AI aiding water infrastructure

The models that are increasingly being deployed to help water engineers keep the taps running. But, as Babak Zolghadr-Asli explains to Chris Smith, up till now, their design has largely focused on the quality of the model’s architecture rather than the quality of the data it’s processing. And, as with any AI model, when it comes to the quality of the output, the old adage “rubbish in, rubbish out” has never been more important. So making sure we have high quality datasets - and data collection methods - is essential to ensure the results and predictions are reliable…

Babak - It is far too easy to misuse these models or misinterpret their results. And imagine a case that you are embedding these models into some of your most vulnerable but most important infrastructures, like your water systems. If these models start producing results that are not necessarily reflective of the real world, then you have to face the consequences.Then you would face problems.

Chris - Is it the model itself, as in the way computationally it's working, that's wrong? Or is it the underpinning data, the information we are training it to use in engineering its own thought processes, that's at fault? Or is it both?

Babak - Very good question. There is this way of thinking that the algorithm is the most important thing. We are putting this in the centre and front and assuming every breakthrough is going to come from improving these models, becoming more efficient with how we are going to do about the computations and so on and so forth. However, there are those who believe that data is also important and should be at the focal point, because the model is only as good as the data that we provide it. So, the idea here in this specific case of water engineering being is that often at times we don't have enough reliable data to train these models. So, essentially they haven't seen the world enough so that they can replicate it in a good way.

Chris - Why is this not already something that people have confronted and dealt with when engineering these models?

Babak - I usually give this example.Imagine that you are a teacher and you want to teach your students. The challenge here is that the only exposure to the real work, your only material to teach this student is that you have a window to the real work and you're asking them to start drawing their real work through this window. Now imagine this window being really small.
Because of this limitation, the student is not being able to pick up every talent during this training session, right? Why haven't we seen any problem? The issue is that usually these models, for natural phenomena, it would take decades to see every pattern.

So for instance, we have to monitor and observe the performance of these models in these long decades to see if they are reliable or not. But we can't obviously wait this long because at the point that we realise we made a mistake, it could be too late. But the signs and the red flags are there because every now and then you would dig deeper in the performance of these models.
You have experts to go into how these models are trained. And if you go deeper, you would realise that the logic inside these models sometimes are not on par with the reality of the work.

Chris - What do you regard as the solution then? Is it just making these models more explainable? Is it having a sooner, more rapid feedback loop so that we can work out if they're going off kilter? Or is it just about enhancing the data that we feed into them so we don't settle for that very small window on the world that you use as a beautiful analogy, we go for the widest vista we can?

Babak - The issue is data. So, first things first, I think we have to manage our expectations and start seeing the problem for what it is. We shouldn't expect these problems to be fixed overnight as well. It would take years if not decades. But this sort of highlights the importance of starting investing in this. Because even if you start to collect more data, as we speak, it would take decades, if not longer to build up enough data sets to see tangible, meaningful changes in the performance of these models.

Convincing both public and private sector to invest in this is more challenging than what you would think, because the return on investment is not necessarily that immediate, or sometimes that apparent. Say in some context, if you start collecting data, you can sell the data either directly or indirectly and get some feedback or financial compensation. But in this case, it's not going to reward you in a financial way.

But if you're looking at this problem as investing in our strategic resources, I think it is far too easier to justify these costs. At the end of the day, I think the best way I can describe it and our relationship with these models is that AI and computational intelligence should be seen as a tool that we need to supervise, not the other way around.