Common sense can make AI smarter

Artificial intelligence (AI)-based solutions are here to stay. A study by Lambda3, a software development company focused on agile concepts and methodologies, reveals that 80% of people in Brazil have already used such technologies at least once. The company’s Chief Data Officer Diego Nogare points out that the remaining 20% ​​refused the application due to lack of information, as they declared that they communicate with applications such as Waze, Google Maps, Facebook and Instagram.

A survey by Ipsos, a research and market intelligence company, World Economic ForumIt showed that 57% of Brazilians believe that services and products with artificial intelligence offer advantages rather than disadvantages.

Companies are aware of this move, as 41% claim they are already using this category of resources in business activities. 2022 Global AI Adoption Indexpowered by IBM’s Morning Consult program; already 34% say they are investigating use.

Much of the success of AI can be explained by advances in the industry. This is because, besides existing in more subtle ways in everyday life (like recommendations from mobile and computer operating systems, chatbots and streaming services, social networks and the like), it stands out in more applications for optimizing the lifespan of electric vehicle batteries and even the film industry. advanced scenarios.

Still, scientists continue to devote themselves to improving AI solutions to make them more efficient at performing the tasks for which they were built – and the answer to smarter AIs may lie in their ability to develop common sense, which is considered a kind of “dark”. area is a matter.

Humanity doesn’t just live by the rules

Artificial intelligence is the discipline that aims to allow machines and computational applications to emulate human intelligence through technology, as described in a paper published by Colorado State University Global. Systems based on it can answer questions, predict situations, and offer strategic solutions through learning through iterative processing (repeated execution of a set of instructions and validation based on results) and experience based on algorithmic training.

Essentially, an AI system works from large datasets by identifying patterns and functions aimed at solving complex problems. While some tools currently convert everything to lyrics and produce all sorts of images, it turns out that one of the limitations to current learning models is the lack of common sense they offer, even if there are claims to be a responsive prototype out there. .

For now, robots act based on algorithms, which are simply commands. On the other hand, common sense, unlike intelligence, is something inherent and natural to man, a skill that cannot be taught.

Because the goal of AI research is to effectively apply problems to real-world situations that are unpredictable, uncertain, and undefined, these science-based solutions always face insurmountable problems, at least at their current stage. obstacles.

Why is common sense important?

When equipped with common sense, machines can cross the boundaries of pre-programmed interactions based on offering alternatives and immediate response, providing better assistance to the consumer as well as guaranteeing greater safety for autonomous cars to operate in unforeseen events on highways. example . However, the implementation of such a talent is not a simple task.

Since 1958, there have been few developments regarding the implementation of the proposal. That year, computer science pioneer John McCarthy published the paper. common sense programs (in free translation) analyzes how logic can be used as a method of representing information in computational memory. The challenge to date is that common sense is not a definitive property, so it is full of hidden variables.

As Michael Witbrock, an AI researcher at the University of Auckland (New Zealand), argues, common sense includes not only social skills and reasoning, but also a “pure understanding of physics.” With common sense, people have a certain knowledge of Physics (like when they avoid approaching a cliff on a windy day) without needing to solve equations. In this sense, it allows planning, forecasting and organizations by taking into account the basic knowledge of abstract elements such as time and space.

Therefore, before considering launching an AI that can infer things based on real situations – an action performed by humans since childhood – it is necessary to evaluate what would actually characterize such a capacity and how close these technologies are to conquering it. . Scientists from IBM, Massachusetts Institute of Technology (MIT), and Harvard University have created a test for just this kind of assessment: AGENT, a promising diagnostic tool.

AGENT: making AI smarter

The acronym AGENT contemplates English terms for action, purpose, efficiency, limitation, and utility, and refers to a large-scale 3D animation dataset inspired by experiments examining cognitive development in children, a set of aiming educational machines.

According to IBM, “videos consist of separate trials (each containing one or more ‘break-in’ videos’ of typical agent behavior in a particular physical environment) paired with ‘test’ videos of the same agent’s behavior in a new environment. Consider the agent’s behavior in the corresponding practice videos. “expected” or “surprising” when received.

For the benchmark, artificial intelligence learning models need to categorize the test videos they feed on as they absorb familiarity-oriented content.

This is a confirmation of the analytical potential of the machine in comparison with the results of large-scale human evaluation tests; success results in the possibility of giving AIs a spectrum of social awareness that enables them to interact with humans in real settings.

This experiment, called AGENT, suggests that integrated representations of how humans plan things are essential for robots to develop a kind of common sense. In this sense, it is necessary to combine the basic understanding of Physics with “cost-benefit based trade-offs”; This means that AIs can ensure an understanding of how humans act “based on utility,” balancing the rewards of their goals with the costs. reach him”.

In the future, IBM believes diagnostics like this can significantly reduce the need to train learning models and enable companies to reduce computing demand, time and money, as well as smarter AI with less effort.