Nov 28
Debates
Patents on Machine Learning Systems
There's a problem. The patent office doesn't consider model training to be an invention
Patents on Neural-Net models
Backpropagation dates from 1986, and
is off-patent now. I don't know if it was ever patented
Cases where patents help tech startups
Microsoft tried to patent the BlueJ Java IDE's "object workbench", too, although they backed down.
1. Contesting patents in court is very expensive ($2 million?)
2. Even contesting these patents via a so-called IPR (a special hearing procedure within the PTO) costs an average of $300-500,000.
3. There is no free mechanism for saying "this patent is fraudulent"
4. Supposedly there is a penalty for willfully failing to list all known prior art, but the PTO almost never applies this.
6,865,185, Patel et al
There are a lot of claims for this patent. Claim 1 recites:
A method for queuing traffic in a wireless
network, comprising:
receiving a stream of packets for transmission in a
wireless network, each packet including a flow identifier uniquely
identifying the flow;
determining for each packet based on the included flow
identifier a location for a corresponding flow, the location comprising at
least one of a sector in the wireless network, a latitude and a longitude
of a mobile device associated with the flow, and a specific beam within a
sector of the wireless network; assigning each packet to
one of a plurality of virtual groups based on the location for the
corresponding flow, the virtual groups comprising discrete transmission
resources, and queuing each packet in an assigned virtual group for
transmission in the wireless network.
So far, so good: packets are assigned to virtual groups based on data. But then many later claims add additional possible data, without specifying how the data should be used. In effect, the patent covers any use of data from a large number of categories:
These are all legitimate network parameters, but they are also extremely
well known.
8,793,151
Claim 1 is about risk analysis. But there are no details about the "'adaptive risk analysis engine".
10,826,941
Claim 1: A method for protecting an enterprise
network, the method comprising, at a system comprising one or more
processors and memory that are remote from the enterprise network:
controlling communications to and from the enterprise network according to
a set of security policies;
controlling endpoint to endpoint connections within the enterprise network
according to the set of security policies;
receiving a request for modifications to one or more policies of the set
of policies;
automatically generating a policy digest formatted according to a
predefined format, the policy digest comprising the modifications, and
storing the policy digest in the memory;
retrieving the policy digest from the memory;
generating one or more calls to one or more system components that control
the communications to and from the enterprise network and the endpoint to
endpoint connections based on the policy digest; and
modifying control of the communications to and from the enterprise network
and the endpoint to endpoint connections based on the one or more calls.
Security policies are well-known. The patent does not define them. Another part of the patent is about central control of security, but this isn't new either.
Claim 1 is the "method" claim and Claim 12 is the nearly identical "system" claim. Most patents do this.
Start with i4i and NTP cases
Apple patents
Paul Graham
KSR v Teleflex
Bilski
Mayo Labs v Prometheus Labs, Myriad Genetics
Abstract patents
Ultramercial
Alice
Lawsuit Venue
Crime