Artificial intelligence is a fresh field ripe for patenting of many forms of inventions, including smaller improvements to existing methods. It is important to keep in mind that an idea need not be revolutionary to be protected by patent and commercialized – incremental improvements are also valuable and can be protected. Many of these incremental improvements could relate to novel inputs and outputs or improvements to data sourcing, formatting, or labelling.

Improvements to Training Methodology

When considering patentable aspects within machine learning techniques, training techniques can be a source of innovation. You may come up with methods that enhance the speed of training, enabling learnable parameter values to converge faster. It could involve the development of entirely new algorithms for training, leading to more efficient computational resource utilization and faster processing.

Additionally, an inventor might devise solutions to address common issues in existing training techniques, such as overfitting or the resource burden that arises with increasing dimensionality. These improvements could be achieved by introducing new objective functions or loss functions and exploring alternative parameter updating approaches beyond conventional gradient descent.

Furthermore, innovation might lie in the way existing model elements are structured to enhance training outcomes, either by accelerating the process or increasing accuracy. For example, an inventor might rearrange filters or layers in their model, potentially creating separate arrangements for training and inference purposes.

Improvements to Data

Your invention could also be improving the training data itself. Some examples of this could be:

• Improved Data Format: One patent claim focuses on configuring the training data into vectors using a block floating-point format. This representation enhances the efficiency of processing the training data, particularly on processors like GPUs, leading to improved training performance.

• Data Sourcing and Labeling: Another invention involves novel techniques for sourcing training data or automatically labeling it for supervised learning. This innovation addresses the challenges of obtaining high-quality labeled data and can lead to more accurate and effective models.

• Reducing Data Labeling Needs: Some techniques in patent claims reduce the necessity for data labeling. This could include semi-supervised learning approaches or methods that require less labeled data while still achieving high model accuracy.

• Data Augmentation: Patent claims also cover data augmentation techniques, where existing data is transformed to create variants with different angles, lighting conditions, and other factors. This process significantly increases the size of the training dataset and enhances the model’s ability to generalize.

• Novel Data Preprocessing: Some inventions involve unique data preprocessing methods that prepare the data before feeding it into the model. These preprocessing techniques can improve the quality and relevance of the training data, ultimately leading to better model performance.

Training techniques offer a broad range of possibilities for innovation within the realm of machine learning, presenting opportunities for inventors to secure patent protection for their advancements in this rapidly evolving field.

Inputs and Outputs

Even in cases where most of the machine learning process uses off-the-shelf components, you can find novel and patentable aspects in how the input and output are handled.

On the input side, the novelty might lie in the type of data being used. Perhaps there is a new data stream that you are leveraging and applying machine learning to extract valuable insights. The representation of the input data could also be innovative, especially if you are incorporating additional data sources or metadata to inform the model’s training or predictions. Efficiently encoding or compressing the data for network usage could also be a patentable aspect.

On the output side, although the predictions might come from a standard model, the post-processing steps that follow could be novel and valuable. How you handle and represent the output data, such as probability distributions, classifications, or object detections, could be a unique and patent-worthy approach. The way you communicate or encode this output information for efficient network transmission may also be an innovative contribution.

Moreover, the patentable aspect could extend to the actions taken based on the model’s predictions. Maybe your post-processing leads to further operations or actions, and the combination of the output and subsequent processing could be considered a novel invention.

Conclusion

As a relatively new field, there are many opportunities to carve out territory in AI. It is likely that many seemingly innocuous improvements may lead to important breakthroughs in the technology so it is important to ensure that you protect whatever ideas you can. If you are unsure, a short consultation with a patent professional could help determine if your idea fulfills the basic requirements for patentability.

Explore other articles in this series

2023

Insights

Intellectual Property

Intellectual Property and Generative AI: Many Questions, Few Answers

2023

Insights

Intellectual Property

Ownership of IP: Do You Own Your Invention?

2023

Insights

Intellectual Property

Patent Searching: How to Find Out if You Really Invented Something