Approaches to Artificial General Intelligence
Artificial General Intelligence, or AGI, is the goal of many, many people, including companies like Google and Amazon, to create a system that has human-level intelligence or higher. AGI will revolutionize the world when (if) it is ever created. Revolutionize, even, is a bit of an understatement. AGI is the dreams of Hollywood, the nightmares of The Terminator, and a tool that will fundamentally change the way we look at the world and the way we look at ourselves. If something exists that is as smart as us, that also has intelligence –that special something that only humans have been able to lay claim to for the past couple millennia — where does our purpose lie?
But AGI is still fairly far off in the future, and an agent that ever possesses AGI needs to meet many, many criteria. These include things like the capability for creativity, perception, memory, learning, and motivation. But to even meet these criteria, this supposed artificial generally intelligent agent needs to have an underlying neural architecture, the foundation for the rest of these abilities. Our brain is a good (the only) example of an architecture that possesses actual intelligence. In the course of the pursuit of AGI, there have been many approaches to the creation of this architecture, and this article aims to cover the core approaches.
*Note: AGI is not a formalized field, in that there is no central dogma or theory. The approaches listed in this article are certainly not all the approaches, and are influenced by my bias.
The Symbolic Approach
The symbolic approach basically utilizes logic networks (if-then statements, etc.) and symbols to learn and increase its knowledge base by manipulating its symbols. Proponents of the symbolic approach argue that the mind exists to manipulate symbols that represent actual aspects of the physical world or themselves. It closely resembles the higher-levels in thinking of the human brain.
In theory, the symbolic approach sounds very nice: symbolic systems have a large representational power and can perform high-level logic and thinking. In reality, most symbolic approaches falls short in learning and lower-level tasks like perception.
The most notable example of the symbolic approach is CYC, started in the 80s, which possesses a huge knowledge base, a powerful logic system (using what it calls inference engines) and an expressive representational language.
The Connectionist Approach
The connectionist approach, also known as the emergentist or sub-symbolic approach, aims to create general intelligence from architectures that resemble the brain, like neural nets. Connectionists expect that higher-level, abstract reasoning will emerge from lower-level, sub-symbolic systems, like neural nets, which has, so far, not happened.
Most of the AI that we are familiar with today is are connectionist “architectures” despite the fact that their purpose is not to create AGI. DeepMind’s AlphaGo, standard convolutional neural networks, and other deep learning systems are good examples.
Yann LeCun, the Chief AI Scientist at Facebook, believes that from pure connectionist architectures, complex behavior and thinking will emerge.
The Hybrid Approach
The hybrid approach, as the name suggests, utilizes both the connectionist and symbolic approaches. Many of the architectures at the forefront of the field utilize the hybrid approach. The CogPrime architecture uses a knowledge representation system known as the AtomSpace to represent both symbolic and sub-symbolic knowledge along with many other systems that act as cognitive processes. CogPrime was used as the neural architecture for Sophia, created by Hanson Robotics and OpenCog: the most expressive robot that exists right now.
Whole-Organism Architecture
There are many people who believe that a true artificial generally intelligent system needs to have a physical body and be able to physically learn and interact in the world. There are no real examples of AI that truly learn through physical interactions and sensory data, but the closest we have right now is Sophia.
The field of AGI contains many more architectures than those listed in this article, which may not follow all of these approaches. Maybe one of them will result in an actual AGI system. If you want to dive deeper into the field of AGI, I recommend going through these sources:
Thanks so much for reading! If you want to talk more about AGI or your work, feel free to contact me at kevn.wanf@gmail.com or through my LinkedIn.
