Here's what actually issued. On March 24, 2020, International Business Machines Corporation was granted US10599977B2, "Cascaded neural networks using test output from the first neural network to train the second neural network," inventors Hiroki Nakano and Masaharu Sakamoto. The CPC mix pairs the network class G06N 3/0454 with a stack of medical-imaging G06T codes (G06T 7/0012 and several 2207 subgroups), pointing at a vision or diagnostic application.
The mechanism is a cascade. Instead of training two models independently, the method uses the output the first network produces on test data as a training input for the second. In effect, the first model becomes a teacher — its predictions, including its uncertainties, shape how the second model learns. The imaging CPCs suggest the worked example is medical image analysis, where a first pass might localize and a second refine.
Why patent a training arrangement? Because the arrangement is the invention. Many machine-learning patents are not about a new neuron or a new layer but about how existing pieces are wired together and trained. A cascade that improves accuracy on a hard task — like reading a scan — is exactly the kind of procedural insight worth a grant.
On scope, the standard caveat: granted and enforceable, but the claims cover this specific cascade-training method, not the general idea of using one model to train another, which appears across the literature under names like distillation and pseudo-labeling. The boundary is the claimed steps.
The takeaway: US10599977B2 is a reminder that enterprise incumbents like IBM patent the plumbing of training pipelines, often anchored to a concrete vertical — here, imaging — rather than the splashy model architectures that dominate headlines.