“Machines are the single most impactful indicators of how human evolution has happened,” said Malavika Peedinti, Data Science Manager at BLP Industry.AI, in her session “Machines Talk, Machines Learn” at The Rising 2020. In this session, Malavika spoke about two kinds of machines — one that is physical and talking, in terms of the data that they churn out, and the other type of machines that are learning patterns.
To start with, Malavika explained the four industrial revolutions that we have gone through over decades — steam-based, electricity-based, computer-based and intelligence-based that “we are currently on,” said Malavika. She further noted that the previous industrial revolutions were hardware-based, whereas the fourth industrial revolution is intelligence-based. “It’s about how you understand the machines and how intelligently are these machines programmed to make their own decisions.”
Problems Of Implementing AI
Malavika further stated that AI is not pixie dust, and many problems arise while actually implementing AI in the core business. “It’s not something that can be sprinkled on a factory, and then we start getting amazing output or productivity from the next day.”
Instead, the problem of implementing AI for industries is twofold.
“The first problem is the way industries look at AI,” said Malavika. Explaining this she said, it gets challenging to make these experienced technicians, who have been working with machines for more than three decades, understand about AI, and how this new technology can augment their understanding of these machines.
Also Read: How COVID Pandemic Highlighted The Limitations Of AI
“The second set of problems are the ones that data scientists face in their real-life work.” Explaining this Malavika said that it’s usually a certain kind of problem like classification and regression that data scientists are trained on. However, in real life when AI is being applied in industries the issues are much different “like optimisation, reduction in quality, and increase in overall production.”
So, according to Malvika, “the kind of terminologies and the kind of problems that are looked for, to solve with AI in the industry are very different. Also, the class of problems is different from the kind of problems that data scientists are usually trained on.”
In this section, Malavika explained a few examples of machines that were termed as “talking machines.” These included wind turbines, seaport crane, solar panel and chillers of a building. Here, Malavika explained that these machines are termed as talking machines, because these machines “give out information continuously at every 10 seconds,” stating different kinds of parameters like temperature, vibration, speed of rotation, etc. And it depends on humans to understand these data and make sense out of it.
To understand the data, and what parameters from that data are essential, data scientists need to go through a three-step process — understand the process flow diagrams, to understand the technical specifications, and to understand the data dictionaries.
“It is basically how the different components of the machine are interacting with each other along with its technical specification of each component individually, and then data dictionaries, which is how the data is stored with any multiplication factor,” said Malavika.
“So, till now, we saw how individual machines are giving out data like temperature and vibration, however when we talk about machine conversations, then things change. This is where the entire change with respect to industrial revolution 4.0 takes place,” said Malavika.
Malavika explained how machines converse with each other. Case in point, a wind farm, which has multiple wind turbines needs to optimise each turbine in order to maximise the output and increase the farm’s productivity. And in these business cases, AI models are the preferred choice for accurate optimisation. “What we need to do is write AI modules that enable these conversations between different machines.”
In another case, Malavika explained how the AI model had been used to optimise a seaport operation and ship unloading process. “So the optimisation problem was to streamline the process of ships’ unloading, and all this was possible through a combination of different kinds of machine learning techniques.”
These machine learning techniques were termed as “learning machines.” Once the talking machines present the data, it’s the work of these learning machines with the help of algorithms to come out with the right kind of predictions and classification for the users.
Challenges Of AI Deployment
Here Malavika explained two most significant challenges of deploying AI into your business — explainable AI and delayed PoC to production.
According to Malvika, making AI explainable is difficult for businesses, and that’s where reliability score comes into the picture. “Not all of the algorithms are so easily explainable, and not all of the results. And that’s why we give out the reliability score. We mention that the machine is reliable for a certain duration, and after that, maintenance would be required.”
The second challenge is the transition of the code from proof of concept to production. According to Malvika, not many projects that are successful at the PoC stage get to reach the production stage because of reasons like longer model development time, insufficient data etc. According to Malavika, the real success of AI will be when more projects are going to the production stage.
Thus, these issues make AI deployment challengeable for businesses in the current era.
To conclude, Malavika said that when an AI model is made explainable and is represented well with, predictive maintenance and data visualisation, the success rate is more. “If these are a few things that are taken care of, then definitely the fourth industrial revolution would be so much more successful.”