Against the backdrop of the heightened tensions that have characterised the geopolitical landscape of the last few years, one major watchword has emerged for Western governments: resilience. But resilience in modern industry cannot be built without AI automation at the edge, something that traditional deep learning cannot accomplish. It’s time for a new approach.
Resilience and Reshoring
Global supply chains built over decades of globalisation have become fragile, and the reliance on others for manufacturing has become a liability instead of a luxury. The global pandemic, wars in Ukraine and the Middle East, and the return of trade tariffs as a geopolitical weapon have all reinforced the same uncomfortable message: security comes from self-reliance.
While restoring sovereign manufacturing and industrial capabilities has become a strategic priority across Europe, there are still major barriers to large-scale reshoring. Competition with lower wage economies and stringent decarbonisation targets mean that a wholesale transfer of industrial infrastructure is not yet viable. While governments look at legislating for supply chain sovereignty, a solution is needed to attract the investment required to make it happen. That solution is automation and optimisation.
Industry’s AI Bottleneck
However, it is with this AI-powered automation that the real challenge lies. The fundamental challenge is that the AI tools we have built to date are almost entirely wrong for the job.
Trillions of dollars have been pumped into large language models and deep learning technologies, resulting in some genuinely game-changing innovations, and tools that can revolutionise office work. But this form of AI is architecturally mismatched to the reality of factory floors, logistics networks and industrial environments. These models are trained on vast datasets, and require constant cloud connectivity to function at their best. Because of their size, they demand significant computational power, and consume energy at a scale that is becoming untenable; they are, by design, centralisedsystems.
Industrial environments are the opposite of this. They are often low-data, or even no-data, consisting of local hardware that is completely ill-equipped to run massive models, and with energy and connectivity restrictions. They are operationally dynamic, with conditions that change in ways that no static training dataset can fully anticipate.
We have seen AI make real progress in dashboards, predictive maintenance, and image-based fault detection, but real-time control and optimisation with physical
AI is the final frontier. The economic opportunity is massive, but these real-time use cases cannot abide the latency or computational requirements of a large model.
Deploying conventional deep learning into these conditions is like sending a deep-sea vessel to navigate a river. The engineering is impressive, but the fit is entirely wrong.
Building Adaptive AI
There is a misconception that the larger the model, the more adaptive it is, because it will have a higher chance of having seen a similar scenario in its training. But reacting to something it has been trained on is not the same as true adaptation. Adaptivity is when a model can react to something it has never seen before and still perform its task effectively.
Again, we can look at the human brain here. The brain does not retrain from scratch when it encounters new situations, nor does it rely on learning set responses to millions of situations. Instead, it refines its existing strategies to solve new problems; it is highly efficient and highly adaptive. This neuroplasticity provides a great template for the specialist neural networks that are needed in industrial settings.
Neuroplastic AI Models at the Edge
Rather than building models that are oversized and rigid, we should implement models that are lean by design and capable of continuous self-refinement at the point of deployment. Instead of a model trained with vast amounts of data before it arrives, we should build models that learn efficiently once they are there.
By building self-tuning mechanisms directly into the model, allowing it to adapt its own neural pathways based on real-time operational conditions, it becomes possible to deploy game-changing AI in environments that would be unviable for conventional deep learning. The sparsity of the networks means that the compute requirements are lower and the energy footprint is smaller. The inherent adaptivity means the need for cloud-based retraining is removed or severely reduced. And the performance, critically, improves with deployment rather than degrading as conditions drift away from training data.
A New Industrial Policy
Developing competitive manufacturing or industrial facilities in a high-cost Western economy requires high levels of automation. It requires energy savings and efficiency gains that are not possible solely with human labour, or even with AI dashboards. Being able to deploy adaptive AI models directly at the edge, in live control and optimisation use cases is essential.
These AI models need to thrive within systems that were built decades ago, without requiring prohibitively expensive hardware overhauls. They need to work in buildings that were not designed as data centres or power generators, connected to networks that were not built for AI workloads. And they need to improve over time without a permanent team of cloud engineers on retainer.
Europe and the West faces a choice that is becoming harder to defer. Conventional deep learning has long been the main focus of AI investment and media coverage, but a one-size-fits-all approach does not translate to reality: these massive networks are exceptional at certain tasks, but completely ill-equipped for others. For AI to be implemented effectively at the edge, enabling the reshoring of industry that will reinforce resilience across Western economies, we must begin to seriously back architectural alternatives to deep learning. Learning from our own brains is a good start.
