, who know how to implement technology, closer to researchers. In a nutshell, we should get over our aversion to funding industry research.
Conclusion
In this story, I argued that a new AI winter is coming and, this time, it has nothing to do with AI. The standards are there, Machine Learning and Deep Learning have delivered great ideas, that we can put into production today.
To achieve that, we need to make some changes: focus on the future, offer different incentives to researchers, and change the funding playbook. We shouldn’t let AI hit another dead-end.
No comments:
Post a Comment