Concepts is All You Need, a path to AGI

In the fascinating world of Artificial Intelligence (AI), we’ve seen some incredible breakthroughs over the last two decades. However, the elusive goal of achieving Artificial General Intelligence (AGI), a form of AI that can perform any cognitive task at a level comparable to a human being, remains unmet. Today, we’re diving into a compelling paper by Peter Voss and Mlađan Jovanović, titled “Concepts is All You Need: A More Direct Path to AGI,” which suggests that the key to AGI might be closer than we think. 

The authors argue that the path to AGI lies in understanding and identifying the core requirements of human-like intelligence. They believe that the current focus on statistical and generative efforts in AI development is somewhat misplaced. Instead, they propose a Cognitive AI approach, which places the central role of concepts in human-like cognition at the forefront.

In their paper, The authors outline an innovative architecture and development plan that could pave the way to full Human-Level AI (HLAI)/ AGI. They emphasize the importance of AGI’s ability to learn new knowledge and skills, interpret and encode knowledge conceptually, learn complex data interactively, and adapt to new situations and environments.

Interestingly, the authors  suggest that the Cognitive AI approach isn’t merely a combination of the first two ‘Waves’ of AI (Rule-based and Statistical learning). Instead, it’s a fresh approach implemented as a cognitive architecture, designed to encompass all the essential structures required for a human-level mind.

One of the intriguing points in the paper is the discussion about the role of senses and actuators in AGI. The authors suggest that effective AGI can be constructed with separate pre-processing mechanisms for visual, tactile, and sound input. They also highlight the importance of vector representation in cognitive functions and the formation of highly abstract concepts.

The authors  propose a graph-based vector datastore as a foundational substrate for all cognitive functions. This vector graph serves as both long-term and short-term memory, providing a fully integrated and high-performance system.

Another fascinating aspect of the paper is the discussion on metacognition and emotions in AGI. The authors argue that what sets human intelligence apart is our ability to think about, and direct our thinking. This ability, along with cognitive emotions such as surprise, certainty, confusion, and boredom, should be an integral part of any workable AGI design.

Towards the end of the paper, The authors provide a roadmap to AGI, which includes activities such as re-integrating multimodal vector pattern learning and matching into the knowledge graph, adding real-time and background abstraction/concept formation, training the system to do basic question-answering, and developing advanced language capabilities.

In conclusion, The authors believe that the slow progress towards AGI is due to a lack of focus on what human-like cognition really requires. They propose a Cognitive AI approach, specifically highlighting the need for effective conceptual knowledge representation, as a more direct path to AGI. This paper is a must-read for anyone interested in the future of AI and the exciting journey towards AGI.

Paper can be found : [2309.01622] Concepts is All You Need: A More Direct Path to AGI (arxiv.org)