AI Common Sense: The Dark Matter Of Intelligence

The term ‘dark matter’ was coined from the mysterious, unidentified cosmic matter found in space, known primarily because of its gravitational effect. Similarly, dark matter in AI refers to poorly understood, unraveled subsets of intelligence. It represents gray areas that limit the ability of AI systems to replicate human recognition, evaluation, decision-making, and reasoning. The introduction of AI meant it hit the ground running, tasked with different sections of intelligence to prove its effectiveness and usefulness. From small analytic tasks demanding little intelligence to problem-solving on bigger scales. 

AI Learning Techniques

One of the greatest breakthroughs of the modern age is AI, programmed systems that learn from large structured data to perform tasks, solve problems, and ultimately ease human activities. AI systems are developed through learning techniques (machine learning) like supervised, unsupervised, semi-supervised, and reinforcement learning, allowing the algorithm to learn from feedback.

Machine learning encompasses supervised learning, which trains the algorithms on labeled datasets; it involves teaching the model by example. Unsupervised learning identifies patterns in data without labels and aims to pinpoint the present structure of the data by grouping it. Semi-supervised learning combines labeled and unlabeled data for improved performance. Lastly, reinforcement learning, in which systems learn by interacting with their environment and receiving feedback. Specifically, these learning processes train the AI system through data (labeled and unlabeled) so it can recognize, analyze, interpret, and adapt to any task. But this learning is limited; why? AI Dark Matter.

For example, in an article published in Quanta Magazine, Gary Marcus, an AI researcher, explained the limitations of AI systems in a prompt he made. He asked GPT-2, ’What happens when you stack kindling and logs in a fireplace and then drop some matches is that you typically start a…’. GPT-2 responded with ‘ick,’ showcasing its lack of intuitive common knowledge of the world that reduces inference. 

AI Dark Matter

Artificial intelligence is inspired by biological intelligence. This intelligence develops its capacity for cognitive reasoning and problem-solving from observation, analysis of patterns, and intuitive inference. additionally, the foundation of machine learning is grounded in neural networks, a set of interconnected networks that mimic the neurons in the human brain. The networks enable AI models to observe patterns, analyse big and small data, make predictions, and give feedback based on reinforcements.

However, humans stand out because of their unique ability to adapt using their generalized knowledge and intuition—some call it common sense. However, this unique ability has been difficult to replicate in AI systems. Humans and animals gather knowledge from observation of different sources and experiences, leading to intuition in vast scenarios. It makes learning seamless and reduces the need for supervised learning. 

Is there hope? 

Remember the matches and log example I cited earlier? I tried it with my GPT-4, and it responded, “Fire, as long as the kindling catches and spreads to the logs.” Does this mean it has unraveled the puzzle? I don’t think so; it probably learned from feedback. But is there any hope? 

According to Meta AI, “We think that one of the most promising ways to build such background knowledge and approximate a form of common sense in AI systems is self-supervised learning (SSL).”

Yes, over several years of research, scientists have developed what seems like the missing jigsaw—self-supervised learning. SSL is a training paradigm that allows AI to learn from unlabeled (raw) data by itself without the need for human interpretation. Moreover, it labels data, observes patterns, and makes judgments on more generalized knowledge. Self-supervised learning offers hope for greater intelligence, as AI can mimic how humans infer meaning from raw experiences. Currently, it is pushing boundaries in:

• Natural language processing and computer vision allow even higher functions in translations and interpretations
• Predictive intelligence with the advent of pretext tasks for representation learning
• Medicine, reducing cost and allowing more precise diagnosis. For more on how AI is revolutionizing medicine, read this.
• Generalized knowledge allows AI to learn from accumulated experience and develop perception and intuition 

There is still room for immense growth as continuous research is ongoing in this field of knowledge. Before the end of the decade, scientists believe AI will have developed comprehensive emotional intelligence to mitigate its present gap. Let’s just hope AI does not learn from our biases, observing unfair and unjust patterns.