Design for AI

Problem Context

The goal of measurement- or diagnostic systems is to provide their user with a classification or quantitative result. When AI is used to process the data that has been collected by the hardware layer of such a system, the overall performance of the system is determined by the amount of AI-usable information that is contained in the captured dataset. Currently the hardware layers are designed using traditional design rules, resulting in the data to be focused towards human-interpretability or based on incomplete physical models of the system. Consequently, the hardware is currently unlikely to capture the most informative dataset. To overcome this limitation, and thus to allow the development of better performing solutions, future systems need to be developed using an integrated design approach that concurrently optimizes the hardware- and AI-based processing layer.

Solution

To gain further experience with this upcoming paradigm shift we used a previously built lab setup to test various approaches for the concurrent design. This included optimization of the physical Spatial Light Modulator (SLM) in the setup, while also using of the simulation environment for its flexibility and high processing speed, thus two parallel development routes were adopted. In the simulation environment, optimization methods for the SLM were tested. It should be noted that, in contrast to last year's approach, the SLM is not a layer of the deep learning model, and therefore it can no longer be (efficiently) optimized using backpropagation. This is essential for translation to the physical setup, where backpropagation of the deep learning gradient through the optical pathway is not feasible.

After exploring multiple routes, the strategy for optimising the SLM was chosen to be based on a genetic algorithm. The genetic algorithm is an optimization method based on natural selection, the process that drives biological evolution. The genetic algorithm uses genetic operators such as crossover (which combines the traits of two solutions to create a new solution) and mutation (which introduces random changes to a solution) to create new candidate solutions.

Results

The main new knowledge is related to the use of hardware optimization techniques, in combination with optimization of a deep learning model. The use of genetic algorithms is effective, but slow, and especially in a physical setup the duration of the optimization is likely to be a bottleneck. We experimented with efficient use of samples, by minimizing the number of samples needed for a single training epoch and maximizing the improvements per epoch. Also, we identified the length of image acquisition in the hardware setup as a key limiting factor and we improved this in our current setup. The optimization was done for an SLM with 400 parameters, but the concept can easily be expanded to a larger number of parameters, even though many hardware systems might have much less parameters that need to be optimized.