Algoritmo ANN | How the artificial neural network works

Share on facebook
Share on twitter
Share on linkedin
Share on telegram
Share on whatsapp

Contents

In the last post (Click here), we talk briefly about the basics of the ANN technique. But before using the technique, an analyst must know, How does the technique actually work? Even when detailed referral is not required, the algorithm framework must be known. This knowledge serves multiple purposes:

  • First, helps us understand the impact of increasing / shrink the data set vertically or horizontally in computational time.
  • Secondly, helps us understand the situations or cases in which the model fits best.
  • In third place, It also helps us to explain why a certain model works better in certain settings or situations..

This post will provide you with a basic understanding of the artificial neural network framework (ANN). We won't go into the actual bypass, but the information provided in this post will be enough for you to appreciate and put the algorithm into practice. At the end of the post, furthermore I will present my views on the three basic purposes of understanding any previously mentioned algorithm.

Neural network formulation

We will start with an understanding of the formulation of a simple hidden layer neural network. A simple neural network can be represented as shown in the next figure:

ANA

Links between nodes are the most crucial finding in an ANN. We will return to “how to find the weight of each link” after discussing the general framework. The only known values ​​in the diagram above are the inputs. Let's call the inputs I1, I2 e I3, hidden states like H1, H2.H3 and H4, outputs as O1 and O2. Link weights can be denoted with the following notation:

W (I1H1) is the weight of the link between nodes I1 and H1.

Below is the framework in which artificial neural networks work (ANN):

ANN flow chart

Few statistical details about the framework

Each linkage calculation in an artificial neural network (ANN) is equivalent. In general, we assume a sigmoid linkage between the input variables and the activation rate of the hidden nodes or between the hidden nodes and the activation rate of the output nodes. Let's prepare the equation to find the activation rate of H1.

Logit (H1) = W (I1H1) * I1 + W (I2H1) * I2 + W (I3H1) * I3 + Constant = f

=> P (H1) = 1 / (1 + e ^ (- f))

Here is what the sigmoid linkage looks like:

SigmoidFunction_701

How are weights recalibrated? A short note

Recalibrating weights is an easy procedure, but long. The only nodes where we know the error rate are the output nodes. The recalibration of weights on the link between the hidden node and the output node is a function of this error rate at the output nodes. But, How do we find the error rate in hidden nodes? It can be statistically shown that:

Error @ H1 = W (H1O1) *[email protected] + W (H1O2) *[email protected]

Using these errors, we can recalibrate the link weights between hidden nodes and input nodes in an equivalent way. Suppose this calculation is performed multiple times for each of the observations in the training set.

The three basic questions

What is the correlation between the time consumed by the algorithm and the volume of data (compared to traditional models such as logistics)?

As mentioned previously, for each observation, ANN performs multiple recalibrations for each link weight. Therefore, the time it takes for the algorithm to increase much faster than other traditional algorithms for the same increase in data volume.

In which situation does the algorithm best fit?

ANN is rarely used for predictive modeling. The reason is that Artificial Neural Networks (ANN) they usually try to fit the bonding too much. ANN is generally used in cases where what has happened in the past is repeated in almost exactly the same way. As an example, let's say we're playing the game of black jack against a computer. A clever ANN-based opponent would be a very good opponent for this case (assuming they can keep the computation time low). Over time, ANN will prepare for all possible card flow cases. And since we are not shuffling cards with a dealer, ANN will be able to memorize each call. Therefore, it is a kind of machine learning technique that has a huge memory. But it doesn't work well in case the scoring population is significantly different compared to the training sample. As an example, if I plan to target a customer for a campaign using their previous response from an ANN. I will probably use the wrong technique, since I might have over-adjusted the linkage between the solution and other predictors.

For the same reason, works very well in cases of image accreditation and voice accreditation.

What makes ANN a very strong model when it comes to memorization?

Artificial neural networks (ANN) have many different coefficients, that you can make the most of. Therefore, can handle much more variability compared to traditional models.

Was the post helpful? Have you used any other machine learning tools recently? Are you planning to use ANN in any of your business problems? If so, tell us how you plan to do it.

If you like what you have just read and want to continue your analytical learning, subscribe to our emails, Follow us on twitter or like ours Facebook page.

Subscribe to our Newsletter

We will not send you SPAM mail. We hate it as much as you.