Deep Learning Interview Questions For Freshers

1. What is Deep Learning?

Deep learning is the branch of machine learning which is based on artificial neural network architecture which makes it capable of learning complex patterns and relationships within data. An artificial neural network or ANN uses layers of interconnected nodes called neurons that work togeather to process and learn from the input data.

In a fully connected Deep neural network, there is an input layer and one or more hidden layers connected one after the other. Each neuron receives input from the previous layer neurons or the input layer. The output of one neuron becomes the input to other neurons in the next layer of the network, and this process continues until the final layer produces the output of the network. The layers of the neural network transform the input data through a series of nonlinear transformations, allowing the network to learn complex representations of the input data.

 Today Deep learning has become one of the most popular and visible areas of machine learning, due to its success in a variety of applications, such as computer vision, natural language processing, and Reinforcement learning.

2. What is an artificial neural network?

An artificial neural network is inspired by the networks and functionalities of human biological neurons. it is also known as neural networks or neural nets. ANN uses layers of interconnected nodes called artificial neurons that work together to process and learn the input data. The starting layer artificial neural network is known as the input layer, it takes input from external input sources and transfers it to the next layer known as the hidden layer where each neuron received inputs from previous layer neurons and computes the weighted sum, and transfers to the next layer neurons. These connections are weighted means effects of the inputs from the previous layer are optimized more or less by assigning different-different weights to each input and it is adjusted during the training process by optimizing these weights for better performance of the model. The output of one neuron becomes the input to other neurons in the next layer of the network, and this process continues until the final layer produces the output of the network.

3. How does Deep Learning differ from Machine Learning?

Machine learning and deep learning both are subsets of artificial intelligence but there are many similarities and differences between them.

4. What are the applications of Deep Learning?

Deep learning has many applications, and it can be broadly divided into computer vision, natural language processing (NLP), and reinforcement learning. 

• Computer vision:  Deep learning employs neural networks with several layers, which enables it used for automated learning and recognition of complex patterns in images. and machines can perform image classification, image segmentation, object detection, and image generation task accurately. It has greatly increased the precision and effectiveness of computer vision algorithms, enabling a variety of uses in industries including healthcare, transportation, and entertainment.
• Natural language processing (NLP): Natural language processing (NLP) gained enormously from deep learning, which has enhanced language modeling, sentiment analysis, and machine translation. Deep learning models have the ability to automatically discover complex linguistic features from text data, enabling more precise and effective processing of inputs in natural language.
• Reinforcement learning: Deep learning is used in reinforcement learning to evaluate the value of various actions in various states, allowing the agent to make better decisions that can maximize the predicted rewards. By learning from these mistakes, an agent eventually raises its performance. Deep learning applications that use reinforcement learning include gaming, robotics, and control systems.

5. What are the challenges in Deep Learning?

Deep learning has made significant advancements in various fields, but there are still some challenges that need to be addressed. Here are some of the main challenges in deep learning:

1. Data availability: It requires large amounts of data to learn from. For using deep learning it’s a big concern to gather as much data for training.
2. Computational Resources: For training the deep learning model, it is computationally expensive because it requires specialized hardware like GPUs and TPUs.
3. Time-consuming: While working on sequential data depending on the computational resource it can take very large even in days or months. 
4. Interpretability: Deep learning models are complex, it works like a black box. it is very difficult to interpret the result.
5. Overfitting: when the model is trained again and again, it becomes too specialized for the training data, leading to overfitting and poor performance on new data.

6. How Biological neurons are similar to the Artificial neural network.

The concept of artificial neural networks comes from biological neurons found in animal brains So they share a lot of similarities in structure and function wise.

• Structure: The structure of artificial neural networks is inspired by biological neurons. A biological neuron has dendrites to receive the signals, a cell body or soma to process them, and an axon to transmit the signal to other neurons.  In artificial neural networks input signals are received by input nodes, hidden layer nodes compute these input signals, and output layer nodes compute the final output by processing the outputs of the hidden layer using activation functions.
• Synapses: In biological neurons, synapses are the connections between neurons that allow for the transmission of signals from dendrites to the cell body and the cell body to the axon like that. In artificial neurons, synapses are termed as the weights which connect the one-layer nodes to the next-layer nodes. The weight value determines the strength between the connections. 
• Learning: In biological neurons, learning occurs in the cell body or soma which has a nucleus that helps to process the signals. If the signals are strong enough to reach the threshold, an action potential is generated that travels through the axons. This is achieved by synaptic plasticity, which is the ability of synapses to strengthen or weaken over time, in response to increases or decreases in their activity. In artificial neural networks, the learning process is called backpropagations, which adjusts the weight between the nodes based on the difference or cost between the predicted and actual outputs.
• Activation: In biological neurons, activation is the firing rate of the neuron which happens when the signals are strong enough to reach the threshold. and in artificial neural networks, activations are done by mathematical functions known as activations functions which map the input to the output.

7. How deep learning is used in supervised, unsupervised as well as reinforcement machine learning?

Deep learning can be used for supervised, unsupervised as well as reinforcement machine learning. it uses a variety of ways to process these.

• Supervised Machine Learning: Supervised machine learning is the machine learning technique in which the neural network learns to make predictions or classify data based on the labeled datasets. Here we input both input features along with the target variables. the neural network learns to make predictions based on the cost or error that comes from the difference between the predicted and the actual target, this process is known as backpropagation.  Deep learning algorithms like Convolutional neural networks, Recurrent neural networks are used for many supervised tasks like image classifications and recognization, sentiment analysis, language translations, etc.
• Unsupervised Machine Learning: Unsupervised machine learning is the machine learning technique in which the neural network learns to discover the patterns or to cluster the dataset based on unlabeled datasets. Here there are no target variables. while the machine has to self-determined the hidden patterns or relationships within the datasets. Deep learning algorithms like autoencoders and generative models are used for unsupervised tasks like clustering, dimensionality reduction, and anomaly detection.
• Reinforcement  Machine Learning: Reinforcement  Machine Learning is the machine learning technique in which an agent learns to make decisions in an environment to maximize a reward signal. The agent interacts with the environment by taking action and observing the resulting rewards. Deep learning can be used to learn policies, or a set of actions, that maximizes the cumulative reward over time. Deep reinforcement learning algorithms like Deep Q networks and Deep Deterministic Policy Gradient (DDPG) are used to reinforce tasks like robotics and game playing etc.

8. What is a Perceptron?

Perceptron is one of the simplest Artificial neural network architectures. It was introduced by Frank Rosenblatt in 1957s. It is the simplest type of feedforward neural network, consisting of a single layer of input nodes that are fully connected to a layer of output nodes. It can learn the linearly separable patterns. it uses slightly different types of artificial neurons known as threshold logic units (TLU). it was first introduced by McCulloch and Walter Pitts in the 1940s. it computes the weighted sum of its inputs and then applies the step function to compare this weighted sum to the threshold. the most common step function used in perceptron is the Heaviside step function.

A perceptron has a single layer of threshold logic units with each TLU connected to all inputs. When all the neurons in a layer are connected to every neuron of the previous layer, it is known as a fully connected layer or dense layer. During training, The weights of the perceptron are adjusted to minimize the difference between the actual and predicted value using the perceptron learning rule i.e 

w_i = w_i + (learning_rate * (true_output - predicted_output) * x_i)

Here, x_i and w_i are the i^th input feature and the weight of the i^th input feature.

9.  What is Multilayer Perceptron? and How it is different from a single-layer perceptron?

A multilayer perceptron (MLP) is an advancement of the single-layer perceptron which uses more than one hidden layer to process the data from input to the final prediction. It consists of multiple layers of interconnected neurons, with multiple nodes present in each layer. The MLP architecture is referred to as the feedforward neural network because data flows in one direction, from the input layer through one or more hidden layers to the output layer.  

The differences between the single-layer perceptron and multilayer perceptron are as follows:

• Architecture: A single-layer perceptron has only one layer of neurons, which takes the input and produces an output. While a multilayer perceptron has one or more hidden layers of neurons between the input and output layers. 
• Complexity: A single-layer perceptron is a simple linear classifier that can only learn linearly separable patterns. While a multilayer perceptron can learn more complex and nonlinear patterns by using nonlinear activation functions in the hidden layers. 
• Learning: Single-layer perceptrons use a simple perceptron learning rule to update their weights during training. While multilayer perceptrons use a more complex backpropagation algorithm to train their weights, which involves both forward propagations of input through the network and backpropagation of errors to update the weights. 
• Output: Single-layer perceptrons produce a binary output, indicating which of two possible classes the input belongs to. Multilayer perceptrons can produce real-valued outputs, allowing them to perform regression tasks in addition to classification.
• Applications: Single-layer perceptrons are suitable for simple linear classification tasks whereas Multilayer perceptrons are more suitable for complex classification tasks where the input data is not linearly separable, as well as for regression tasks where the output is continuous variables.

10. What are Feedforward Neural Networks?

A feedforward neural network (FNN) is a type of artificial neural network, in which the neurons are arranged in layers, and the information flows only in one direction, from the input layer to the output layer, without any feedback connections. The term “feedforward” means information flows forward through the neural network in a single direction from the input layer through one or more hidden layers to the output layer without any loops or cycles.

In a feedforward neural network (FNN) the weight is updated after the forward pass. During the forward pass, the input is fed and it computes the prediction after the series of nonlinear transformations to the input. then it is compared with the actual output and errors are calculated.

During the backward pass also known as backpropagation, Based on the differences, the error is first propagated back to the output layer, where the gradient of the loss function with respect to the output is computed. This gradient is then propagated backward through the network to compute the gradient of the loss function with respect to the weights and biases of each layer. Here chain rules of calculus are applied with respect to weight and bias to find the gradient. These gradients are then used to update the weights and biases of the network so that it can improve its performance on the given task. 

11. What is GPU?

A graphics processing unit, sometimes known as a GPU, is a specialized electronic circuit designed to render graphics and images on a computer or other digital device fast and effectively.

Originally developed for use in video games and other graphical applications, GPUs have grown in significance in a number of disciplines, such as artificial intelligence, machine learning, and scientific research, where they are used to speed up computationally demanding tasks like training deep neural networks.

One of the main benefits of GPUs is their capacity for parallel computation, which uses a significant number of processing cores to speed up complicated calculations. Since high-dimensional data manipulations and matrix operations are frequently used in machine learning and other data-driven applications, these activities are particularly well suited for them.

12. What are the different layers in ANN? What is the notation for representing a node of a particular layer?

There are commonly three different types of layers in an artificial neural network (ANN):

• Input Layer: This is the layer that receives the input data and passes it on to the next layer. The input layer is typically not counted as one of the hidden layers of the network. 
• Hidden Layers: The input layer is the one that receives input data and transfers it to the next layer. Usually, the input layer is not included in the list of the hidden layers of the neural network. 
• Output Layer: This is the output-producing layer of the network. A binary classification problem might only have one output neuron, but a multi-class classification problem might have numerous output neurons, one for each class. The number of neurons in the output layer depends on the type of problem being solved. 

We commonly use a notation like  to represent a node of a specific layer in an ANN, where L denotes the layer number and i denotes the node’s index inside that layer. For instance, the input layer’s first node may be written as  whereas the third hidden layer’s second node might be written as  With this notation, it is simple to refer to specific network nodes to understand the structure of the network as a whole.

13. What is forward and backward propagation?

In deep learning and neural networks, In the forward pass or propagation, The input data propagates through the input layer to the hidden layer to the output layer. During this process, each layer of the neural network performs a series of mathematical operations on the input data and transfers it to the next layer until the output is generated.

Once the forward propagation is complete, the backward propagation, also known as backpropagation or back prop, is started. During the backward pass, the generated output is compared to the actual output and based on the differences between them the error is measured and it is propagated backward through the neural network layer. Where the gradient of the loss function with respect to the output is computed. This gradient is then propagated backward through the network to compute the gradient of the loss function with respect to the weights and biases of each layer. Here chain rules of calculus are applied with respect to weight and bias to find the gradient. These gradients are then used to update the weights and biases of the network so that it can improve its performance on the given task. 

 In simple terms, the forward pass involves feeding input data into the neural network to produce an output,  while the backward pass refers to utilizing the output to compute the error and modify the network’s weights and biases.

14. What is the cost function in deep learning?

The cost function is the mathematical function that is used to measure the quality of prediction during training in deep neural networks. It measures the differences between the generated output of the forward pass of the neural network to the actual outputs, which are known as losses or errors. During the training process, the weights of the network are adjusted to minimize the losses. which is achieved by computing the gradient of the cost function with respect to weights and biases using backpropagation algorithms.

 The cost function is also known as the loss function or objective function. In deep learning, different -different types of cost functions are used depending on the type of problem and neural network used.  Some of the common cost functions are as follows:

• Binary Cross-Entropy for binary classification measures the difference between the predicted probability of the positive outcome and the actual outcome.
• Categorical Cross-Entropy for multi-class classification measures the difference between the predicted probability and the actual probability distribution.
• Sparse Categorical Cross-Entropy for multi-class classification is used when the actual label is an integer rather than in a one-hot encoded vector.
• Kullback-Leibler Divergence (KL Divergence) is used in generative learning like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), it measures the differences between two probability distributions.
• Mean Squared Error for regression to measure the average squared difference between actual and predicted outputs.

15. What are activation functions in deep learning and where it is used?

Deep learning uses activation functions, which are mathematical operations that are performed on each neuron’s output in a neural network to provide nonlinearity to the network. The goal of activation functions is to inject non-linearity into the network so that it can learn the more complex relationships between the input and output variables. 

In other words, the activation function in neural networks takes the output of the preceding linear operation (which is usually the weighted sum of input values i.e w*x+b) and mapped it to a desired range because the repeated application of weighted sum (i.e w*x +b) will result in a polynomial function. The activation function transformed the linear output into non-linear output which makes the neural network capable to approximate more complex tasks.

In deep learning, To compute the gradients of the loss function with respect to the network weights during backpropagation, activation functions must be differentiable. As a result, the network may use gradient descent or other optimization techniques to find the optimal weights to minimize the loss function.

Although several activation functions, such as ReLU, and Hardtanh, contain point discontinuities, they are still differentiable almost everywhere. The gradient is not defined at the point of discontinuity, This does not have a substantial impact on the network’s overall gradient because the gradient at these points is normally set to zero or a small value.

16. What are the different different types of activation functions used in deep learning?

In deep learning, several different-different types of activation functions are used. Each of them has its own strength and weakness. Some of the most common activation functions are as follows.

• Sigmoid function: It maps any value between 0 and 1. It is mainly used in binary classification problems. where it maps the output of the preceding hidden layer into the probability value.
• Softmax function: It is the extension of the sigmoid function used for multi-class classification problems in the output layer of the neural network, where it maps the output of the previous layer into a probability distribution across the classes, giving each class a probability value between 0 and 1 with the sum of the probabilities over all classes is equal to 1. The class which has the highest probability value is considered as the predicted class.
• ReLU (Rectified Linear Unit) function: It is a non-linear function that returns the input value for positive inputs and 0 for negative inputs. Deep neural networks frequently employ this function since it is both straightforward and effective.
• Leaky ReLU function: It is similar to the ReLU function, but it adds a small slope for negative input values to prevent dead neurons.
• Tanh (hyperbolic tangent) function: It is a non-linear activations function that maps the input’s value between -1 to 1. It is similar to the sigmoid function but it provides both positive and negative results. It is mainly used for regression tasks, where the output will be continuous values.

17. How do neural networks learn from the data?

In neural networks, there is a method known as backpropagation is used while training the neural network for adjusting weights and biases of the neural network. It computes the gradient of the cost functions with respect to the parameters of the neural network and then updates the network parameters in the opposite direction of the gradient using optimization algorithms with the aim of minimizing the losses.

During the training, in forward pass the input data passes through the network and generates output. then the cost function compares this generated output to the actual output. then the backpropagation computes the gradient of the cost function with respect to the output of the neural network. This gradient is then propagated backward through the network to compute the gradient of the loss function with respect to the weights and biases of each layer. Here chain rules of differentiations are applied with respect to the parameters of each layer to find the gradient. 

Once the gradient is computed, The optimization algorithms are used to update the parameters of the network. Some of the most common optimization algorithms are stochastic gradient descent (SGD), mini-batch, etc.

The goal of the training process is to minimize the cost function by adjusting the weights and biases during the backpropagation.

18. How the number of hidden layers and number of neurons per hidden layer are selected?

There is no one-size-fits-all solution to this problem, hence choosing the number of hidden layers and neurons per hidden layer in a neural network is often dependent on practical observations and experimentation. There are, however, a few general principles and heuristics that may be applied as a base.

• The number of hidden layers can be determined by the complexity of the problem being solved. Simple problems can be solved with just one hidden layer whereas more complicated problems may require two or more hidden levels. However adding more layers also increases the risk of overfitting, so the number of layers should be chosen based on the trade-off between model complexity and generalization performance.
• The number of neurons per hidden layer can be determined based on the number of input features and the desired level of model complexity. There is no hard and fast rule, and the number of neurons can be adjusted based on the results of experimentation and validation.

In practice, it is often useful to start with a simple model and gradually increase its complexity until the desired performance is achieved. This process can involve adding more hidden layers or neurons or experimenting with different architectures and hyperparameters. It is also important to regularly monitor the training and validation performance to detect overfitting and adjust the model accordingly.

19. What is overfitting and how to avoid it?

Overfitting is a problem in machine learning that occurs when the model learns to fit the training data too close to the point that it starts catching up on noise and unimportant patterns. Because of this, the model performs well on training data but badly on fresh, untested data, resulting in poor generalization performance.

To avoid overfitting in deep learning we can use the following techniques:

1. Simplify the model: Overfitting may be less likely in a simpler model with fewer layers and parameters. In practical applications, it is frequently beneficial, to begin with a simple model and progressively increase its complexity until the desired performance is attained.
2. Regularization: Regularization is a technique used in machine learning to prevent the overfitting of a model by adding a penalty term, it imposes the constraint on the weight of the model. Some of the most common regularization techniques are as follows:
   1. L1 and L2 regularization: L1 regularization sparse the model by equating many model weights equal to 0 while L2 regularization constrains the weight of the neural network connection.
   2. Dropout: Dropout is a technique that randomly drops out or disables some of the randomly selected neurons. It is applied after the activation functions of the hidden layer. Typically, it is set to a small value like 0.2 or 0.25. For the dropout value of 0.20, Each neuron in the previously hidden layer has a 20% chance of being inactive. It is only operational during the training process.
   3. Max-Norm Regularization: It constrains the magnitude of the weights in a neural network by setting a maximum limit (or norm) on the weights of the neurons, such that their values cannot exceed this limit.
3. Data augmentation: By applying various transformations, such as rotating or flipping images, to new training data, it is possible to teach the model to become more robust to changes in the input data.
4. Increasing the amount of training data: By increasing the amount of data can provide the model with a diverse set of examples to learn from, which can be helpful to prevent overfitting.
5. Early stopping: This involves keeping track of the model’s performance on a validation set during training and terminating the training process when the validation loss stops decreasing.

20. Define epoch, iterations, and batches.

A complete cycle of deep learning model training utilizing the entire training dataset is called an epoch. Each training sample in the dataset is processed by the model during a single epoch, and its weights and biases are adjusted in response to the estimated loss or error. The number of epochs will range from 1 to infinite. User input determines it. It is always an Integral value.

Iteration refers to the procedure of running a batch of data through the model, figuring out the loss, and changing the model’s parameters. Depending on the number of batches in the dataset, one or more iterations can be possible within a single epoch.

A batch in deep learning is a subset of the training data that is used to modify the weights of a model during training. In batch training, the entire training set is divided into smaller groups, and the model is updated after analyzing each batch.  An epoch can be made up of one or more batches.

• The batch size will be more than one and always less than the number of samples.
• Batch size is a hyperparameter, it is set by the user. where the number of iterations per epoch is calculated by dividing the total number of training samples by the individual batch size.

Deep learning training datasets are often separated into smaller batches, and the model analyses each batch sequentially, one at a time, throughout each epoch.  On the validation dataset, the model performance can be assessed after each epoch. This helps in monitoring the model’s progress. 

For example: Let’s use 5000 training samples in the training dataset. Furthermore, we want to divide the dataset into 100 batches. If we choose to use five epochs, the total number of iterations will be as follows:

Total number of training samples = 5000
Batch size = 100
Total number of iterations=Total number of training samples/Batch size=5000/100=50
Total number of iterations = 50
One epoch = 50 iterations
Total number of iterations in 5 epochs = 50*5 = 250 iterations.

21. Define the learning rate in Deep Learning.

The learning rate in deep learning is a hyperparameter that controls how frequently the optimizer adjusts the neural network’s weights when it is being trained. It determines the step size to which the optimizer frequently updates the model parameters with respect to the loss function. so, that losses can be minimized during training.

With the high learning rate, the model may converge fast, but it may also overshoot or bounce around the ideal solution. On the other hand, a low learning rate might make the model converge slowly, but it could also produce a solution that is more accurate.

Choosing the appropriate learning rate is crucial for the successful training of deep neural networks.

22. What is the cross-entropy loss function?

Cross-entropy is the commonly used loss function in deep learning for classification problems. The cross-entropy loss measures the difference between the real probability distribution and the predicted probability distribution over the classes.

The formula for the Cross-Entropy loss function for the K classes will be:

Here, Y and  are actual and predicted values for a single instance. k represents a particular class and is a subset of K.

23. What is gradient descent?

Gradient descent is the core of the learning process in machine learning and deep learning. It is the method used to minimize the cost or loss function by iteratively adjusting the model parameters i.e. weight and biases of the neural layer. The objective is to reduce this disparity, which is represented by the cost function as the difference between the model’s anticipated output and the actual output.

The gradient is the vector of its partial derivatives with respect to its inputs, which indicates the direction of the steepest ascent (positive gradient) or steepest descent (negative gradient) of the function. 

In deep learning, The gradient is the partial derivative of the objective or cost function with respect to its model parameters i.e. weights or biases, and this gradient is used to update the model’s parameters in the direction of the negative gradient so that it can reduce the cost function and increase the performance of the model. The magnitude of the update is determined by the learning rate, which controls the step size of the update.

24. How do you optimize a Deep Learning model?

A Deep Learning model may be optimized by changing its parameters and hyperparameters to increase its performance on a particular task. Here are a few typical methods for deep learning model optimization:

• Choosing the right architecture
• Adjusting the learning rate
• Regularization
• Data augmentation
• Transfer learning
• Hyperparameter tuning

25. Define Batch, Stochastic, and Mini gradient descent.

There are several variants of gradient descent that differ in the way the step size or learning rate is chosen and the way the updates are made. Here are some popular variants:

• Batch Gradient Descent: In batch gradient descent, To update the model parameters values like weight and bias, the entire training dataset is used to compute the gradient and update the parameters at each iteration. This can be slow for large datasets but may lead to a more accurate model. It is effective for convex or relatively smooth error manifolds because it moves directly toward an optimal solution by taking a large step in the direction of the negative gradient of the cost function. However, it can be slow for large datasets because it computes the gradient and updates the parameters using the entire training dataset at each iteration. This can result in longer training times and higher computational costs.
• Stochastic Gradient Descent (SGD): In SGD, only one training example is used to compute the gradient and update the parameters at each iteration. This can be faster than batch gradient descent but may lead to more noise in the updates.
• Mini-batch Gradient Descent: In Mini-batch gradient descent a small batch of training examples is used to compute the gradient and update the parameters at each iteration. This can be a good compromise between batch gradient descent and Stochastic Gradient Descent, as it can be faster than batch gradient descent and less noisy than Stochastic Gradient Descent.

26. What are the different types of Neural Networks?

There are different-different types of neural networks used in deep learning. Some of the most important neural network architectures are as follows;

1. Feedforward Neural Networks (FFNNs)
2. Convolutional Neural Networks (CNNs)
3. Recurrent Neural Networks (RNNs)
4. Long Short-Term Memory Networks (LSTMs)
5. Gated Recurrent Units (GRU)
6. Autoencoder Neural Networks
7. Attention Mechanism
8. Generative Adversarial Networks (GANs)
9. Transformers
10. Deep Belief Networks (DBNs)

27. What is the difference between Shallow Networks and Deep Networks?

Deep networks and shallow networks are two types of artificial neural networks that can learn from data and perform tasks such as classification, regression, clustering, and generation.

• Shallow networks: A shallow network has a single hidden layer between the input and output layers, whereas a deep network has several hidden layers. Because they have fewer parameters, they are easier to train and less computationally expensive than deep networks. Shallow networks are appropriate for basic or low-complexity tasks where the input-output relationships are relatively straightforward and do not require extensive feature representation.
• Deep Networks: Deep networks, also known as deep neural networks, can be identified by the presence of many hidden layers between the input and output layers. The presence of multiple layers enables deep networks to learn hierarchical data representations, capturing detailed patterns and characteristics at different levels of abstraction. It has a higher capacity for feature extraction and can learn more complex and nuanced relationships in the data. It has given state-of-the-art results in many machine learning and AI tasks.

28. What is a Deep Learning framework?

A deep learning framework is a collection of software libraries and tools that provide programmers a better deep learning model development and training possibilities. It offers a high-level interface for creating and training deep neural networks in addition to lower-level abstractions for implementing special functions and topologies. TensorFlow, PyTorch, Keras, Caffe, and MXNet are a few of the well-known frameworks for deep learning.

29. What do you mean by vanishing or exploding gradient descent problem?

Deep neural networks experience the vanishing or exploding gradient descent problem when the gradients of the cost function with respect to the parameters of the model either become too small (vanishing) or too big (exploding) during training.

In the case of vanishing gradient descent, The adjustments to the weights and biases made during the backpropagation phase are no longer meaningful because of very small values. As a result, the model could perform poorly because it fails to pick up on key aspects of the data.

In the case of exploding gradient descent, The model surpasses its optimal levels and fails to converge to a reasonable solution because the updates to the weights and biases get too big.

Some of the techniques like Weight initialization, normalization methods, and careful selection of activation functions can be used to deal with these problems.

30. What is Gradient Clipping?

Gradient clipping is a technique used to prevent the exploding gradient problem during the training of deep neural networks. It involves rescaling the gradient when its norm exceeds a certain threshold. The idea is to clip the gradient, i.e., set a maximum value for the norm of the gradient, so that it does not become too large during the training process. This technique ensures that the gradients don’t become too large and prevent the model from diverging. Gradient clipping is commonly used in recurrent neural networks (RNNs) to prevent the exploding gradient problem.

Deep learning is a part of machine learning that is based on the artificial neural network with multiple layers to learn from and make predictions on data. An artificial neural network is based on the structure and working of the Biological neuron which is found in the brain.

This Deep Learning Interview Questions and answers cover all the basic to advanced Interview questions on Deep learning which will give you the confidence to stand in a tech interview. These Deep Learning Interview Questions are suggested questions by highly experienced data scientist professionals. So, this will definitely give you an edge in your Deep Learning Interview.

• Deep Learning Interview Questions for Freshers
• Deep Learning Interview Questions for Experienced