How precise does a patent have to be when it is describing the usage of a neural network?

Question

While preparing a presentation about metadata, I stumbled upon the following patent: Automatic extraction of metadata using a neural network.

I have no experience with patents, but the strategies for metadata creation presented in the patent seem very generic to me - not only because the imagery kind of says: "We put text into a NN and it gives us the data" (see the cover of the document) - but also because the interesting parts of the pseudo code are very abstract (see part 15 of the document).

How specific do these kinds of patents need/tend to be? How similar must my approach be so that it counts as "the same invention"?

Answer 1

I am not a lawyer, but I believe since the Alice decision in 2014, many applications of existing and well known algorithms are much less patentable. Thus, it is quite possible that this patent would not have been granted if filed since Alice. In any case, this patent's priority date is 4-30-1998 so it should be expired by now.

As will most cases of software related patents, the specifics are very important and I doubt general guidance is of value. If you have a specific idea you want to patent, I would highly recommend consulting with a patent attorney well versed in the field.

Answer 2

There are two distinct parts of a patent, the specification and the claims. The claims are what nails down the scope of the protected invention. This claim 1 seems reasonably specific. To infringe the claim, each step a-e must be performed.

1. A method of automatically extracting metadata from a document, the method comprising: (a) providing: a computer readable document including blocks comprised of words, an authority list, including common uses of a set of words, and a neural network trained to extract metadata from compounds; (b) locating authority information associated with the words by comparing the words with the authority list; (c) creating compounds, a first of the compounds describing a first of the blocks and including: first-block words, descriptive information associated with one of the first-block and the first block words, and authority information associated with one first-block word; (d) processing the compounds through the neural network to generate metadata guesses; and (e) deriving the metadata from the metadata guesses.

Answer 3

My firm has been patenting and applying for patents in the fields of AI, machine learning, deep learning, computer vision, NLP, and other related technologies -- mostly related to life sciences and healthcare. What we've found is that there are a lot of "generic" patents issues such as the one you describe. However, the USPTO is increasing its scrutiny, and the European Patent Office (EPO) has had a pretty high bar for such patents for years. In a nutshell, what we have found, to obtain patents on AI-enabled technologies, is to describe new pre-processing and post-processing steps, new combinations of data, new combinations of nets and algos, and all encompassed by claims to a "system". By drafting patent applications this way, we have overcome the "Alice" rejections mentioned above, as well as the USPTO "two-step" analysis. This is all an oversimplification, of course, and each invention will be a unique circumstance. However, this is a pretty good guideline to what should be allowed as patentable subject matter. Hope this helps!