Robotics, analytics, data science, machine learning and AI

“AI” is all the rage these days; no proposal to do something clever with information technology is going to get funding unless it has the magic words “Artificial Intelligence” in it – preferably oft-repeated. Conversely, if those sacred words are present and repeated in a mantra-like fashion, the level of investment scrutiny goes down a little, and the eagerness to put money in rises sharply. This is, of course, the recipe for, if not a full-blown economic bubble, at least a bit of boom-and-bust as expectations are (over) inflated and then rapidly deflated. Governments should not be promoting boom-and-bust – some even declared an end to it a while ago. And they should not plough public funds into large-scale, high-risk investment, development projects based on vague hopes of stimulating the economy to slightly higher growth using some “silver bullet” technology.

So, it is worthwhile to take a practical, down-to-earth look at AI to examine where it might be useful and add value rather than just swallow it up and put aside the hyperbolic rhetoric about radical change in the economy and workplace.

Robotics and data processing

There are two strands to this topic: one is robotics and the other data processing.

Robotics

Robotics, or automation, is using machines to make changes in the world; to do the things that humans used to do. For example, technically you don’t need builders to build a house. A house can be built in hours by a hybrid machine, something like a cross between a 3D printer and a crane. Of course, in practice such machines are very expensive, but, in principle, they can build houses faster and better and cheaper than any group of humans. In practice, therefore, you end up with one or two people using the machine to build 4 houses in, at most, a few days.

This is what automation does – it allows a small number of people to use machines to do the work of a large number of people, and so a significant increase in productivity (work done per person). Behind the scenes, of course, there was a massive investment of time, money and human intelligence to develop the machines in a sort of “10 years of practice went into that 2 minutes of performance” sense. [Assuming the investors had the vision to invest in the machine and not the traditional house-builders.]

Automation can also be used for improving security by securing supply chains, reducing errors and minimising cyber attacks. This sort of automation does not require any huge amount of “artificial intelligence” – most of it is ‘programmed’, ‘mechanical’, ‘robotic’ (sic). The machine just needs to be equipped with the sensors and rules to stop it from doing the stupid things – like trying to put the floor on top of the roof, or building the walls over the drain covers, or driving over people.

Data processing

That brings us nicely onto the data processing strand. It is well-known that data is not information; for it to become information, it has to be analysed and interpreted, by which meaning or semantics are attached to the processed data. Several decades ago such data analysis was carried out by people using tools like spreadsheets – and some regressive companies still do this sort of thing in a regressive reverse-automation (getting people to do things the machines can do better, faster, cheaper) – but most organisations realised this was just a programmatic task – and so automated it with software robots operating across large-scale stores of data.

Thus, “analytics” was born, and the people who had previously been analysing the data moved up the stack and created the software machines (data analysis programs) to do it for them. In doing so, this group also raised their productivity just like the house-builders. But like the house building machine, the data analysis machine doesn’t need much (if any) intelligence built into it – it is just following the pre-programmed rules mechanically; all the intelligence stays with the programmers.

Data science and AI

“Data science and AI” takes a significant step up on both the intelligence and data processing ladders. It is still founded in analytics, but it differs in that it brings to bear a model of the world to which the data pertains and tries to use this model to generate more incisive and insightful interpretations of the data, attaching richer and more fruitful semantics or meanings to the data. Hence, you get:

• “Descriptive analyses” – semantic information seeking to describe what is going on in the world.

• “Diagnostic analyses” – semantic information trying to find out what’s going wrong (or right!).

• “Predictive analyses” – semantic information that tries to tell you how the world will be.

• “Prescriptive analyses” – semantic information that tries to tell you how you should respond to things happening in the world.

Data scientists are very clever folk who are adept at creating little software robots capable of extracting real, meaningful information out of your aggregated data sets. This is a step up in applied intelligence, but not in the software doing the data processing, which is pretty much as dumb as ever it was. The intelligence still resides in the data scientists, although a bit of it, and their knowledge of the world, may be embedded in the data science models, but not a lot of ‘intelligence’ in the data processing software itself.

The fundamental issue with data science with respect to artificial intelligence is that the ‘understanding’ of the world is not encoded into software; the model is ‘external’ and ‘static’. Machine Learning changes that, and this is the start of “artificial intelligence”.

AI and pattern recognition

The classic example is pattern recognition in the form of breast cancer diagnosis based on imaging data. [See this study on the “Use of artificial intelligence for mammographic image analysis in breast cancer screening”]. To begin with, lots of data are collected in the form of images and the subsequent (human) diagnoses of breast cancer. This database, generated by the application of human intelligence, which itself has to be trained using traditional ‘social’ means, is used to train a pattern recognition ‘network’ (that simulates a human neural network) so that eventually the learning algorithm learns to recognise, in images, the detailed features characteristic of a positive breast cancer indication even if those characteristics are below the ‘cognitive threshold’ of the humans that contributed to the database.

Once trained, the algorithm has ‘learned’ and can be deployed to analyse new images for the positive diagnosis patterns and, reflexively, add the positive ones to the database for better training, thus closing the learning loop. If the learning loop is effective enough, it can even slightly modify the model to be more effective, even though it has no ‘understanding’ at all of the concepts represented in the model. Nevertheless, this is a (very) limited form of intelligence based on learning through interaction with (data representing) the world.

The ‘top-of-the-tree’ is real (sic) “artificial intelligence” (AI), which is not “artificial general intelligence” (AGI). [It is “artificial general intelligence, which is potentially an existential threat to humanity, but even the most advanced AI today is decades away from even a basic AGI capability.] However, it does have a more developed and dynamic model of the world or, at least, its domain of application internal to its operations and modifiable in the light of ‘information’ extracted from its large volumes of data.

With this level of capability, the AI-enabled system can recognise a wider range of situations, from even low-level data and make more logical, probabilistic inferences and projections. But just to be clear, the AI has no real understanding of the concepts represented in the model, and though it may change the model based on the results of its data processing, those changes are very incremental and it cannot invent new concepts or ‘think’ (it’s not really thinking) differently from the thousands of humans whose thoughts are captured in the data. [So even the most advanced AIs available today are not that intelligent, judged according to human criteria for intelligence.] There is nothing magical about AI any more than there is about human or animal intelligence, despite the magical thinking built into Cartesian philosophies of Mind.

Robotics, AI and defence

Oddly enough, this also brings us back to the issue of robotics. One very valid application of AI, in defence, is to use it to enable “command-and-control” of ‘swarms’ of uncrewed vehicles to act as a “force multiplier” to enable a smaller number of soldiers or sailors to do the work of many, just like the house-builders we started with. However, these robotic and autonomous systems, to use the jargon, need to be somewhat smarter in their “situational awareness” and “course of action” selection than a house-building robot. There's no use trying to use a small swarm of Uncrewed Underwater Vehicles to track a foreign submarine if they go wandering off to follow a whale or an active sonar decoy at the first opportunity. They need a developed form of pattern recognition and AI to recognise and avoid deception and to adapt their behaviours appropriately. In short, they must be able to accommodate Clausewitz’s perpetual fog of war.

The consequence of these factors and considerations is that AIs need to be deployed at the edge where they can sense change in the world, analyse it, make decisions and make changes in the world using whatever ‘effector’ sub-systems they have at their disposal. The AIs may well not be alone in this; they may be ‘partnered’ (in an unequal partnership) with human intelligences in “Intelligent Complex Adaptive Systems-of-Systems” (ICASOSs), or systems of sociotechnical systems, to use the modern systems engineering jargon. But systems affecting change in the world are necessarily at the cloud edge and not in the core, where the big data centres, data stores and large data-sets necessary for AI training are currently found.

Therefore, we need a novel mode of AI utilisation; one where the training of an AI using very large data-sets in the Core is separated from the deployment of an AI to the Edge to carry out Operational functions, including assisting in rapid decision-making.

Infrastructure-as-Code

The model of “Infrastructure-as-Code” serves here: AIs are shipped out to their Base-of-Operations, wherever it may be in the world, just like military human intelligences are; except that they are shipped in minutes over the networks that form the communications infrastructure of the cloud (and any military operation these days – networks of operational capability are formed over and above communications networks).

Final thoughts on data science, robotics and AI

Discussions at Defence AI Centre (DAIC) this week highlighted a critical truth; AI is only as valuable as its practical application. While the potential for AI in defence is vast, its real impact depends on how well it integrates with existing systems, enhances decision-making, and operates reliably in complex environments. The conference reinforced that the focus must shift from AI as a buzzword to AI as a tool that augments human capability, strengthens operational resilience, and is deployed responsibly at the edge. As defence continues to navigate the challenges of AI adoption, the priority must remain on developing adaptive, accountable, and effective systems, rather than chasing technological hype.