AI has existed for several decades and has shown increasing levels of maturity from time to time. It has also shown the promise and aptness to solve the problems of that day. There were growing investments by companies to adopt these technologies to gain a competitive edge. But it soon ended up with massive disappointment and loss of trust in not being able to keep up with the expectations. These couple of periods in history when AI slumped into darkness were termed THE AI WINTER. Now we are again in the era of similar hype with companies seeing a renewed promise and interest in Deep Learning (DL) technologies.

But who is to be blamed for those winters?

The SOTA progresses through 4 different groups to reach the end-users; Academicians, Engineers, Executives, and Customers.

Academicians and Researchers are the ones who come up with SOTA algorithms and techniques. Academics demand students to publish improvements over existing methods to fulfill their course completion requirements. Industries also resort to publications and patents to counter other company patents/licenses. Though every publication would have an incremental improvement, very few of them will actually be groundbreaking, a stepping stone, and a benchmark for the future. Even though this group is well aware of the shortcomings of the SOTA techniques, it is not well understood by the rest of the groups. There are two reasons for this. Anyone who comes up with a breakthrough wants the technology to be adopted and cited, so why would the shortcomings be highlighted? And even if one decides to highlight it, it would not be worth explaining to any of the other groups because academicians would already know of the drawbacks and the other groups would not comprehend it well. Shortcomings are like backlogs that researchers are trying to address. When they publish a paper, just one of the many backlogs is what will be addressed and also highlighted. It does not make sense to keep repeating all the other backlogs in every paper published. Backlogs might also change over time.

Engineers as they are the ones who create solutions or build systems from these algorithms. It is the responsibility of an engineer to apply the right kind of algorithm to the right problem. They should not try to punch above the algorithm's weight and undermine it. Engineers first create demos, POCs, or POVs which would be a demonstration of the algorithm on a limited scope of the problem in controlled or manipulatable environments. Due to this reason, the results will generally be very impressive and also achieved in a short span of time. Even though this group is well aware of the challenges that lie ahead in building an industrial system out of it, the next-in-line group is so overwhelmed by the initial results that they generally don’t heed to any narratives about it. Improving the system beyond a “certain point” generally follows the law of diminishing returns wrt the time spent on it. This “certain point” depends on the potential of the algorithm and this potential was different before every AI winter.

The baton then goes to the Executives. The executive group makes sure the initial results are impressive because that is what is to be presented to the end user/customer. The executives pretend these results to be close to the end product though from their experience they know it is generally far from it. They have to paint this picture either because the competitors do something similar or they are desperate to meet the targets set by their company. If they don’t do so, their career is at stake. The executives take this risk while pitching in the solution because they assume they can buy time later to polish it further. They also bet on the probabilistic nature of such deals going through. If the executives commit beyond the potential of the SOTA, they are in for a treat from the end user.

Customers are the end users or consumers of any technology. Customers always have high expectations in terms of accuracy, performance, and value for money. They are the most demanding of all and care the least about the underlying technology and most about whether it serves their purpose. Customers can come up with crazy use cases which may or may not be solvable with the existing SOTA. But since this group knows technology the least, they leave it to the Executives to decide on it. The executives bank on Engineers and Engineers on SOTA research. So the technology movement from Research to User is not necessarily a one-way street.

Engineers play a significant role in this ecosystem. They are the ones who make or break the technology. Winters start to commence when Engineers come up with solutions by blindly fitting algorithms to every problem they encounter or Executives agree to provide a solution to any and every requirement from the Customer.

DL is no doubt the most sophisticated learning algorithm to date. Reduced cost of storage, bandwidth, and compute combined with widespread digitization fuelled its growth beyond expectations. But like the greed for money, time, etc, expectation is not an absolute quantity. We always expect more than what is available to us. But there is always a limit to the scope and applicability of an algorithm and should not be taken for granted as an ultimate one size fits all solution. If we continue to do so, we are bound to hit another AI winter.

It is my own article