“
Let me use a simple everyday problem to explain. Suppose I were to try to determine what type of coffee you like. I do not have that information stored in my extensive database, so I would need to learn this fact. I would start with a prior understanding of the coffee preferences of people of your approximate age, gender, and general physiology. Then, I would brew you a coffee. I would analyze your reaction to drinking the beverage. In the short-term, your facial expressions, noises you make, your level of happiness, excitation, etcetera. After this, I would analyze your sleep patterns, elimination rates, urine production, bowel movements, and general health levels. I would interpret this information in the light of other data that may confound my assessment of your coffee preference, such as what events had occurred in the day that may affect your general levels of happiness, levels of trauma, your concomitant food intake, your emotional status. Using each piece of information, I would update my prior understanding, generating a posterior probabilistic model of your individual coffee preference, adjusting for any confounding factors, of course. This posterior model becomes my new prior understanding of your particular coffee predilections. Each day, I could then adjust your coffee: hotter, colder, stronger, weaker, sweeter, more milk, less milk, etcetera — knowing, of course, that many of these factors are correlated. For instance, people who like sweeter coffee generally prefer more milk and vice versa. Then, I would assess your reaction to the new formulation and adjust my probabilistic model as more data are collected, thereby maximizing my understanding of your coffee preference. As you can appreciate, this is simply a rudimentary and idealized example of my learning process.
”
”