The hard part wasn't the code. The hard part was the thinking that produced it (medium.com)

In the latest installment of "The Engineer’s Journey," the author reflects on the critical realization that their AI model, initially designed for classification, needed a fundamental shift to succeed in its real-world application focused on similarity search. This transformation was prompted by acknowledging a mismatch between the training objective—classifying images into fixed categories—and the ultimate goal of finding the closest known example based on learned embeddings. The discovery that the model's architecture was inappropriate led to a deep dive into similarity learning strategies, ultimately steering the author toward the ArcFace approach, initially developed for face recognition. ArcFace offers a significant advantage by using angular margins rather than traditional linear decision boundaries, making it more suitable for scenarios with high intra-class variation, like differentiating between plant species or faces under varying conditions. This approach not only aligned the training and serving objectives, enhancing model effectiveness but also allowed for a more efficient use of resources. By leveraging the SubCenter variant of ArcFace, the author was able to incorporate multiple embedding clusters for each class, accommodating varying appearances within species and optimizing the model's performance within the revised mobile size budget. This transformative thinking and engineering decision underscore the project's evolution and set the stage for the upcoming testing phase.