Even though the definition of Artificial general intelligence (AGI) is a bit vague, mostly, it is argued to be a machine intelligence that performs cognitive tasks such as language, learning, decision making, perception at the human level. Most of us are very familiar with it from science fiction movies such as The Matrix, 2001: A Space Odyssey. Even though we do not have such intelligence yet, I believe that we will probably start to have the first instances soon! However, are the current technique called neural networks is the answer to get AGI?
I think that even though neural networks are extremely powerful, it is not enough for us to get us to AGI. However, the neural networks with some structure and improvements might be the answer. For example, classical artificial intelligence (AI) was generally implemented symbolically without any learning. However, it focused on humans’ very fundamental aspects, such as perception, and provided one of the critical steps to form perception for a machine. Then, we improved it by creating distributed representations at different levels of abstractions and developed techniques like neural networks and deep learning. In other words, we made some improvements in structure. Likewise, currently on our way to AGI, we are building on top of deep learning concepts to solve problems such as reasoning, planning, or even higher level of cognitive tasks like consciousness.
Biological brains are nothing but neural networks functioning in some structures or architectures. Similarly, to get human-level intelligence with higher-level cognitive tasks, we also need to develop different types of neural network architectures. For example, when we put some structure in the networks, we get properties like attention and prior knowledge that resulted in massive success in the machine translation field. Another example comes from the convolutional neural network structure representing different features of an image at different levels in network architecture to perceive an image more efficiently.
Consequently, what we are doing/aiming for now is, indeed, a work of evolution that is hardwiring the neural networks with some structures and computational improvements to get human-level efficiency in machine intelligence. In my next post, I will address how neuroscience can inspire such a process.