The AI Gaps
I attended NeurIPS 2018 (Thirty-second Conference on Neural Information Processing Systems) in Montreal, Canada last week. NeurIPS, formally called NIPS, is one of the most important and popular conferences in machine learning and computational neuroscience held every December. It includes invited talks as well as oral and poster presentations of refereed papers, followed by parallel-track workshops. NeurIPS was the biggest AI conference this year with more than 8,000 registered participants. The first batch of 2,000 tickets sold out in less than 12 minutes!
As an AI researcher, I enjoyed the conference a lot. I had a chance to meet many of my peers as well as the pioneers in this filed. There were many interesting talks, papers presented orally and papers presented through posters. It was a great learning experience for me. But, as a practitioner and someone who has worked in the IT industry for a long period of time, I observed two main issues/gaps in the AI field. I will try to explain my observation below hoping for a positive impact.
Applied vs. Pure Research
A lot of great papers were submitted to the main conference this year. It was exciting to see so many researchers are working on new ideas. I heard the organizers tried to accept many applied papers this year too, but didn’t feel enough to me. Surely, we shouldn’t limit research to only current applications but they shouldn’t be ignored either. There seems to be a big gap between applied and pure research in this field. Unfortunately, there is a strong competition among AI researchers to publish as many papers as possible. As a result, majority of the papers are making only a little contribution and experiments are done on a limited set of datasets or synthetic data. Those who explore how these algorithms will generalize in real life scenarios, are minorities. As an applied researcher, I understand how challenging it is to get good results on real data and why many researchers prefer otherwise.
The gap between applied and pure research was even more evident when I visited the NeurIPS exhibition where many companies attended to attract talents. In addition to the usual exhibitors like Google, Facebook, Amazon, NVIDIA, Alibaba, there were hedge funds, investment companies, banks, a few consulting companies, a couple of incubators like NextAI, and only a couple of giant e-commerce companies. Talking to a few non-tech companies there, I realized most of them don’t have a clear vision how to take advantage of AI (some don’t even understand the difference between AI and data analytics). Looked like they are building AI teams just to be ahead of the competition not with a real purpose. However, what surprised me the most was the absence of other type of businesses that can take advantage of AI to optimize their operations. Companies like airlines, transportation companies, manufacturers, retail companies, government agencies, and many more.
The demo sections of NeurIPS were even more depressing. These are the companies or startups that used AI for their applications. Most of the works I saw were very primitive. Many of them didn’t even need the type of AI solutions currently explored in the research community. Also, the week before NeurIPS, I was in Google’s SOCML 2018 (Self Organizing Machine Learning …) and attended a few of the discussion groups including “AI in Supply Chain” and “AI in non-tech Companies”. Only handful of people attended these two groups comparing to big crowds that were attending very technical groups like GAN (Generative Adversarial Networks). All of the above were confirming that there is a big gap between applied and pure research in AI.
One thing that makes me worried is that many of the businesses are just re-branding their traditional data analytics or BI as AI while they are missing the real benefits a true AI strategy can bring to them. I know one of the big retail chains in Canada that consider them a leader in AI but all they do is data analytics. This topic by itself deserves another blog post.
Poor vs. Rich
The nature of this gap is very different from the previous one, but I thought it deserves to be mentioned too.
One advantage of working in IT was that you didn’t need a lot of resources to implement your ideas by developing a software, mobile app or a website. But that’s not the case in the era of Big Data and Deep Learning. Teaching deep models, in a reasonable time, requires a lot of computing power especially expensive GPUs. Although, designing a good model requires the knowledge and the experience of the AI expert, but many times good models are discovered by trying different architectures and hyper parameters. If you don’t have access to large computer clusters or GPU arrays, then it won’t be feasible to achieve a lot. Even some universities in the developed countries cannot afford to have enough resources, never mind of those in developing countries.
The fact is that many of the good works in this field is done by big companies like Google, Facebook, Amazon, and so on that have access to both data and computing resources so much that they are even ahead of universities. The good news is that most of them are open sourcing their works for everyone to benefit. However, Not everyone in this field can work for those companies and consequently, we are not giving the chance to everyone to contribute and to bring new perspectives. Starting an AI business is also not easy and requires considerable funding from the beginning. This is definitely a problem that requires attention.
What is the Solution?
We are in the early phases of AI revolution and I don’t think anybody has a full understanding of the problems and more importantly the solutions to those problems. However, I could think of the following few things to start the conversation:
- I think we, the AI community, have the obligation to educate public, policy makers, business leaders and NGOs to understand the potential of AI and how it can be applied in their organizations to increase efficiency. This is specially the task of those who have experienced both sides (industry and research).
- Universities and governments should pay more attention to applied AI. More cooperation between industry and universities is required. The existence of entities like Vector Institute in Toronto is a good step toward that goal. I was also happy to hear that government of Canada announced that it will invest 230$ M in Scale.AI, that is Canada’s artificial intelligence (AI) Supercluster dedicated to building the next-generation supply chain and boosting industry performance by leveraging AI technologies. But I think we need to see more of these good initiatives.
- We need more applied and industry specific conferences, competitions and workshops. For my part, I have already started discussions with the potential people to create a momentum. But this requires more collaboration among the community and the support of regional or federal governments.
- Governments need to invest more in computational resources for local universities and colleges allowing them to get ahead of the private companies.
- Multi-national companies and organizations and leading universities should collaborate with their counterparts in developing countries to give more researchers opportunity to be active in this field.
