1. What Does Understanding for an AI System Entail?
Herbert Simon (1977) suggests that understanding involves a symbiotic relationship between a system, knowledge, and tasks. Over the years, we can liken AI's journey toward comprehension to a musical fugue—a two-part harmony:
Early AI systems primarily grasped the syntax of languages, using this knowledge to encode and store information presented in natural language, responding to queries with apparent understanding.
A second kind of knowledge emerged: inference rules, enabling systems to derive implicit information and expand their capabilities.
The third kind introduces intensional definitions of objects and relations for real-world or simulated environments.
The fourth kind delves into understanding the reasons behind actions, retaining knowledge about goal hierarchies to answer "why" questions.
The fifth kind brings algorithms for problem-solving into play, with success measured by efficiency and explanatory power.
Finally, the sixth kind empowers systems to create representations and operators for new, unfamiliar problems—a genuine test of understanding.
2. How Can We Make AI Systems Understand the Real World?
To reach human-level learning and thinking, AI must extend beyond current paradigms. Lake et al. (2016) advocate for building AI systems that construct causal models of the world, supporting explanation and understanding rather than mere pattern recognition. It involves grounding learning in intuitive theories of physics and psychology and harnessing compositionality and learning-to-learn for rapid knowledge acquisition.
Recent advances in deep learning have propelled AI forward. Reverse engineering human intelligence informs AI, particularly in domains where humans excel. To build more human-like machines, we call for the emphasis on developmental "start-up software", incorporating intuitive physics and psychology as foundational elements. Learning, especially model building, plays a pivotal role. Models, resembling causal structures, underpin our understanding and guide actions.
An intriguing aspect is the interplay between model-based and model-free learning. While the former is essential for planning and adapting to new tasks, the latter offers rapid decision-making. The challenge lies in integrating these approaches effectively.
3. The Future of Human-level AI
The future of neural networks promises a departure from current paradigms. These networks could possess intuitive physics, theory of mind, and causal reasoning abilities, reflecting more human-like learning patterns. Incorporating structure and inductive biases into networks or learning from previous experience could lead to profound changes.
Cognitive science underscores the importance of early inductive biases and rich, theory-like knowledge structures. If we want AI to mimic human learning and thought, it should tackle tasks similar to those human learners face. Integrating psychological insights with deep neural networks, including selective attention, augmented working memory, and experience reply, represents a promising trend.
These cognitive ingredients can address core AI challenges with practical applications, such as scene understanding, autonomous agents, creative design, and more.
According to John McCarthy (2007), in the quest for human-level AI, we must consider the common sense informatic situation—dealing with incomplete knowledge, nonmonotonic reasoning, and context. The path to AI intelligence is intricate, demanding introspection, heuristics, elaboration tolerance and some mental model of the consequences of actions.
In conclusion, AI's journey towards understanding the world mirrors our cognitive development. By embracing these cognitive ingredients and infusing them into AI systems, we can pave the way for machines that genuinely comprehend the world, making AI more relatable, adaptable, and capable of addressing complex real-world challenges.
References:
Please feel free to subscribe to my newsletter, if you are interested, to learn more about my work https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=6891134047146397696
If you have any valuable feedback, thoughts or ideas, please don't hesitate to comment.
Comments