Chapter One why are we interested in machine learning? Nine

Understanding the Brain

According to MARR1 (1982), Understanding the information processing system requires three layers of analysis:

1. computational theory is related to the abstract definition of computational goals and tasks
2. representations and algorithms are detailed descriptions of how inputs and outputs are displayed, and the algorithms that are converted from input to output.
3. Hardware Implementation is the actual physical implementation of the system.

The basic idea of multilayer analysis is that for the same computational theory, there may be multiple representations and multiple algorithms used to manipulate the symbols in the representation. Similarly, there may be a variety of hardware implementations for any given representation and algorithm. For either theory, we can use one of several algorithms, and the same algorithm can have different hardware implementations.

For example: ' 6 ', ' VI ', ' 110 ' are three different representations of the number 6, namely Arabic numerals, Roman alphabet, and binary notation. Different representations are used, and the addition algorithm is not the same. Digital computers use binary notation and have circuits that add to this notation, which is a special hardware implementation. In the abacus, the use of different methods to represent numbers, and the use of different instruction sets for addition operations, this is another kind of hardware implementation. When we add the two numbers in mind, we use another representation method and its corresponding algorithm, which is implemented by neurons-that is, we, the Abacus and the digital computer-Achieve the same computational theory: addition.

A typical example is the difference between a natural vehicle and a man-made aircraft. The sparrow beats its wings and the plane does not flap its wings but uses a jet engine. Sparrows and airplanes are two hardware implementations that are constructed for different purposes to accommodate different constraints. But they all realized the same theory, that aerodynamics.

From this point of view, we can say that the brain is a hardware implementation of learning. If from this particular implementation, we can do reverse engineering and extract the use of the representation and algorithm, if we can get the computational theory, then we can use another representation method and algorithm, and so on, using more suitable for the method and the limitations of the hardware implementation. We want our implementations to be cheaper, faster and more accurate.

Just as the aerodynamic theory was discovered, the aircraft we initially tried to build looked more like birds, and it was expected that the first attempt to construct a brain-capable structure would look like a brain with a large network of processing units. In fact, in the 4th chapter, we will discuss the artificial neural network composed of interconnected processing units, and how such networks are to be learned-this is the representation and algorithm layer. Finally, when we discover the intelligent computational theory, we can find that neurons and highlighting are implementation details, just as we realize that feathers are used to fly.

MARR1, founder of the Theory of visual computing, D.marr was originally from England. He arrived in the United States in 1973 and was invited to create and lead a research group with PhD students in the MIT AI Lab, focusing on the exploration of visual theory. The theory of Visual Computing was published in 1982.

Chapter One why are we interested in machine learning? Nine