In the 1930s, the Swiss psychologist Jean Piaget made an interesting discovery about his children. The children, who were still very young, made rapid learning and sensory progress when they interacted with their environment. They applied their experiential knowledge to new situations in order to understand them.
However, their already established first ideas about the world were initially only very slowly changed or replaced by new information. Today we would say “updated”.
For example, if one of the children was already familiar with the schema of a dog and then encountered a cat for the first time, he or she would also consider it a dog, since this animal walks on four legs and has fur. Only additional experience through more frequent encounters with the cat led to the child being able to distinguish the cat from the dog based on various criteria.
Piaget called this assimilation (knowledge is absorbed through experience) and accommodation (knowledge is changed and replaced).
Assimilation and Failed Accommodation in Humans
Interestingly, the machine learning process is not all that different from the one we go through as children. In fact, comparisons of untrained machines with small children appear quite frequently in scientific publications.
What both have in common is a relatively limited or non-existent wealth of experience. Data on which conclusions could be drawn are initially sparse or non-existent.
For example, a machine with a manageable amount of data would initially classify cats as dogs with great confidence. The only difference is that the machine sees the animals as columns and rows of data and makes the conclusion through calculations and algorithms. With increasing experience, through more data input and additional information about cats, the ability to reliably distinguish the two animals from each other would improve significantly.
Assimilation and failed accommodation of the machine
The learning process of a machine learning model always goes through four steps. This process can be found in all ML models and applications; from social media feeds to self-driving cars.
data acquisition/hypothesis (assimilation)
calculation error
Parameter updates/error minimization (accommodation)
evaluation
First, a hypothesis is calculated based on the lithuania phone number data available data, which should describe the existing observations as well as possible. There are usually deviations, which are calculated in the second step using an error function.
Hypothesis and error calculation (red lines) in the linear model
The actual learning then takes place in step number three, in which the calculated errors are minimized until an optimal value is reached. A fourth step is limited to the area of supervised learning, where the output is already known to us. By splitting it into a training and a test data set, we can evaluate our model here using previously unknown data.
What can be derived from this process is the conclusion that a machine learning model is more precise and accurate the more experience it has. In other words: the more training data is available, the more accurate the prediction . Gigantic amounts of data with positive and negative data are required. This is the great competitive advantage of tech companies like Google and Facebook. Since it cannot be assumed that they will ever share their knowledge and data with us (even if politicians insist on it), we as companies have to take action ourselves.