Neural Network Exam Questions and Answers

In this blog, we are going to learn different Artificial Neural Network exam questions and answers, in brief, to understand very easily.

Artificial Neural Networks (ANNs) are the heart of machine learning and machine intelligence, used in many real-world applications like face recognition, speech recognition, industrial inspection and diagnosis, fault detection and prediction, stock market prediction, medical diagnosis, etc.

1. What are Artificial Neural Networks?

Artificial Neural Networks (ANNs) are a group of algorithms inspired by biological neural networks.

In ANNs, an information processing system is made up of a large number of highly interconnected processing elements that can store information and process it in real-time.

Although commonly referred to as neural networks or artificial neural networks, ANNs do not simulate neurons in any way.

Related Article: What is an Artificial Neural Network (ANN)?

2. Different Types of Artificial Neural Networks

ANNs are divided into several categories according to how they are designed. These categories include Perceptron, Multi-Layer Perceptron (MLP), Radial Basis Function (RBF), Self-Organizing Map (SOM), Hopfield Neural Network, Kohonen map, Associative memory neural network, Brain-Computer Interface (BCI), and Learning Vector Quantization Neural Networks (LVQ).

It contains different categories of Artificial Neural Networks like Feedforward neural networks, Recurrent neural networks, Convolutional neural networks, Deep learning Neural networks, etc.

Each type has its characteristics which makes it unique from others in terms of application areas as well as training techniques.

3. How do Neural Networks Work?

Imagine a neural network as a human brain that has several layers of neurons. When we feed input data to a neuron it fires or sends an output depending on its connection with other neurons.

Similarly, when we feed input data to an artificial neuron, it fires or sends out an output depending on its connection with other neurons in its layer.

The output of one neuron is used as input for another neuron in the next layer which makes it more complex.

4. What is Perceptron Learning Algorithm?

Back in 1957, Frank Rosenblatt came up with a learning algorithm for Neural Networks known as Perceptron.

Perceptron is an algorithm that divides its training data into two sets: positive and negative sets.

A positive set of data is those from which we want our neural network to learn whereas a negative set contains that data from which we don’t want our neural network to learn.

The neural network tries to find out a line or curve that separates both these sets by itself. This line or curve is called the decision boundary.

5. What is Delta Rule?

Delta Rule is a modified recursive least squares method of updating neuron weights in a neural network. 

It was developed by Paul Werbos in 1974. The delta rule has been used extensively for control systems, pattern recognition, time series prediction, data fitting, etc.

The main advantage of using the delta rule over the backpropagation algorithm is that it reduces computation time significantly since it uses only one pass through training data, unlike backpropagation which needs two passes.

6. What is the Back Propagation Learning Algorithm?

The backpropagation learning algorithm is one of the neural network learning algorithms which makes machines self-learn how to do a certain task.

It is also called backpropagation neural network training or BPNN training. This BPNN training will be done by iteratively changing the weights of connections among neurons based on errors between target output and actual output generated by a neural network.

For more details on Backpropagation, please check out our previous post here: What is Backpropagation in Neural Networks?

7. What is the Generalized Delta Rule?

The Generalized Delta Rule is one of three possible learning methods in Artificial Neural Networks (ANNs). The other two are Hebbian Learning, a type of Reinforcement Learning (RL), and Error Back-Propagation, a type of Supervised Learning.

A neural network with these capabilities for parallel distributed processing (PDP) has many applications, including pattern recognition or classification problems such as: identifying credit card fraud.

8. What is a Neural Network?

To understand how ANNs work, we must first understand what neurons are. Neurons are biological cells that process information and communicate with each other through electrical signals.

In neural networks, artificial neurons replace biological ones, Each neuron receives input from multiple sources and passes its output on to multiple targets via weighted connections called synapses.

A neural network is made up of interconnected processing elements that can be simulated with an electronic device (i.e., a computer). The neurons are organized in layers, with connections between each layer.

That means a neural network can take information from one layer, process it through the second layer of neurons, and then transmit it to a third layer where more computations occur.

This way, they perform tasks such as image recognition or voice recognition because they replicate how a human brain works.

9. Applications Of ANN Or NN

ANN or Neural Networks is one of the top choices for artificial intelligence algorithms.

Since its inception, it has given many successful applications in different fields like Robotics, Image Processing, NLP, Pattern Recognition, and Classification just to name a few.

Let’s take a look at some of the artificial Neural Networks (ANN) that were used by NASA to control spacecraft on Mars.

It is also used by Google for image recognition and classification using Deep Learning Algorithms. It is also used by Facebook to identify faces in uploaded photos from users.

10. What do you mean by Perceptron?

A perceptron is a linear classifier based on supervised learning, where each sample is mapped into a fixed number of classes represented by using decision hyperplanes in feature space.

It can be shown that no single-layer neural network can classify all patterns correctly. The solution proposed by Frank Rosenblatt in 1958 was to add more neurons or layers between input and output layers.

Such networks are called multilayer perceptrons (MLP). multilayer perceptrons (MLP) consist of an input layer, one or more hidden layers, and an output layer.

11. What are the different types of Perceptrons?

A perceptron is a linear classifier that maps an input vector x=(x1,x2,….,xn) onto an output of value 1 (indicating positive class) or 0 (indicating negative class).

There are three types of Perceptrons: Multilayer perceptron, Single layer perceptron and Binary perception.

In a multilayer perceptron, there are two layers between the input and output layer. In a single layer perceptron, there is only one layer in between the input and output layer. In binary perceptron, all weights wij = 0 for i > j .

12. Neural Network (NN) and Artificial Intelligence (AI)

A neural network (NN) is an artificial intelligence (AI) system that attempts to simulate the functional aspects of biological neural networks and can be used to solve problems more efficiently than an expert human.

An NN consists of an interconnected group of artificial neurons that process information by their dynamic state response to external inputs.

Neural networks learn (progressively improve performance) through experience, rather than from pre-programmed knowledge.

13. What do you mean by Cost Function?

Before going into Artificial Neural networks, we need to learn what is the cost function. The general term of the cost function is an evaluation criterion for optimization.

For example: In the case of Linear Regression, Mean Squared Error is used as a cost function for finding optimal weights of the linear regression model by minimizing its value.

And in the case of Logistic Regression, Cross-Entropy is used as a cost function for finding optimal weights of the logistic regression model by minimizing its value.

And in the case of Neural Networks or Deep Learning models, it’s a loss function that can be defined using different metrics like MSE (Mean Squared Error), Accuracy (Accuracy), etc.

14. Activation Functions used in Neural Networks

The Sigmoid, Tanh, Rectifier Linear Unit (ReLU), and Hyperbolic Tangent functions are different activation functions used in neural networks.

Sigmoid: The sigmoid function is also known as the logistic function. It is a type of nonlinear mathematical function that outputs values between 0 and 1. It has an S-shaped curve that looks like an S on a graph with the x-axis representing input values and the y-axis representing output values.

Tanh: The tanh function is similar to sigmoid function but its range is [-1,1] instead of [0,1]. The hyperbolic tangent function is defined as: $tanh(x) = \frac$ Where $e$ represents Euler’s number. This means that if you apply the tanh function to any real value, you will get a value between -1 and 1.

ReLU: The rectifier linear unit or ReLU is another activation function that maps negative numbers to zero and positive numbers to one. This means it turns off neurons that are not active at all.

Maxout function: The max out function is a nonlinear activation function that was introduced by Geoffrey Hinton in his paper titled Rectified Linear Units Improve Restricted Boltzmann Machines. It is similar to a rectifier linear unit but it outputs values of 1 if all its inputs are 1, otherwise, it outputs 0. It can be represented as: $max(x_1, x_2, \dots, x_n)$ Where $x_i$ represents input value.

Exponential Linear unit (ELU) function: The exponential linear unit or ELU is another activation function that was introduced by Nair and Hinton in their paper titled Rectified Linear Units Can Replace Global Pooling in Deep Neural Networks. It is similar to the rectifier linear unit but instead of outputting values between 0 and 1, it outputs values between -1 and 1.

15. What is the use of Loss Functions?

The loss function is used in any Machine Learning method to give feedback on how good or bad is our model doing for a given input.

It allows us to find out if our training has been successful. In some cases, it can be minimized to minimize prediction errors.

Many loss functions have been introduced over time, from very basic ones such as Mean Squared Error (MSE) to more complex ones such as Cross-Entropy Loss (CEL).

16. What are Weights and Biases in Neural Networks?

Neural networks are made up of various layers, each serving a specific purpose in creating a network that can learn.

The first layer is called an input layer, which represents information on a problem to be solved.

The last layer is called an output layer, which determines the solution to be used. In between these two layers are hidden layers where data processing happens.

All inputs pass through these hidden layers before reaching an output, whether it’s predicting an outcome or identifying faces in images.

17. What is Hyperparameters or Hyperparameters Tuning?

How we set our hyperparameters is a crucial part of training an Artificial Neural Network.

There are two types of hyperparameters: those that must be tuned manually, known as Tunable Hyperparameters, such as regularization rate or learning rate.

Others have sensible defaults and therefore need not be tuned manually, known as Non-Tunable Hyperparameters.

Such parameters may be a choice between multiple loss functions or metrics for performance evaluation.

In both cases, it’s important to evaluate several values for each parameter before deciding on what to use.

For example, there’s no universal value for a neural network’s learning rate (also called momentum).

You might find that one works better than another depending on your dataset and task at hand.

18. Backpropagation Algorithm & Error Function

The Backpropagation algorithm is a gradient descent type of optimization method for use in neural networks.

It does what it says on the tin, that is to say, it propagates errors back through our network from output nodes to hidden nodes.

Then from hidden nodes to input nodes to determine how much each element of each layer needs to be adjusted so as we don’t overshoot or undershoot an error target we want to achieve by a certain end-state.

This procedure is known as back-propagation because it starts at one end (the output) and works its way back through all layers until it reaches its starting point (the input).

The error function calculates a single value representing how well your neural network did at predicting some outcome you were trying to predict.

19. What is Data Normalization?

Neural Network questions answers, data normalization is used to represent variables in a way that reduces or eliminates redundancy.

Several techniques have been developed for performing data normalization; here we focus on one technique: scaling.

The basic idea behind scaling is to scale each of the input variables to a mean of zero and a standard deviation of one before passing them through an artificial neural network.

Scaling ensures that all inputs are presented to the neural network with equal weight. This can be useful when dealing with noisy data sets where some inputs may be more important than others.

Scaling also allows you to use linear regression as your loss function even if your original dataset contains non-linear relationships between inputs and outputs.

In practice, it is often best to perform both normalizations and scaling on your dataset before feeding it into an artificial neural network (ANN).

20. What are Feedforward Neural Networks?

Feedforward neural networks (FNNs) are part of artificial neural networks also a special case of Multi-Layer Perceptrons (MLPs).

Feedforward means that only data flows in one direction through layers. The input layer is called an ‘input layer’ because it receives information from outside its network.

The output layer is called an ‘output layer’ because it produces information from outside its network.

A hidden layer is called a ‘hidden layer’ because it neither receives nor sends information to or from any other neural network.

Hidden layers can be used to model complex relationships between inputs and outputs, as well as to perform nonlinear transformations on their inputs.

Related Article: Feed Forward Neural Networks Ultimate Guide Explained

21. What are Recurrent Neural Networks (RNN)?

The architecture of an RNN cell is quite similar to that of a feed-forward NN with some modifications.

As we know, in feed-forward networks, when we pass an input through layer after layer of neurons, at each layer we apply a nonlinear transformation to map it into a space which makes it easier for us to predict values at subsequent layers. The output from one neuron is fed as input to another neuron.

In RNNs, however, we are dealing with sequences rather than vectors or scalars. So how do we use neural networks to model sequences? The answer lies in using recurrent connections between neurons (i.e., connections that loop back on themselves). This way, every time step t+1 becomes dependent on time step t.

22. What are Convolutional Neural Networks (CNN)?

A convolutional neural network (CNN, or ConvNet) is a class of deep, feed-forward artificial neural networks, most commonly applied to analyzing visual imagery.

CNN has been popularized by its success in automated image captioning tasks, where they directly address the classification of images at multiple levels of abstraction.

In other words, CCNNsare just like real human brains so we can use them for tasks such as speech recognition and object detection.

They consist of an input and an output layer, with one or more hidden layers between them. Each connection within a layer is assigned a weight, which indicates how much that input will contribute to computing the output from that neuron.

23. What is AutoEncoders?

AutoEncoder is a neural network that consists of two separate layers – an encoding layer, and a decoding layer.

Encoding takes place in hidden layers of one Neural Network while Decoding takes place in hidden layers of another Neural Network.

The two Neural Networks are connected through their inputs/outputs, hence it’s called Auto-Encoders.

This type of Neural Network can be used for dimensionality reduction and feature learning. It learns to transform input data into output data with fewer dimensions than input data.

This technique can be used for image compression, as well as for learning latent space representations from unlabeled data like text or speech signals, etc.

Related Article: Autoencoders: Introduction to Neural Networks

24. What do you mean by Boltzmann Machine?

It is a type of associative memory, also known as a Boltzmann machine, that can be used to solve machine learning problems.

Associative memory is a neural network architecture where each node (neuron) in a layer connects to each node in another layer.

In an associative memory architecture, if input A activates neuron X in layer 1, it also activates neuron Y in layer 2.

Thus, neurons in different layers are associated with one another. The activation of a neuron in layer 1 causes all its associated neurons in layer 2 to activate.

The strength of association between two neurons is represented by their connection weight or synaptic weight.

Neural networks that use such associations are called associative neural networks. Such architectures have been used for unsupervised learning, feature extraction, dimensionality reduction, information retrieval, and other applications.

They have been applied successfully to problems such as classification and clustering, compression, anomaly detection, and prediction of non-linear time series data.

25. What are Optimizers?

The first step in implementing any Neural Network is to create a Model. In DL4J, there are three ways to create a model: by specifying its architecture manually, by training it from data (using Stochastic Gradient Descent), or by loading weights pre-trained on other data.

The simplest way to create a model is by specifying its architecture manually – here you decide how many neurons make up each layer, what their activation functions are, etc.

Backpropagation refers to computing gradients concerning all parameters in an optimization problem.

26. Epoch, Batch, and Iteration in Neural Networks

A neural network is trained in a supervised manner to predict output for a given input(s).

It learns by adjusting its weights according to some error signal (also called cost function or loss function).

The backpropagation algorithm adjusts these weights using three steps: epoch, batch, and iteration.

When we train a neural network using multiple epochs (one pass through all data) with each epoch split into multiple batches (multiple passes over subsets of data), then each such batch is called an iteration.

In other words, one epoch = one batch = one iteration.

In practice, neural networks are typically trained using hundreds or thousands of iterations.


Artificial Neural Network also known as ANNs, is a kind of machine learning algorithm and above all Neural Network, exam questions and answers can give you a detailed idea about them.

In which computer software or hardware components are modeled on biological neural networks in order to perform tasks such as data classification, predictive analysis, function approximation, and processing of inputs.

So it can be said that Neural Networks and Deep Learning are similar concepts.

If you want to learn more about ANNs, then refer to our post-neural network exam questions answers where we have discussed some important questions related to Neural Networks with their answers.

Leave a Reply

Your email address will not be published.