I'm trying to somehow put everything in my head, I'm interested in which networks are better suited for tasks and why

I know that some tasks are not solved by networks at all, but it is easier to say the final goal than to abstract the network I would like to find.

Suppose the tasks are as follows:

  1. Voice recognition
  2. Learning musical taste and searching for similar songs
  3. Temperature control of the inertial system (very fast training is needed right in the process, otherwise something may overheat or overcool)
  4. Recognition of graphs (say, spectrum)
  5. To learn to recognize a person from a camera one by one by his photo.

    1 answer 1

    There are several main types of neural networks that are used to solve some types of tasks. Very, very briefly :

    FFNN The direct distribution network is the simplest and is mainly used for analyzing data in which a repeating pattern can be identified.

    RNN Recurrent Neural Network - a network that makes a conclusion based not only on real data but also on past ones. Used where it is important to take into account the influence of several factors.

    LSTM Network with long and short memory is a subspecies of recurrent networks, but unlike conventional recurrent networks, it can remember data for a long period in other words to understand the context.

    CNN Convolutional Neural Network - NA which processes data using filters and direct distribution network.

    Answering your question:

    1 and 2 - LSTM . All that is connected with speech processing and text analysis needs context, and by default a sentence is a consistent arrangement of words that are related to each other in meaning.

    3 - FFNN . I did not quite understand the essence, but if the logic is to set the temperature, depending on a set of factors, then the network of direct distribution is best suited.

    4 - Either FFNN or RNN , depending on the complexity and "predictability" of the graph.

    5 - CNN . All that is connected with object recognition and photo or video analysis, a convolutional neural network is always used.

    • Thank you, I thought nobody would answer me. - Fangog
    • At the expense of 3 points, I meant the analogue of the PID on the neural network - Fangog
    • At the expense of 4 points, can you answer what is meant by complexity? And predictability. Suppose there are spectra that differ from each other quite a bit (by 0.5%), are susceptible to noise and are slightly distorted with time. What is better to choose? - Fangog
    • I can’t tell you about FID because I don’t know all aspects of the device. At the expense of 4 points, in this case it is better to take the FFNN. - Arnis Shaykh
    • Another question, I want to generate light music on the LSTM network, I want to feed recordings from concerts (picture + sound) to the input, and get, for example, a simple square that changes color depending on the melody itself, how should I organize the architecture of such an application? - Fangog