Dynamic Vector Router system hits 97.17% on MNIST

The write-up frames its approach as a response to two common pain points: the cost of training or fine-tuning huge models without high-end GPUs, and the difficulty of retraining an entire model whenever new classes or data arrive. Instead of designing perfect clusters or experts up front, it starts with rough vector-space grouping and lets small CNNs grow as data streams in.

The streaming simulation on 60,000 MNIST training images shows models being created on the fly, with clusters labeled cluster_001 through cluster_010 each accumulating about 3,000 images and reporting per-cluster accuracies such as 98.7% and 99.4%. The test evaluation then adds an uncertainty filter: samples that fall below the similarity threshold are returned as Unknown rather than forced into a class.

The author positions this as a practical alternative to training one giant model, noting that additional learning only touches the relevant cluster's model rather than the whole system. The prototype is limited to MNIST, so how it scales to more complex image data remains to be seen.