6. Validation using MNIST
In this section, we'll look at the results following the validation of IIC using the MNIST test dataset. After running the cluster prediction on the test dataset, the linear assignment problem assigns a label to each cluster, essentially converting the clustering into classification. We computed the classification accuracy, as shown in Table 13.6.1. The IIC accuracy is higher than the 99.3% reported in the paper. However, it should be noted that not every training results in a high-accuracy classification.
Sometimes, we have to run the training multiple times since it appears that the optimization is stuck in a local minimum. Furthermore, we do not obtain the same level of performance for all heads in multi-head IIC models. Table 13.6.1 reports the best performing head.
| Number of heads | 1 | ...
