The authors simply implement a continued fraction library in Pytorch and call the backward() function on the resulting computation graph.
That is, they chain linear neural network layers and use the reciprocal (not RELU ) as the primary non-linearity.
The authors reinvent the wheel countless times:
1. They rename continued fractions and call them ‘ladders’.
2. They label basic division ‘The 1/z nonlinearity’.
3. Ultimately, they take the well-defined concept of Generalized Continued Fractions and call them CoFrNets and got a patent.
IBM's lawyers can strip out all the buzzword garbage if they feel litigious and sue anyone whose written a continued fraction library.
Because, that's what the patent (without all the buzzwords) protects.
You sent me down a rabbit hole. In trying to track it down for myself I read a couple of others that I thought might be it, and was stunned by how obtuse these patents are.
What sort of leverage does this stuff provide? You mentioned "charge rent". What does that look like?
Honestly, I don't even know where to begin. It's insane IBM owns the patent to continued fractions.
If you wrote a continued fraction class in Pytorch and called backwards (or even differentiated the power series) then you're infringing on their copyright.
Here it is: https://patents.justia.com/patent/20230401438
On Google Patents: https://patents.google.com/patent/US20230401438A1/en
The authors simply implement a continued fraction library in Pytorch and call the backward() function on the resulting computation graph.
That is, they chain linear neural network layers and use the reciprocal (not RELU ) as the primary non-linearity.
The authors reinvent the wheel countless times:
1. They rename continued fractions and call them ‘ladders’. 2. They label basic division ‘The 1/z nonlinearity’. 3. Ultimately, they take the well-defined concept of Generalized Continued Fractions and call them CoFrNets and got a patent.
IBM's lawyers can strip out all the buzzword garbage if they feel litigious and sue anyone whose written a continued fraction library. Because, that's what the patent (without all the buzzwords) protects.
Thanks for that. That is patently absurd.
You sent me down a rabbit hole. In trying to track it down for myself I read a couple of others that I thought might be it, and was stunned by how obtuse these patents are.
What sort of leverage does this stuff provide? You mentioned "charge rent". What does that look like?
Honestly, I don't even know where to begin. It's insane IBM owns the patent to continued fractions.
If you wrote a continued fraction class in Pytorch and called backwards (or even differentiated the power series) then you're infringing on their copyright.