Across the portfolio, 2020 was a year of broad, incumbent-driven accumulation in neural-network training. A PatentBear sweep for neural-network training methods in the 2020 publication window surfaces a familiar roster of assignees clustering in CPC G06N 3/08 (learning methods) and G06N 3/0454 (deep architectures): International Business Machines, Samsung Electronics, Google, Adobe, and Intel all appear repeatedly. The standout structural fact is breadth — no single assignee dominates; the field is a crowd of large players each staking method claims.
A representative exemplar is IBM's US10839226B2, simply titled "Neural network training," granted November 17, 2020, with CPC G06N 3/08 and a vision code G06K 9/00758. It is the kind of foundational training-method grant that recurs across the year — narrow in its specific claims, broad in the assignee's intent to hold position across the training stack.
Two caveats are load-bearing for anyone reading these counts. First, they are filing-date sensitive: a 2020 grant reflects an application filed years earlier, so the landscape describes who was investing in the late-2010s, not who is leading now. Second, a count of grants is not a measure of market position — owning many training-method patents does not, by itself, establish dominance in any product market. It establishes a defensive and cross-licensing posture.
The whitespace observation: in 2020 the cluster is heavily about training convolutional and recurrent vision/speech models. The transformer-and-generative wave that would dominate later filings is barely visible in the granted record yet, precisely because of the prosecution lag. The 2020 landscape is, in effect, a snapshot of the pre-transformer training era frozen into enforceable rights.
For the IP reader, the practical read is this: if you are mapping who owns training-method IP, start with the incumbents — IBM, Samsung, Google, Intel, Adobe — and anchor any portfolio claim to specific grants like US10839226B2 rather than to the aggregate counts, which are sensitive to how and when you query.