Conclusion
- The Federal Circuit decision in April provides an initial perspective on benchmarks to ensure successful claims for artificial intelligence and machine learning algorithms.
- Improvements in technology are likely to be a key factor in successful claims, but it is still unclear whether improvements in machine learning will be sufficient.
- A claim that focuses on how algorithms can be applied beyond the improvements it provides in machine learning is more likely to succeed given the current legal situation.
Although patent applications for artificial intelligence and machine learning have been surged in recent years, legal guidance on the patentability of AI and machine learning algorithms remains lacking.
US Court of Appeals for the First Critical AI Judgment of the Federal Circuit Redive analysis v. Fox Additionally, the latest guidance from US patent and trademark offices on machine learning models remains an important unanswered question. Can the invention be patented if it uses an improved (i.e. non-genetic) machine learning model but does not make other technical contributions?
General software patent eligibility has been in liquidity since the US Supreme Court Alice Corp.Pty. v. ClsBank Int'l 2014 opinion Alicethe court explained that the use of “generic computers” does not otherwise translate abstract ideas or non-technical ideas (such as business methods) into patentable correspondence.
At the time, some practitioners feared that this was the end of all software patents. However, in 2016, the Federal Circuit explained Enfish, LLCv. MicrosoftCorp. The software is also subject to a patent, even if it is used to improve the functionality of a computer. in Enfishfound that the software has improved the way “computers store and retrieve data in memory” using the new “data structures.”
That same year, the court decision McRo, Inc. v. BandaiNamco Games Am. Inc., The software explained that it could be subject to patents if it offers improvements in other technical fields, such as “3-D computer-generated lip synchronization” in computer animation. This set of case law dictated that claims that provide “specific technical solutions to technical problems” are eligible to obtain a patent.
But where does machine learning fit into this patent eligibility situation? Does the Federal Circuit simply classify it as traditional software or do they perceive it as having a distinct nature?
Machine learning models are inherently more technically advanced than traditional software programs. Unlike traditional software programs that contain a given set of instructions based on a particular set of inputs, machine learning algorithms use training data and iterative optimization steps to improve the ability of the model to respond to a particular set of inputs.
Humans cannot easily predict what a machine learning model outputs using a particular set of inputs. The technically complex nature of machine learning algorithms raises important questions. Should the technical improvement thresholds be diverged from one another for such algorithms to be patent qualifying?
AI Guidance
In July 2024, in front of the Federal Circuit recently Decisions, PTO issued patent eligibility guidance based on existing software law, providing examples of patent eligibility for machine learning.
- One example shows that patent applicants cannot request the use of artificial neural networks (type of machine learning models) for the general detection of anomalies in a dataset. However, according to the PTO, applicants can get the claim that they use artificial neural networks to detect and block network traffic anomalies (Prevents malicious attacks).
- In another example, the general use of deep neural networks to resolve mathematical formulas related to speech analysis was not considered eligible. However, we improved speech analysis by using the same machine learning algorithms to separate the desired audio signal from external or background audio.
Considering the software eligibility case for the Federal Circuit at the time, this guidance provided no surprise. It matches the cases such as MacrosPTO found that patent claims are eligible if machine learning algorithms provide technical improvements to network traffic and voice signal analysis.
This guidance is agnostic to machine learning innovations. It says nothing about whether improvements in the performance or structure of the machine learning model will lead to eligible claims.
Recent analysis
Although the PTO guidance appears to be the opposite, the patent applicant continued to file claims for the non-technical use of the machine learning model. Recent patents were one such example. We use known machine learning technologies to optimize advertising revenue for live events and TV scheduling.
The oral discussion focused on the complexity of machine learning itself, rather than the recent revenue optimization. He emphasized that computers that run machine learning models are “special purpose computers” that find connections between inputs that humans do not.
The Federal Circuit was not convinced. “We do not argue that a patent is eligible due to the fact that it uses existing machine learning techniques) performs tasks performed at higher speeds and efficiency than humans could previously accomplish.”
The court found the claim to be ineligible because it “just argued for the application of machine learning in general.” Therefore, the results were closely tracked Alice– “Adds to “advertise on a computer” is not enough, such as “state abstract ideas” advertising revenue optimization.
Applying common (non-existent) machine learning models to well-known tasks was not met. Alice threshold.
I'll track it closely Alice In holdings, the Federal Circuit focuses on machine learning innovation. recently Nevertheless, the patent eligibility enquiries were notable. The court was very interested in technical improvements to machine learning algorithms. That could have been provided in itself.
Many times, the courts juxtaposed ideas for technical improvements with machine learning models. For example, the court discussed the “requirements for machine learning models.” “Iteratively Trained” or dynamically adjusted… The patent does not represent a technical improvement. ”
The court also worried that the patent “does not portray the steps that machine learning technology achieves improvements,” as it was concerned that the patent “does not assert a specific way to “improve mathematical algorithms or improve machine learning.”
The language appears to indicate that improvements to the machine learning algorithm itself may be a type of technical improvement that qualifies patents. But both recently Both the opinions and PTO's previous AI guidance explicitly discuss whether the types of improvements or innovations in machine learning algorithms can save abstract ideas that are otherwise unsuitable.
The model gap has been improved
There appears to be an important gap in current machine learning eligibility guidance as to what improvements constitute a sufficient “technical improvement.”
Can someone claim an improved machine learning model that does not necessarily have an improvement to the overall computer (or other technical) functionality?
For example, if a machine learning model uses an architecture or training algorithm (such as optimization methods or loss features) that are linked to improve the performance metrics of the model, is it sufficient for patentability? Or should the improved performance of the model be reflected in the improved performance of the computer itself to achieve eligibility?
Best Practices
Considering the ambiguity that follows that follows: Recent analysis PTO's recent AI guidance practitioners should adopt a strategic approach when drafting machine learning-related patent applications. There are several practices that allow you to maximize your patent eligibility while navigating through these uncertain bodies of water.
Strategically bill your draft by focusing on technical applications. Wherever possible, patent practitioners should emphasize “technical improvements” of their claims by reciting specific downstream technical applications of improved machine learning models.
A likely approach that can often be cautious is to focus independent claims on core machine learning innovations with dependency claims that know some critical applications. This will maintain broader protection against co-innovation in independent claims, but includes dependent claims that are likely to withstand eligibility challenges under current eligibility laws.
Describe improvements in machine learning performance. Both the court and the examiner have given clear preferences to detailed technical explanations regarding the declaration of improvement in general. This specification should provide a comprehensive explanation of how claimed innovations improve machine learning-related performance metrics.
This specification also requires the establishment of a clear nexus using extensive and conclusion statements between the innovative features of machine learning algorithms and improvements in machine learning-related performance.
Connect machine learning improvements to broader technological advances. Improvements in machine learning frequently lead to secondary technical benefits beyond the system itself. These connections can strengthen the eligibility debate by linking machine learning innovations to a clear eligibility precedent by establishing established technical domains.
Specifications need to clarify in sufficient detail how AI improvements contribute to technical improvements, such as improved computational efficiency, network security, data integrity, hardware system operation, and technical issues such as image processing and signal analysis.
Keep patent family pending. Applicants should consider holding applicants' families on hold through ongoing applications in order to benefit from potentially favorable changes to the law.
For example, applicants who are forced to narrow down their claims by including specific downstream applications may be able to later ensure broader protection against core machine learning innovations if the law evolves to recognize improvements in machine learning as inherently technical.
This approach requires caution in supporting written explanations for broader claims. The specification should comprehensively describe machine learning innovations independent of the application, and detail specific implementations to support narrower claims when necessary.
During the Federal Circuit recently The decision suggests that advances in machine learning algorithms may be sufficient to qualify a claim, and does not provide a definitive answer to this question. The PTO's latest guidance on AI is silent on the issue.
In light of this lack of clarity, practitioners should try to highlight both improvements in the AI model itself and related technical improvements (or other technical areas) of computer functions. Such an approach increases the likelihood of benefits and benefits from potential changes in the law. recently Only hints were provided.
This article is based on Bloomberg Act, Bloomberg Tax, Bloomberg Government, or its owner, Bloomberg Industry Group, Inc. It does not necessarily reflect the opinions of the
Author information
Dr. Lidiya Mishchenko is a special advisor to Mololamken, focusing on appeal and on intellectual property and technology-related litigation.
Pooya Shoghi is an attorney for the AI committee of Lee & Hayes and specializes in patent prosecution and strategic counseling for software and electronic technology.
Written for us: Author's Guidelines
