What is so special about neural networks?

The case for neural networks in general as an approach to AI is based on a similar argument as that for logic-based approaches. In the latter case, it was thought that in order to achieve human-level intelligence, we need to simulate higher-level thought processes, and in particular, manipulation of symbols representing certain concrete or abstract concepts using logical rules.

The argument for neural networks is that by simulating the lower-level, “subsymbolic” data processing on the level of neurons and neural networks, intelligence will emerge. This all sounds very reasonable but keep in mind that in order to build flying machines, we don’t build airplanes that flap their wings, or that are made of bones, muscle, and feather. Likewise, in artificial neural networks, the internal mechanism of the neurons is usually ignored and the artificial neurons are often much simpler than their natural counterparts. The electro-chemical signaling mechanisms between natural neurons are also mostly ignored in artificial models when the goal is to build AI systems rather than to simulate biological systems.

Compared to how computers traditionally work, neural networks have certain special features:

Neural network key feature 1

For one, in a traditional computer, information is processed in a central processor (aptly named the central processing unit, or CPU for short) which can only focus on doing one thing at a time. The CPU can retrieve data to be processed from the computer’s memory, and store the result in the memory. Thus, data storage and processing are handled by two separate components of the computer: the memory and the CPU. In neural networks, the system consists of a large number of neurons, each of which can process information on its own so that instead of having a CPU process each piece of information one after the other, the neurons process vast amounts of information simultaneously.

Neural network key feature 2

The second difference is that data storage (memory) and processing isn’t separated like in traditional computers. The neurons both store and process information so that there is no need to retrieve data from the memory for processing. The data can be stored short term in the neurons themselves (they either fire or not at any given time) or for longer term storage, in the connections between the neurons – their so called weights, which we will discuss below.

Because of these two differences, neural networks and traditional computers are suited for somewhat different tasks. Even though it is entirely possible to simulate neural networks in traditional computers, which was the way they were used for a long time, their maximum capacity is achieved only when we use special hardware (computer devices) that can process many pieces of information at the same time. This is called parallel processing. Incidentally, graphics processors (or graphics processing units, GPUs) have this capability and they have become a cost-effective solution for running massive deep learning methods.