The role of any AI system is to make decisions on behalf of humans; decisions that are either very rudimentary or beyond human capability. To make these decisions, the system is composed of complex transformation functions that map inputs to outputs taking noise and control into account.

Let’s take a simple example of switching on and off a street light depending on the amount of light in the environment. Traditionally there used to be a physical switch, which a human would physically handle whenever light had to be switched on or off. Though this method has no issue, it is such an inefficient use of humans as a resource. It is also unwanted because we do not want systems to be person dependent. What if he fell sick one day? You need to have a backup or live in the dark. 

A simple person-independent solution to this is to have a light sensor that would produce a voltage/current proportional to the amount of ambient light. We would heuristically decide on a value at which to switch on and off and program an electronic switch to do it for us. The transformation function for this system can be represented as two, step functions in hysteresis form as shown below.

Let's assume that the transformation function of the systems that we design can be written in a general form as shown below,

f(.) can be a linear or a non-linear function. To build an AI model, we basically extract some key properties from the data called features, decide on a set of transformation 

functions to map this data, and come up with a model, which is basically the weights (W) or coefficients that gives a specific shape to the function.

Whether Man or Machine handles each of these blocks decides what kind of AI system we are dealing with. There are two more factors along with these three blocks, Data and Domain. Both data and domain are related. It is the knowledge of the domain that brings meaning to the data.

A computer at its core is binary in nature. Any data, whether it is audio, image or text is represented this way. If an audio signal that is 1D, is reformatted into a 2D signal and given to the display as an image, the computer would never complain. It is the human viewer who knows that there is something wrong with it. Irrespective of how much progress AI makes, data and domain are here to stay with humans.

Coming back to the general form of the equation, in the case of heuristic-based AI, a human decides on what features to look for in the data, a human decides the transformation function, and a human also decides on the rules to evaluate the input. A heuristic-based system is the most explainable because everything lies in the hands of a human and he designed it according to his interpretation. This is also the most basic form of AI. Though the computer does the computation, the control lies in the hands of a human.

Heuristic systems are very easy to fool because they learn from human interpretation and not from data. Human interpretation is person dependent and can vary, while a given data always conveys the same information. In the street light example, you could easily trick the system into switching off the light by flashing a torch at the sensor. The decision function in the case of a heuristic-based system is also not optimal.

In traditional Machine Learning based AI systems, a human still decides on what features are to be considered for modeling the system. He also decides on what family of transformation functions to use, but the machine finds the optimal coefficients or that function. In heuristic systems, because humans cannot image multidimensional data, data is usually treated as being uncorrelated to one another, while in ML-based systems a machine finds the optimal rules by taking the entire multidimensional data into account.

In Neural-Network based AI systems, features can still be computed by humans, but the transformation and rules are completely offloaded to the machine.

A Deep Learning based system takes it one step further, it moves the control of all the 3 blocks to the machine.

You can see how AI has progressed over the years from being fully under the control of humans to being mostly controlled by machines now. State-of-the-art deep learning techniques automatically find features in the data, figure out the mapping between input and output, and set their own rules. They also do this in multiple layers, which is where the name comes from and the features in the underlying layers are so abstract that it makes it difficult to come up with a human interpretation for its decisions. Machines work in such high dimensions that even if we want to, it is not humanly possible to do this kind of heavy lifting.

None