Abstract

Multi-modal patent claim strategies include the use of system, method and dataset claims to provide the strongest protection for AI/ML inventions. System claims broadly cover implementations. Method claims expose technical processes underpinning trained models. Dataset claims cover proprietary training data. Together, they can be combined to protect the inventive concept underlying your AI/ML innovation. Although such protection requires robust technical disclosures that satisfy enablement requirements and bolster support for subject matter eligibility, these can be generated cost effectively so that AI/ML startups can gain an edge through valuable patents that reward their breakthrough innovations.

Introduction

Early-stage AI/ML companies face a unique opportunity and challenge: machine learning innovations that could transform industries but uncertainty regarding how best to protect their valuable intellectual property. Traditional patenting strategies are usually impractical due to the exorbitant costs and uncertain eligibility requirements for software patents that typically rely on functional claiming.

Alternative strategies like multi-modal claiming through systems, methods and datasets coupled with robust technical disclosures offer AI/ML startups the best shot at patents that fulfill their goal: just compensation for transformational technologies by blocking would-be copyists, attracting investors who value IP assets and positioning for lucrative partnering or exit opportunities.

Obtaining such patents requires carefully crafted claims that precisely yet broadly cover key innovations while navigating evolving legal interpretations. Early input from an advisor experienced in AI/ML can guide startup teams through unfamiliar territory toward intellectual property protection that maximizes commercial potential and market position. Proactively filing provisional patent applications early in product development anchors inventions on record before patent bars run and embeds valuable know-how within specifications that can clarify significant value for investors, partners and acquirers.

This paper explains how patents can be cost-effectively leveraged to bring your machine learning model to market with the intellectual armor it needs to succeed. My goal is to help startups navigate complex legal requirements while maintaining a fluid, conversational tone that resonates with technical founders.

Coded in Complexity: The Law of Software Patenting

When compared to other technologies, software is especially difficult to patent. The five requirements for patentability – subject matter eligibility, utility, novelty, nonobviousness and enablement – are theoretically the same for all technologies but the courts and USPTO have formulated special tests for software. Since the early 2000s, the law of patent eligibility has become increasingly unclear, making the outcome for patent applicants highly uncertain. The cost-benefit analysis has increasingly caused startup founders to take the decision not to pursue patent protection. I intend to make the case for at least filing provisional patent applications at the early stage.

Court decisions and USPTO guidance have not focused on AI/ML but we can still draw key takeaway points for patenting such inventions:

• Patent applications in this area should be examined based on the specific techniques and machine learning processes described in the application, not merely on the results obtained through use of machine learning.
• Examiners should evaluate whether the specification provides sufficient technical details about techniques, model optimization, hyperparameters, data processing, etc. Disclosure of results without details of how they were achieved may not support eligibility or enablement.
• Describe multi-stage training of learning algorithms, including details of feature determination, data cleansing, model optimization, hyperparameter tuning, etc. The more technical depth, the stronger the application.
• Explain how data was collected, labeled (if supervised learning), tested for biases and prepared for use. Dataset curation and transformation techniques provide “structure” that supports eligibility and definiteness.
• Demonstrate possession of the full scope of claimed functionality – not just what has been built and tested so far.
• Be prepared to refine claims throughout prosecution to accurately capture the current state of the invention as its implementation progresses. Avoid attempting to claim the invention at a future time or level of development.

In summary, focus on technical details of the specific machine learning techniques, processes, hyperparameters, datasets and algorithms that implement the functions of an AI model – not just the results achieved. Providing this “structure” through robust disclosures provides the best foundation for strong, valid patent claims for AI/ML inventions.

Multi-Modal Claiming

Describing AI/ML components in terms of their underlying models, algorithms and datasets offers an alternative for “structural claiming”. Claims covering a trained model’s architecture, optimization techniques and data processing go beyond functions to detail structure. Specifying training datasets with labeled examples imparts further structure.

Multi-modal claims covering systems, methods and datasets leverage different advantages. System claims broadly encompass implementations. Method claims describe technical processes for creating/using models. Dataset claims provide strong protection for proprietary training data. All work together to protect an AI/ML invention from different angles.

How do multi-modal claiming strategies differ from traditional patent claiming strategies?

Multi-modal claiming strategies for AI/ML-related inventions differ from traditional patent claiming strategies in several ways:

1. Inclusion of dataset claims: Claiming the proprietary training datasets used to develop AI/ML models goes beyond what is typically covered in traditional patents. Dataset claims provide a unique angle of protection that captures a key aspect of AI/ML inventions-the data used to create and optimize the models.
2. Emphasis on techniques for creating/optimizing models: AI/ML claims often describe the technical techniques, algorithms and optimization methods used to generate models, rather than just focusing on the functionality of the trained models. This reveals more of the “structure” underlying AI/ML inventions.
3. Focus on model architecture: AI/ML claims frequently specify details of model architecture like layers, nodes, connections, parameterization, etc. This level of technical detail goes beyond what is typically found in functional claims for conventional software inventions.
4. Increased reliance on method claims: Alongside or instead of system claims, AI/ML claims often include method claims that precisely outline the technical processes for creating, training and using ML models. This reveals the “recipe” behind the invention.
5. Disclosure of feature engineering/data processing techniques: AI/ML specifications commonly describe the techniques used to transform raw data into “features” suitable for machine learning, like feature selection, generation, expansion, reduction, etc. This reveals crucial pre-processing of the data.

In summary, multi-modal claiming strategies aim to capture the key technical details, data processing steps, model architectures and proprietary training data that underlie modern AI/ML-related inventions. While retaining the traditional benefits of system and product claims, multi-modal strategies also leverage method and dataset claims to provide a more complete picture of what truly constitutes the “invention” for these emerging technologies.

Challenges in Drafting Multi-modal Claims

Here are key challenges in drafting multi-modal claims for AI/ML inventions:

1. Selecting appropriate “structure”: Determining what constitutes sufficient “structure” for an AI/ML invention, including debate over whether datasets, algorithms and model architecture qualify. Guidance from the USPTO and courts remains limited.
2. Scope of dataset claims: Defining the appropriate scope of dataset claims, including the types and amounts of labeled data needed to support possession of the invention. The boundaries of what constitutes an enabling dataset are still being defined.
3. Changes over time: AI/ML techniques, model architectures and datasets evolve rapidly over time. Claims must capture the invention at a point in time while remaining broad enough to cover future improvements.
4. Lack of standardization: The absence of standardized language, datasets and techniques for describing AI/ML inventions. Practitioners must devise custom ways of articulating inventions that examiners will find clear and comprehensive.
5. Defining “inventive concept”: Identifying the aspects of an AI/ML invention that truly represent the “inventive concept” worthy of patent protection, rather than routine or conventional elements. This impacts what is claimed.
6. Inherent variability: The inherent variability in outputs from AI/ML models due to random initialization, optimization techniques, etc. Claims must capture the invention despite output variability.
7. Enablement of datasets: Meeting the enablement requirements for large, proprietary datasets. The specification must teach how to acquire or recreate a representative dataset without undue experimentation. (See below for further discussion.)

In summary, there are challenges to overcome in effectively drafting multi-modal claims for AI/ML inventions, including navigating evolving legal requirements, a lack of standardization, defining what constitutes “structure” and demonstrating proper enablement of datasets.

Meeting the Enablement Requirement for Proprietary Datasets

Here are strategies for meeting the enablement requirements for large, proprietary datasets:

1. Provide a representative dataset sample: Include a small sample of records from the full dataset in the specification to demonstrate that the invention works as claimed. The sample should represent the types, attributes and coverage of data in the full dataset.
2. Describe data collection and curation techniques: Explain the techniques used to collect, filter, organize and label the dataset, such as web scraping, crowdsourcing, data cleaning algorithms, etc. This shows how one of ordinary skill could recreate a similar dataset.
3. Disclose data transformations and preprocessing: Document any transformations, feature engineering or normalization performed on the raw data before using it to train the ML model. This shows how the data was made suitable for machine learning.
4. List external sources of similar data: Identify external sources (public or commercial) where data with similar attributes, coverage and quality are available. This demonstrates that one of ordinary skill would not face undue experimentation in acquiring an equivalent dataset.
5. Include details on labeling process: For labeled datasets used in supervised learning, describe the process by which examples were labeled and the quality control measures taken. This reveals how labels were reliably assigned.
6. Refer to related publications/patents: Direct the reader to previously published research papers, patents or other disclosures that describe datasets like the one used to develop the invention. This demonstrates that one of ordinary skill would be enabled to access equivalent data.

Patent vs Trade Secret Protection

While both patenting and trade secrecy offer advantages for protecting AI/ML innovations, they differ fundamentally in their requirements regarding disclosure, limitations on term and enforceability. Practitioners must weigh the risks and benefits for their specific inventions to determine the optimal intellectual property strategy.

Here are some of the pros and cons of patenting AI/ML inventions versus keeping them as a trade secret.

	Patent	Trade Secret
Pros	Public disclosure of the invention in exchange for exclusivity: By disclosing details of the invention in a patent, competitors are put on notice and cannot freely use or implement the idea. This provides a legal monopoly for 20 years.Blocking patents: Patents can block competitors from implementing similar technologies, providing a strategic advantage. Competitors must license or “invent around” the patent.Licensing income: Patented inventions can be licensed to generate ongoing revenue streams. This may be difficult with trade secrets.Marketing advantage: Patents can signal legitimacy, credibility and competitive edge, which can help attract investors, talent and customers.	No disclosure of details: Technical details remain completely confidential, making reverse engineering or “inventing around” difficult for competitors.No costs: There are no application fees, maintenance fees or litigation costs associated with trade secrets.Perpetual duration: Trade secrets can remain protected indefinitely so long as secrecy is maintained. They are not time-limited like patents.
Cons	Cost: Patent applications and maintenance fees incur ongoing costs. Patents must be enforced through litigation to achieve their full value, which is also expensive.Disclosure of details: Patents necessarily disclose technical details of the invention that could aid competitors. Trade secrets can maintain complete confidentiality indefinitely.Limited life: Patents last for 20 years from filing, after which point the invention enters the public domain. Trade secrets can be kept indefinitely.	No legal monopoly: Trade secrets provide no legal recourse against independent invention, discovery by reverse engineering, or inadvertent disclosure. Patents do.Difficult to enforce: It can be difficult to prove trade secret misappropriation in court and determine appropriate damages. Patent infringement is easier to prove.Requires constant protection: Trade secrets demand constant vigilance, monitoring and security procedures to maintain secrecy. This ongoing effort can be costly and impractical.

For AI/ML startups, key factors determine the best IP strategy:

• Competitive landscape: In a “winner-take-all” market, patents provide an edge competitors must navigate. Trade secrecy may work best in fragmented spaces.
• Capital needs: Patents can attract investors who value IP assets, but maintaining patents requires cash. Trade secrecy avoids costs but limits partner/licensing opportunities.
• Team expertise: Startups early in development with limited tech know-how benefit from the structure patents require. Mature teams focus internally using trade secrecy.
• Time horizon: Patents give 20 years of protection but take time to obtain. Trade secrecy provides immediate protection while secrecy can be maintained.
• Revenue model: Licensing opportunities make patents an asset for startups pursuing such strategies. Trade secrets rely more on secrecy and transactions’ confidential nature.
• Necessity of disclosure: Startups seeking funding, partnerships or acquisitions benefit from patents’ “teaching” of the invention to explain value. Trade secrecy limits what can be shared.

Conclusion

AI/ML inventions offer huge promise for breakthroughs across many fields, but obtaining suitable intellectual property protection remains a challenge due to their functional nature. Multi-modal claiming strategies through system, method and dataset claims provide the strongest coverage by revealing various aspects of the “inventive concept.” Robust technical disclosures that satisfy enablement requirements and bolster support for subject matter eligibility further strengthen protection.

For early-stage AI/ML companies, proactively filing protective provisional patent applications that employ these strategies represents the first step toward valuable patents that reward their hard-won innovations. Starting the patenting process early avoids losing valuable intellectual property rights due to statutory bars or public disclosures.

The attached appendices include additional comments relating to specific industry verticals including medical devices, agriculture, and engineering and construction.

Appendix 1: Medical Devices

Here are some considerations for patenting AI/ML innovations in the medical device space:

• Medical device makers are increasingly incorporating AI/ML technologies to assist with tasks like automated diagnosis, monitoring vital signs, image recognition and personalized treatment recommendations.
• Examples include AI-powered ECG monitors, imaging systems for detecting anomalies, vital signs analyzers and surgical robots that use computer vision. The applications of AI/ML in medtech are vast and growing rapidly.
• Functional claiming remains common in medtech patents, especially for devices that leverage AI to perform automated tasks. However, the same challenges regarding eligibility, definiteness and scope apply.
• To strengthen patents, medical device companies are employing strategies like multi-modal claiming covering systems, methods and datasets. Method claims reveal the technical processes underlying AI models.
• Dataset claims provide protection for proprietary training data used to develop AI algorithms. This can include labeled examples of MRIs, CT scans, ECG readings and other medical data.
• For human-involving inventions, demonstrating improvements to technology or significant additional steps may support eligibility under Vanda. Describing ML techniques, model optimization and data preparation provides “something more.”
• The USPTO’s guidance (see below) applies equally to medtech AI innovations. Examiners focus on technical details of model development, not just functionality. Disclosing feature engineering, preprocessing, hyperparameter tuning, etc. supports enablement.
• Overall, while patenting AI/ML innovations for medical devices presents unique considerations, strategies like multi-modal claiming, emphasizing improvements to technology and providing robust technical disclosures can help secure valid, defensible patents.

In summary, AI/ML innovations increasingly assist medical professionals. Patent strategies must demonstrate “something more”.

Appendix 2: Agriculture

Considerations for patenting AI/ML innovations in agriculture include:

• AI and machine learning are increasingly being applied in agriculture for tasks like crop monitoring and yield forecasting, soil analysis, irrigation management, pest detection and precision spraying.
• Examples of agricultural AI/ML applications include drone-based crop monitoring, soil moisture sensors with AI analytics, computer vision systems for detecting diseases and anomalies in crops, and autonomous ground vehicles for precision weed control and spraying.
• Many of the same challenges regarding functional claiming, eligibility, definiteness and scope apply to agricultural AI/ML inventions as for other domains. However, patent protection remains critical for recouping investments in R&D.
• Multi-modal claiming strategies covering systems, methods and datasets can help strengthen agricultural AI/ML patents. Method claims reveal the technical processes behind trained models, which are increasingly key assets for agribusinesses.
• Dataset claims provide protection for proprietary labeled training data, such as images of crops with annotated regions of disease, soil samples tagged with nutrient contents, and satellite imagery mapped to yield forecasts.
• Demonstrating improvements to the function, way or results of a prior art agricultural process – for example through AI-enabled precision or automation – can support eligibility.
• Patent applications for agricultural AI/ML should emphasize technical details like data preprocessing, feature generation, model optimization techniques, and hyperparameter tuning to provide well-supported disclosures.
• With careful application of these strategies and disclosures that reveal the technical “structure” behind AI/ML innovations, agribusinesses stand to gain valuable intellectual property protection for their transformative AI-powered solutions.

Appendix 3: Engineering/Construction

Considerations for patenting AI/ML innovations in engineering and construction include:

• AI and ML are increasingly applied to challenges in engineering and construction for applications like project scheduling optimization, materials defect detection, fault diagnosis, equipment monitoring and failure prediction.
• Examples of construction AI/ML innovations include computer vision systems for detecting cracks in structures, defects in 3D printed parts, and wear of metal components; predictive maintenance models for heavy machinery; and optimization algorithms for scheduling tasks and allocating resources.
• The use of drones, robots, and IoT sensor data in construction also provides vast amounts of data that can be used to train ML models for automation, optimization and safety improvement.
• Multi-modal claiming strategies through dataset, system and method claims remain important for patenting construction AI/ML inventions. Method claims detailing technical processes and dataset claims encompassing labeled training data strengthen protection.
• Emphasizing technical details of model training, optimization techniques, hyperparameter tuning, and data preprocessing/feature engineering in disclosures provides “structure” that satisfies enablement requirements.
• Demonstrating improvements to safety, efficiency, precision or speed through AI/ML compared to conventional engineering processes can support subject matter eligibility when coupled with robust technical disclosures.
• Successfully patented construction AI/ML inventions reveal not just outputs or results, but how labeled training data and technical processes underlying the models allow them to generate superior outputs compared to traditional techniques.
• With strategic claiming approaches that leverage the advantages of systems, methods and datasets, and robust disclosures, AI/ML innovations offer transformative potential for engineering and construction through valuable intellectual property protection.