Abstract background

Neural network: how it works, 3 advantages and 4 disadvantages

A neural network is an imitation of our brain, in structure and in operation. It weighs many vague and incomplete signals at once and learns the right weights itself from examples. That makes it strong in pattern recognition, but also uncertain and hard to influence: a self-learning network can see relationships that are not there.

Key takeaways

  • A neural network predicts on the basis of a weighted combination of all factors and accepts that the prediction is a best guess.
  • You use it when no single characteristic and no fixed combination of characteristics is decisive, so that a decision tree does not work.
  • It consists of an input layer, a hidden layer and an output layer; learning means adjusting weights and thresholds until the output is right.
  • Advantages: it handles vague and incomplete information, adapts to changing circumstances and excels in pattern recognition.
  • Disadvantages: uncertain outcomes, a learning phase is needed, strong dependence on data quality and hardly any influence on the inner workings.

What is a neural network?

A neural network is a model, inspired by the brain, that consists of layers of connected artificial neurons and learns from examples which weights to give to combinations of input signals in order to arrive at the best possible prediction.

Neural networks are one of the methods of predictive analytics and the engine behind much of modern artificial intelligence. They are not just another algorithm: because of its self-learning character a neural network can quickly get out of hand.

Introduction: who is going to play football?

Next to a district of a large city lies a complex of sports fields. At a certain moment, boys all over the district are cycling towards it, apparently to play football. None of them carries a sign saying what he is going to do. Can you determine from a number of characteristics whether they will indeed play football?

A boy carrying a new football will probably play, but you cannot be sure: he may have bought the ball as a present, or found it. A boy in a tracksuit with football boots gives no certainty either, nor does a group of boys with rucksacks. Only when you see them arrive, walk onto the pitch and kick a ball do you know who is playing. In practice you often need to know in advance. Because none of the characteristics is decisive, and neither is a fixed combination, you cannot use the decision tree method. You need a method that works with combinations of characteristics: a neural network.

The problem a neural network solves

Suppose data is known for a group of boys who all seem to be on their way to the sports complex: ball or no ball, with or without a group, sports clothing, football boots, bicycle. Inspection of the first twelve boys leads to these observations:

  • Having a ball is no guarantee of playing football.
  • Going along with a group with a ball results in playing football.
  • The combination of clothing and football boots results in playing football.
  • Having a bicycle has no predictive value.

It looks as if predicting is easy. But when the prediction for the next four boys is compared with their real behaviour, it suddenly turns out that boots and clothing are not decisive, and neither is going along with a group with a ball. Four ways are open to improve the prediction:

  1. Use more factors. The question is which.
  2. Use data from more boys. There is a good chance that predicting then gets harder: adding four boys already undermined the criteria.
  3. Accept that the prediction is not always right, as long as it is a best guess.
  4. Base the prediction on a weighted average of all factors. The only difficulty is assigning the right weight to each factor.

In fact only options 3 and 4 offer a solution, and a combination of the two gives the best prediction. That is precisely what a neural network does.

How a neural network works

Computers are very good at processing clearly defined data according to fixed protocols. One of the strengths of the human brain is working with vague and incomplete data according to often poorly defined methods. The difference lies in the way a neuron, the processing unit of the brain, handles data. A neuron collects signals from various sources. If the sum of all incoming signals exceeds a preset value, the neuron emits a signal at its output; the greater the sum, the greater the outgoing signal. By distributing information over neurons and processing outgoing signals in other neurons, the brain can handle poorly defined and incomplete information.

With a neural network a digital computer simulates this. The network must assign the right weights to the connections, in one of two ways:

  • Fixed: the weights for the signals are set permanently. The network is then only applicable to a strictly delimited task.
  • Adaptive: the weights arise by presenting combinations of input signals and (known) output signals.

Learning with and without supervision

By changing the weights of the connections, the network makes sure the answer is right, and repeats this for a large number of combinations. This “learning” continues until the changes in weights fall below a preset limit. From then on the network can be used for new, unknown situations.

Supervised versus unsupervised learning

The circumstances largely determine the strategy.

With supervision

Without supervision

Data used

Input and known output

Input only

When

Usually offline, beforehand

During use

Character

Tested before deployment

Training on the job

Risk

Depends on labelled examples

Fully self-learning, less control

The three layers of a neural network

  1. Input layer
  2. Hidden layer (also called processing layer)
  3. Output layer

Take an adaptive network that learns with supervision. Initially the weights of all connections are equal, as are the thresholds above which a neuron becomes active. By presenting the network with a large number of situations with a known result, the weights and thresholds are adjusted so that the output is correct each time. As soon as the changes fall below a preset value, the network is ready for use. An adaptive network can be readjusted later with a “refresher course”: again a large number of situations, for example combinations on which it still performs less well.

The advantages and disadvantages

3 advantages

  • It can handle vague and incomplete information well.
  • Its settings can adapt to changing circumstances and requirements.
  • It is at its best where a degree of pattern recognition is wanted, such as in image material.

4 disadvantages

  • The outcome has a certain uncertainty that is not always desirable.
  • Before it is usable it has to go through a learning phase, online or offline.
  • The quality of the outcomes depends strongly on the quality of the data used for learning.
  • Except with a fixed network, the user has hardly any influence on the internal weights and thresholds.

The limitations of this algorithm

Although a neural network copes well with vague and incomplete information, its outcome can be incomplete too. You will not find the certainty of a decision tree in a neural network, and not every data set or assignment is suitable. The art is to choose the assignments it is good at: in (almost) all cases, searching for patterns. Neural networks are used with great success in face recognition and in detecting patterns in photos or film footage. They can also score well in detecting (rare) diseases, because the information is often incomplete and the symptoms overlap.

How Diamond Decisions applies neural networks

Diamond Decisions only uses a neural network when simpler, explainable methods fall short and the pattern is worth it. After inspiration, business modeling determines what a wrong prediction costs, data analytics safeguards data quality and tests the network against a simple baseline, and the decision makers agree in advance where a human checks the outcome.

Talk to the author

Jack EsselinkJack Esselink
Diamond Decisions

Discuss your question with Jack

Tell us which decision needs to improve. In a no-obligation conversation we show you where your data delivers the most.

Schedule an introduction →

Step 3 of our approach

Diamond Data Analytics

A data warehouse that starts with your most important decisions and is ready for conversational AI: a working prototype within 24 hours.

Explore Data Analytics →

Frequently asked questions

01

What is a neural network?

A model inspired by the brain, consisting of layers of connected artificial neurons, that learns from examples which weights to give to combinations of input signals in order to predict as well as possible.

02

When do you use a neural network instead of a decision tree?

When no single characteristic and no fixed combination is decisive and you are looking for patterns in vague or incomplete data, such as images. If the outcome has to be explainable and certain, a decision tree is the better choice.

03

What are the layers of a neural network?

An input layer, one or more hidden (processing) layers and an output layer. Learning takes place by adjusting the weights of the connections and the thresholds of the neurons.

04

What are the main disadvantages of neural networks?

Uncertain outcomes, the need for a learning phase, strong dependence on the quality of the training data and hardly any insight into or influence on the internal workings.

05

Does an SME need its own neural network?

Rarely. Most SMEs benefit more from existing AI services that are built on neural networks, and from simple, explainable models on their own data. Building your own only pays off with a lot of data and a clear pattern-recognition task.

Further reading

About the author

Jack Esselink

Jack Esselink

Jack is a very experienced trainer and speaker on data and AI. As Data Science & Machine Learning Evangelist at IBM, he gave training courses, workshops and keynotes on every continent. His motto: “It’s all about data.”

About Diamond Decisions

Diamond Decisions helps organizations become better by enabling better decisions. We don’t start with data, dashboards or AI, but with the decision that needs to improve, and work back from there to KPIs, data and technology. We do that in four steps: inspiration and alignment, business modeling, analytics and data platform and data-driven decision makers. That way BI and AI don’t remain an IT project, but deliver value every day.

Schrijf je in voor de Masterclass