Early AI systems tended to be really good at only one thing. They might, for example, be terrific at classification problems, such as distinguishing a cat from a dog or a West Highland Terrier from a Great Dane. Or they might be outstandingly good at translating English to Hindi.
Those were the days of long ago (i.e., a very few years). Today, leading AI systems are multimodal and multifunctional. Multimodal AI systems cleverly fuse input data from a variety of nominally disparate sources. Such sources might include the obvious: text or images or sound. Or more esoteric data such as radar or lidar sensor outputs or the results of medical imaging. Often all of these sources are converted into tokens or ‘feature vectors’ which can be collectively ingested to train a large foundational AI model based on a transformer or similar architecture.
In contrast, multifunctional AI systems fuse more than one AI skill together to deliver the required system behavior. Well known examples include self-driving cars or vehicles or articulated robots. A robot must, for example, master real-time control of the dynamics of its limbs and actuators by using AI techniques such as reinforcement learning or similar methods. It may further apply vision processing to detect and classify the objects in its surrounding area by using convolutional neural networks (CNNs), and it may use AI-based simultaneous location and mapping (SLAM) methods to navigate around them.
But what happens when AI methods are combined not within an individual machine but rather to a specific industry or vertically integrated supply chain or production system? Do the best AI solutions for that industry then also need to be ‘vertically integrated’? And what does this mean for their multimodal and multifunctional capabilities?
This question has been occupying some of SineWave’s attention recently. As pragmatists who see investment value in real-world production companies, one of the domains where we have addressed this question is the world of industrial systems. Consider the following three examples of multimodal and multifunctional industrial AI systems that we have recently assessed.
- Industrial Safety – Safety is a critical consideration in the operation of large industrial plants such as oil refineries or steel-making facilities. An AI safety system operating in such a plant might use convolutional neural networks to assess thermal images of reactor vessels to identify hot spots or other points of failure. Or it might apply an AI transformer model to assess the acoustic signature of fluid flowing within a pipeline transporting a reactant under pressure to predict stress fractures or structural failure within the pipeline. Or the system might use an AI graph neural network to evaluate the pose and physical position of workers within the plant. Are they, for example, endangering themselves and others by not using the required safety equipment or behaving irresponsibly when working at height? All of these methods can be fused together into a single safety-enhancing AI solution that is expert about, say, mining operations.
- Industrial Quality Optimization – Production quality is a perennial goal of industrial plant managers. Returned products and recalls are expensive, reputationally damaging and occasionally even dangerous. Today, however, a knowledgeable and adventurous plant manager can use a vertically integrated AI system to optimize production quality. Such a system can handle a variety of inter-related tasks, ranging from the use of a convolutional neural network to assess and classify images ultrasound images of a weld, to an AI transformer model used to assess worker performance in parts placement within a moving production system such as an engine or chassis assembly line. And just as quality optimization is often cast within a company’s total production philosophy (see e.g., the Toyota Production System) the resulting AI subsystems optimize quality best when fused together.
- Industrial Production Management – Many modern production systems rely on complex, multi-stage reactions to generate products ranging from biologicals such as anti-viral vaccines and drugs to advanced organic and inorganic pharmacological products. The cost of the reagents and processing means this must be done at the absolute maximum of process efficiency and yield, but the limitations of contemporary plant management systems means that production control of reagent ratios, temperature, agitation level etc. is suboptimum. Here, a vertically integrated AI system can plan production using AI-based time-series prediction methods and can use related and closely co-designed neural networks to manage the real-time behavior of each stage of the production process pipeline to optimize yield.
These are just a few examples of where we at SineWave have been seeking value in ‘vertically integrated AI’ recently; there are many others. When fused together into a single coherent AI system, and when trained by access to authentic plant operational data, such solutions are broadly expert and can offer huge value in human safety, reduced plant downtime and operational efficiency. We believe that there is real and present value in companies that can both master the different AI techniques needed to build vertical AI applications and who also have the necessary skills to partner with industrial companies who can provide the critical training data needed to train such systems. The age of vertically integrated, industry-specific AI is indeed upon us.