Data has a better idea – how does machine learn?
Machine learning surrounds us in everyday life. By 2020 according to Gartner chatbots will take over the customer service in 85%. 41% of consumers believe that AI will improve their quality of life. The Netflix portal saves about $ 1 billion a year using a machine learning algorithm that personalizes the suggestions for following viewing titles based on our past decisions. No need to convince anyone that ML is not the future, but the present. And what’s more… companies are investing intensively in ML because they realize how much profits can it bring. So it is worth to ask a basic question: how does machine learn?
Machine learning consists of three main elements: model, parameters, learner. The model is a system that creates predictions/identifications. Parameters determined by data provided to the system are factors by which the model formulates decisions. And last element: learning – the system adjusts the model by observing the differences between predictions and actual results. The algorithm should learn i.e. improve themselves the basis of observation of differences in results, update parameters based on new data, improve performance and gather experience.
In ML, several types of algorithms can be used. What are the basic differences between them?
Machine learning tasks are typically classified into several categories:
This model involves providing a set of input and output data. That means that model gets information about the expected results. What we expect from model is that after providing the right amount of data, the algorithm will be able to classify or predict results from new set of data and improve results on the base of gather experience.
Possibilities of application – risk management, fraud detection, personalization of interaction, speech and text recognition.
This model involves providing a set of input. And it must find a pattern based on the analysis of the received data and provide us with the output on its basis. The key task of the system is to identify complex processes and patterns on their own, without the need for human guidance. In this kind of learning, we can see a similarity to the operation of the human brain – conclusions are drawn based on analysis and observation.
Possibilities of application – recognition of similar objects
In this kind of learning, we do not provide the model with any kind of data. The only information it is a signal – a prize in the case of an accurate choice and a penalty in the case of erroneously taken action. The key here is interaction with the environment. The agent assesses the state of the environment and after observation takes action that is rewarded or punished. An example illustrating the operation of this kind of algorithm can be chess playing: a good selection of moves is rewarded with a victory or penalized with a failure in a party. An undoubted advantage of this approach is the fact that in this case the model can move in a completely unknown environment and following the trial and error method will be able to get act better.
Possibilities of application – planning of strategies in games, optimization of network traffic throughput.
*According to Google