Reducing AI’s Computational Burden: The Case for Hybrid AI

Why Reducing Uncertainty in AI Matters

DeepSeek is the most publicised result of these efforts, but other forms of neural networks have achieved even greater reductions in processing. We will discuss a case where elimination of uncertainty within the numerical method results in better performance, and how that has been achieved.

The Rise of Hybrid AI Solutions

Neural networks are adapted to produce specific solutions to a problem. They can incorporate whatever rules-of-the-game that we wish to add. They can also be structured to make the calculations easier to perform by a computer, either by using tables of precomputed results, or by eliminating any redundancies within the mathematics. One type of hybrid neural network is the Physics Informed Neural Network (PINN).

How PINNs Improve AI Efficiency

3D image reconstruction is determined by optical laws, which constrain the range of inputs and possible outputs of the network. The restrictions are not contentious and do not result in a bias. Efficiencies come from using fundamental laws, such as light travels by the shortest route, or they can be classical image processing methods that eliminate redundant AI calculations. Constraints can be introduced within the neural network, or by modifying the data submitted to the network.

The application of these techniques to the 2D -> 3D reconstruction problem has resulted in a sixtyfold reduction in processing effort compared with current AI techniques.

Hybrid AI solutions, such as PINNs, offer a way of incorporating intellectual property within the AI process. SMEs with successful classical solutions for their chosen vertical market should first identify where AI adds value for their clients and determine whether a targeted AI solution will meet their needs.