Artificial Intelligence has changed the world in ways unimaginable. From reducing human labor and time – to making hazardous tasks seem like a piece of cake; AI has surely disrupted the general scheme of things. Today, there is probably no sphere wherein AI does not find any application. Be it Amazon’s Alexa or Google’s Siri, AI has transformed how things get done. With the growth in applications of AI, IP system encourages innovations in the area of AI that are created by humans. The future holds potential for Artificial Intelligence, and it is definitely going to scale new heights. Which is why, one must grab every chance to patent AI inventions.

This article discusses the underlying technology behind the workings of Artificial Intelligence, its algorithms, and different types. It also discusses the need to patent AI technologies, the AI trend and how describing technical effects of the AI invention can lead to AI patent grant.

Types of AI

Artificial Intelligence can be broadly categorized into two types – based on capabilities and the other based on functionality.

Artificial Intelligence type 1 – Based on Capabilities

Based on Capabilities, AI can be further subdivided into:

1. Narrow or weak AI: Narrow or Weak Artificial Intelligence encompasses almost all the existing AIs. It refers to such AI machines that perform a specific task autonomously. Such AI systems have limited or narrow competencies. Narrow or Weak AIs find application in a number of fields. Many smartphone applications include features such as face recognition, voice-based personal assistants, a search tool, etc. Other applications include vision recognition in self-driving vehicles, recommendation engines suggesting products based on purchase history, or the like. These features are goal-oriented and singular tasks that have narrow parameters and contexts. The narrow AI does not mimic or replicate human intelligence and merely simulates human behaviour on a narrow range of parameters and contexts.

2. General or strong AI: It is a concept of a machine with general intelligence that mimics human intelligence and/or human behaviour, with the ability to learn and apply its intelligence to solve any problem. This strong AI uses a theory of mind AI framework, which refers to the ability to discern needs, emotions, beliefs, and thought processes of other intelligent entitles. The theory of mind-level AI is not about replication or simulation but about training machines to truly understand humans.

One of the most notable attempts at achieving strong AI is Fujitsu-built “K-computer.” However, considering it took 40 minutes to simulate a single second of a neural activity, it is difficult to determine whether or not the strong AI will be achieved in the foreseeable future. As image and facial recognition technology advances, there may be an improvement in the ability of machines to learn and see.

3. Artificial superintelligence: Artificial superintelligence is a hypothetical AI that does not just mimic or understand human intelligence and behaviour. It is where machines become self-aware and surpass the capacity of human intelligence and ability. Artificial superintelligence, while able to replicate the capabilities of human beings, would be exceedingly better because of its overwhelmingly greater memory. 

Artificial Intelligence type 2 – Based on Functionality

1. Reactive Machines: Mostly basic and rather primitive, these AI systems have limited capabilities. These machines do not have memory storage, and they try to react to the present stimuli as per possible best action. This also means that they do not have the capability to learn. They cannot “learn” from their past experiences and therefore cannot use them to better their present actions. These machines can only be trusted for responding to a limited set of instructions. They cannot ‘improve’ their operations based on their memory. One famous example of such an AI machine is IBM’s Deep Blue, the one that beat Grandmaster Garry Kasparov in Chess.

2. Limited Memory: Limited memory machines are capable of learning from past experiences and storing data but for a very short duration of time. These can be thought of as improvements over reactive AI machines. In addition to functioning as a reactive machine, it is also capable of learning from its past. It uses its past memory and volumes of data to train itself for solving future problems.

3. Theory of Mind: Although researchers are still engaged in innovating, this new type of machine is believed to have “human-like” emotions, beliefs and would be able to behave just like humans in society. While the previous-mentioned machines are a reality, this type of AI system is still a concept. It is the next level of AI system that is still a ‘work under progress.’

4. Self-Aware: These types of machines exist only hypothetically. Believed to be the future of Artificial Intelligence, these super-intelligent machines will have their own consciousness and would be extremely “self-aware,” just like Human Beings or even better.

Agents and environments

Agent and Environment are two significant portions of any AI. Before we begin to understand how an AI works, it is crucial to grasp an understanding of “an agent and an environment.” An agent is something that can perceive while an environment is what it perceives. An agent perceives the environment through sensors and acts upon it through effectors. Examples of the agent include a program, a chatbot, a robot, etc. A ‘human agent’ has its sensory organs as ‘sensors’ and organs such as legs and hands as effectors. Similarly, a ‘robotic agent’ has cameras and infrared sensors that are parallel to sensors, while it has motors and actuators for effectors.

The agent and environment vary depending on application areas. For instance, in the case of a robot in a room, the robot is the agent while the room is the environment. Similarly, in the case of a vehicle on the road, the vehicle is the agent while the latter is the environment. Likewise, in the case of a software program, the program is the agent while the data and rules are the environment.

Thus, the agent is basically a solution for solving a problem, i.e., the environment. In order to work in the environment, the agent needs intelligence, which is where the AI comes into the picture. The AI helps in designing an “agent program” for the agent. The agent program runs on a computation device that is equipped with sensors and actuators. This agent program with the architecture that comprises the sensors and the actuators contributes a structure to the agent. Architecture is the machinery on which execution is done by the agent, while an agent program is an implementation of an agent function.

That is,

Agent = Architecture + agent program

An example of the agent program is the “autonomous driving” program for “autonomous vehicle” equipped with computers, sensors, or the like, which is the architecture. The sensors in the architecture perceive ‘sensor data’ from the environment. The sensor data is inputted to the agent program for returning “actions” that are performed via the actuators.

Based on the interaction between the agent and the environment and its applicability, agents can be categorized into different types:

1. Simple-reflex agent
2. Model-based reflex agent
3. Goal-based agent
4. Utility-based agent
5.  Learning agent

1. The simple-reflex agent works on the condition-action rule. It receives percepts from the environment via the sensors and performs actions in a fully observable environment based on the said rule. “Percepts” are inputs that an intelligent agent is receiving at any moment. The condition-action rule refers to the fulfillment of action in response to a condition. For example, a room cleaner agent works only where there is dirt in a room. In other words, for a room-cleaner to perform its ‘action,’ there should be dirt in the room, which is the necessary ‘condition.’

The simple-reflex agent is rather simplistic in nature. It has limited intelligence with no capability to adapt to changes in the environment. This inefficiency may be overcome by the model-based reflex agent.

2. The model-based reflex agent uses a model (knowledge about the environment) to perform actions in a partially observable environment. This model is the information about “how are things in the world right now.” It also maintains an internal state that represents the current situation or current state based on a percept history. The percept history is a history of all that the agent has perceived to date. The state is updated based on information of how the world evolves and how the agent’s actions affect the world.

There may be a time when the knowledge of the current state of the environment may not always be sufficient for the agent to decide what actions are to be performed unless there is a goal to achieve. The capabilities of the model-based agent can be further expanded to the goal-based agent.

3. A goal-based agent is comparatively more flexible than reflex-based agents or model-based agents: the reason being an explicit modeling of the knowledge supporting a decision. A goal describes a set of desirous situations. The goal-based agent knows its goal that describes desirable situations and has the capability to select an action for achieving the goal. For this, the goal-based agent performs searching and planning to select the action from a long sequence of possible actions. The selected action that the goal-based agent performs is intended to reduce its distance from the goal, which makes the agent proactive.

Sometimes, the goal achieved by the goal-based agent may not be the best as the world is uncertain. So, a measure of success for achieving the goal may be helpful to determine the best way to achieve a goal.

4. The utility-based agent is similar to that of the goal-based agent with a capability for utility measurement. The utility measurement gives a measure of success at a given state for achieving the goal. That is, the utility-based agent has the capability to select actions based on the utility for each state such that the expected utility is maximized. In particular, the utility describes how “happy” the agent is that may be determined using a utility function. The utility function maps a state onto a real number which describes the associated degree of happiness. 

The environment is not only uncertain, but it can be dynamic too. As the environment expands, amount of tasks to be performed by the agent may also increase. This may require the agent to learn new actions so as to work in environments that are unknown, which results in another type of agent, i.e., the learning agent.

5. The learning agent has learning capabilities to learn from past experiences and act with basic knowledge. Eventually, the learning agent builds the ability to act and adapt automatically through learning. In particular, the learning agent has a learning element to learn from the environment, a critic component to take feedback from critics that describes how well the agent is performing based on a predefined performance standard, a performance element to select an action and a problem generator to suggest actions that will lead to new and informative experiences.

The learning agent, therefore, learns, analyzes performance, and looks for new ways to improve.

The Algorithm

In order to fully understand what runs an AI machine, it is imperative to know an AI’s algorithm. Simply put, algorithms are step-by-step instructions that a machine or system needs in order to perform the desired function. One should think of it as an instruction manual that lets a machine know exactly what to do and when to do it. Artificial intelligence is powered by algorithms, using techniques such as machine learning, deep learning and data science.

Machine learning (ML) algorithms feed data to an AI system using statistical techniques. The statistical techniques enable the AI system to learn and progressively improve at performing tasks without having to be specifically programmed to do so. Programming concepts, such as functional programming, attribute-based programming, and object-oriented programming may be used to program an application to react to every input and output, which has downsides. In a real-world scenario, there are too many inputs and outputs that may not be predictable. The predictability may be achieved through machine learning algorithms.

A large amount of data is fed to machine learning algorithms as input data. Initially, the input data is pre-processed, i.e., cleaned to remove unwanted or redundant data and achieve an accurate outcome. The Machine learning algorithms crunch the input data and produce a model that has “rules” created based on the input data. Further, the validity of the model may be tested. For instance, if there is 100% of the input data, then the data is split into, say 70% of data for creating the model and 30% of test data for testing that evaluating the model. A model based on the 70% data is created. The 30% of test data is then played against the model, predicting values and determining how accurate the predictions are. Furthermore, the data may be randomized to a new subset of data for improving the results. All these steps of pre-processing data (structured or semi-structured), selecting appropriate machine learning algorithms, and creating right parameters for creating the most accurate set of data, are achieved through data science techniques.

The input data fed to the ML algorithms may be of different modalities, such as: image, video, audio, etc. The different modality of data may be unstructured and unlabelled that are processed and analysed using deep learning techniques, without the need for human labelling. This can be achieved using neural networks, such as artificial neural networks that mimic how human brain works. The neural networks help in finding patterns in data, recognizing the relationship between variables, clustering, and classification.

Nowadays, deep learning and neural networks are at the forefront of AI research and may hold the key to a future where AI can function at levels that are unheard of today. With the rise in AI applications, patenting AI-related applications has also increased.

Patenting AI inventions

AI invention patents are needed to:

• protect technical innovations and technical solutions to problems
• Prevent others from exploiting an inventor’s AI invention
• help the inventor mark out legal exclusivity around a patented product
• earn a good return on investment through licensing of the patented product

Challenges in protecting AI inventions

• Most of the AI-related applications, such as the core AI involve advanced mathematics and mathematical models. This challenges eligibility for patent protection of the AI inventions.
• In AI inventions that are related to the trained AI models, claiming variations and ranges may be challenging. There may be objections if buzzwords and marketing terms are used for the trained AI models in claims.
• Lack of technical information or insufficient of technical information to define or describe technical effects of the AI invention

Strategy to overcome the challenges

Most of the AI inventions face a patentability issue due to use of computer systems and algorithms. Also, law surrounding subject matter eligibility rapidly evolves which adds the factor of patentability issue.

In January 2019, the USPTO issued “Revised Patent Subject Matter Eligibility Guidance” to address the changes in law and stem the many patent application rejections. In October 2019, an updated of the revised subject matter eligibility guidance “Patent Eligibility Guidance Update” was released. In the updated revised subject matter eligibility guidance, “examples” are included that help in formulating a strategy for overcoming the challenges. Since the release of January 2019 guidelines, rejections of AI related patent applications have dropped from 60% to about 32%.

Making AI invention functionality practical

The AI invention directed to method should have technical means, like a machine, such as a computer or a device to add a technical character to its subject matter as a whole and prevent the AI invention from being an “abstract” that may exclude from patentability.

For example: if the AI invention is to improve a technique for analysing data such as, improves K-nearest neighbours (kNN) classification with no advantage to a technical field, then the AI invention may face rejection. However, if the AI invention applies kNN classification to achieve a specific improvement to a specific technical system then chances of obtaining a patent protecting the AI invention increases.

Thus, a complete description of technical features describing technical effects of the AI invention frames the AI invention in a way to convince patent offices to grant a patent.

Rising Trend for patenting AI inventions

In a U.S. Patent and Trademark Office (USPTO) report titled “Inventing AI: Tracing the diffusion of AI with U.S patents”, it is mentioned that AI technologies and systems as defined by the U.S. National Institute of Standards and Technology (NIST) “comprise software and/or hardware that can learn to solve complex problems, make predictions or undertake tasks that require human-like sensing (such as vision, speech, and touch), perception, cognition, planning, learning, communication, or physical action.”

There is a shift in the patenting behaviour from theoretical to practical applications. The trend of AI related inventions shifting from theory to practical implementation can be seen from the WIPO Technology Trends 2019 – Artificial Intelligence report. The report states that

• 40% of AI-related patents correspond to machine learning technique
• 49% of AI-related patents correspond to computer vision
• 42% of AI-related patents correspond are used in telecom, transportation or life sciences.

Conclusion

Intelligence is perceived as the “capacity to think and solve problems.” This capacity, which used to be the sole prerogative of humans is not limited to us – as intelligence can also be demonstrated by machines or computers, also popularly known as “Artificial Intelligence” or “AI”. With advanced development in technology, AI is increasingly driving important developments across many fields and industries. In most of the applications, AI is used to assist, reduce, or eliminate human involvement in existing processes. Some of the popular applications of AI are autonomous vehicles, advanced manufacturing processes, medical diagnostic tools, e-commerce, etc. It is making an impact on the scale and speed of innovation in almost all areas.

Sagacious IP, a leader in global IP research and consulting, has helped multiple clients secure AI patents. We have an adept team of over 300 professionals who specialize in various domains. Click here to know more. At Sagacious, IP professionals have expert skills for drafting computer-implemented inventions involving AI-related application. We ensure that the drafts for such computer-implemented inventions involving AI-related application include:

• Technical solution to a technical problem
• Clear identification of technical advantageous features and why they are advantageous.
• A disclosure of at least one set of training data and details of how it has been trained
• Illustration through examples or use cases that describe how the technical problem is solved.

– Oinam Binarani Devi (ICT Drafting) and the Editorial Team

Having Queries? Contact Us Now!