Deep Learning Interview Questions

If you are a deep learning engineer, you need to have a thorough understanding not only of coding but also of each of the components that go into creating a successful deep learning algorithm.

Example: "The primary function of a neural network is to receive a set of inputs, perform complex calculations and then use the output to solve the problem. A neural network is used for a range of applications. One example is classification; there are many classifiers available today, such as random forest, decision trees, support vector, logistic regression and so on, and of course neural networks."

Each sheet contains neurons called “nodes,” performing various operations. Neural Networks are used in deep learning algorithms like CNN, RNN, GAN, etc.

Following are some of the applications of deep learning :

Following are the advantages of neural networks :

Following are the disadvantages of neural networks :

Both shallow and deep networks are good enough and capable of approximating any function. But for the same level of accuracy, deeper networks can be much more efficient in terms of computation and number of parameters. Deeper networks can create deep representations. At every layer, the network learns a new, more abstract representation of the input.

Overfitting is the most common issue which occurs in deep learning. It usually occurs when a deep learning algorithm apprehends the sound of specific data. It also appears when the particular algorithm is well suitable for the data and shows up when the algorithm or model represents high variance and low bias.

Backpropagation can be divided into the following steps :

A perceptron is similar to the actual neuron in the human brain. It receives inputs from various entities and applies functions to these inputs, which transform them to be the output.

A perceptron is mainly used to perform binary classification where it sees an input, computes functions based on the weights of the input, and outputs the required transformation.

Machine Learning is powerful in a way that it is sufficient to solve most of the problems. However, Deep Learning gets an upper hand when it comes to working with data that has a large number of dimensions. With data that is large in size, a Deep Learning model can easily work with it as it is built to handle this.

Activation function translates the inputs into outputs. Activation function decides whether a neuron should be activated or not by calculating the weighted sum and further adding bias with it. The purpose of the activation function is to introduce non-linearity into the output of a neuron.

There can be many Activation functions like :

Data visualisation libraries help in understanding complex ideas by using visual elements such as graphs, charts, maps and more. The visualisation tools help you to recognise patterns, trends, outliers and more, making it possible to design your data according to the requirement. Popular data visualisation libraries include D3, React-Vis, Chart.js, vx, and more.

16 .

Overfitting is a type of modelling error which results in the failure to predict future observations effectively or fit additional data in the existing model. It occurs when a function is too closely fit to a limited set of data points and usually ends with more parameters than the data can accommodate. It is common for huge data sets to have some anomalies, so when this data is used for any kind of modelling, it can result in inaccuracies in the analysis.

Overfitting can be prevented by following a few methods namely :

Remove features :

Regularisation :

Ensembling :

The deep learning frameworks and tools are :

Single layer perceptron is the first proposed neural model created. The content of the local memory of the neuron consists of a vector of weights. The computation of a single layer perceptron is performed over the calculation of sum of the input vector each with the value multiplied by corresponding element of vector of the weights. The value which is displayed in the output will be the input of an activation function.

As in Neural Networks, MLPs have an input layer, a hidden layer, and an output layer. It has the same structure as a single layer perceptron with one or more hidden layers. A single layer perceptron can classify only linear separable classes with binary output (0,1), but MLP can classify nonlinear classes.

Except for the input layer, each node in the other layers uses a nonlinear activation function. This means the input layers, the data coming in, and the activation function is based upon all nodes and weights being added together, producing the output. MLP uses a supervised learning method called “backpropagation.” In backpropagation, the neural network calculates the error with the help of cost function. It propagates this error backward from where it came (adjusts the weights to train the model more accurately).

One of the most basic Deep Learning models is a Boltzmann Machine, resembling a simplified version of the Multi-Layer Perceptron. This model features a visible input layer and a hidden layer -- just a two-layer neural net that makes stochastic decisions as to whether a neuron should be on or off. Nodes are connected across layers, but no two nodes of the same layer are connected.

Also referred to as “**loss**” or “**error**,” cost function is a measure to evaluate how good your model’s performance is. It’s used to compute the error of the output layer during backpropagation. We push that error backward through the neural network and use that during the different training functions.

Gradient Descent is an optimal algorithm to minimize the cost function or to minimize an error. The aim is to find the local-global minima of a function. This determines the direction the model should take to reduce the error.

This network is made up of numerous tiny neural networks, rather than being a single network. The sub-networks combine to form a larger neural network, which operates independently to achieve a common goal. These networks are extremely useful for breaking down a large-small problem into smaller chunks and then solving it.

Convolutional Neural Networks are mostly used in computer vision. In contrast to fully linked layers in MLPs, one or more convolution layers extract simple characteristics from input by performing convolution operations in CNN models. Each layer is made up of nonlinear functions of weighted sums at various coordinates of spatially close subsets of the previous layer's outputs, allowing the weights to be reused.

The AI system learns to automatically extract the properties of these inputs to fulfill a specific task, such as picture classification, face identification, and image semantic segmentation, given a sequence of images or videos from the actual world.

Recurrent Neural Networks were created to solve the sequential input data time-series problem. RNN's input is made up of the current input and prior samples. As a result, the node connections create a directed graph. Furthermore, each neuron in an RNN has an internal memory that stores the information from previous samples' computations. Because of their superiority in processing data with a variable input length, RNN models are commonly employed in natural language processing (NLP). The goal of AI in this case is to create a system that can understand human-spoken natural languages, such as natural language modeling, word embedding, and machine translation.

Each successive layer in an RNN is made up of nonlinear functions of weighted sums of outputs and the preceding state. As a result, the basic unit of RNN is termed "cell," and each cell is made up of layers and a succession of cells that allow recurrent neural network models to be processed sequentially.

Fourier transform package is highly efficient for analyzing, maintaining, and managing a large databases. The software is created with a high-quality feature known as the special portrayal. One can effectively utilize it to generate real-time array data, which is extremely helpful for processing all categories of signals.

In neural networking, weight initialization is one of the essential factors. A bad weight initialization prevents a network from learning. On the other side, a good weight initialization helps in giving a quicker convergence and a better overall error. Biases can be initialized to zero. The standard rule for setting the weights is to be close to zero without being too small.

If the set of weights in the network is put to a zero, then all the neurons at each layer will start producing the same output and the same gradients during backpropagation.

As a result, the network cannot learn at all because there is no source of asymmetry between neurons. That is the reason why we need to add randomness to the weight initialization process.

Now, this can be answered in** two ways**. If you are on a phone interview, you cannot perform all the calculus in writing and show the interviewer. In such cases, it best to explain it as such:

Deep Learning goes right from the simplest data structures like lists to complicated ones like computation graphs.

The **softmax** function is used to calculate the probability distribution of the event over 'n' different events. One of the main advantages of using softmax is the output probabilities range. The range will be between 0 to 1, and the sum of all the probabilities will be equal to one. When the softmax function is used for multi-classification model, it returns the probabilities of each class, and the target class will have a high probability.

Swish is a new, self-gated activation function. Researchers at Google discovered the Swish function. According to their paper, it performs better than ReLU with a similar level of computational efficiency.

Autoencoder is an artificial neural network. It can learn representation for a set of data without any supervision. The network automatically learns by copying its input to the output; typically,internet representation consists of smaller dimensions than the input vector. As a result, they can learn efficient ways of representing the data. Autoencoder consists of two parts; an encoder tries to fit the inputs to the internal representation, and a decoder converts the internal state to the outputs.

Dropout is a cheap regulation technique used for reducing overfitting in neural networks. We randomly drop out a set of nodes at each training step. As a result, we create a different model for each training case, and all of these models share weights. It's a form of model averaging.

Tensors are nothing but a de facto for representing the data in deep learning. They are just multidimensional arrays, which allows us to represent the data having higher dimensions. In general, we deal with high dimensional data sets where dimensions refer to different features present in the data set.

A Boltzmann machine (also known as stochastic Hopfield network with hidden units) is a type of recurrent neural network. In a Boltzmann machine, nodes make binary decisions with some bias. Boltzmann machines can be strung together to create more sophisticated systems such as deep belief networks. Boltzmann Machines can be used to optimize the solution to a problem.

Some important points about Boltzmann Machine :

The loss function is used as a measure of accuracy to see if a neural network has learned accurately from the training data or not. This is done by comparing the training dataset to the testing dataset.

The loss function is a primary measure of the performance of the neural network. In Deep Learning, a good performing network will have a low loss function at all times when training.

Autoencoders are artificial neural networks that learn without any supervision. Here, these networks have the ability to automatically learn by mapping the inputs to the corresponding outputs.

Autoencoders, as the name suggests, consist of two entities :

* **Encoder :** Used to fit the input into an internal computation state

* **Decoder :** Used to convert the computational state back into the output

There are five main steps that are used to initialize and use the gradient descent algorithm :

Hyperparameters can be trained using four components as shown below :

Transfer learning is a learning technique that allows data scientists to use what they've learned from a previous machine learning model that was used for a similar task. The ability of humans to transfer their knowledge is used as an example in this learning. You can learn to operate other two-wheeled vehicles more simply if you learn to ride a bicycle. A model trained for autonomous automobile driving can also be used for autonomous truck driving. The features and weights can be used to train the new model, allowing it to be reused. When there is limited data, transfer learning works effectively for quickly training a model.

In the above image, the first diagram represents training a model from scratch while the second diagram represents using a model already trained on cats and dogs to classify the different class of vehicles, thereby representing transfer learning.

Following are the advantages of transfer learning :

A tensor is a multidimensional array that represents a generalization of vectors and matrices. It is one of the key data structures used in deep learning. Tensors are represented as n-dimensional arrays of base data types. The data type of each element in the Tensor is the same, and the data type is always known. It's possible that only a portion of the shape (that is, the number of dimensions and the size of each dimension) is known. Most operations yield fully-known tensors if their inputs are likewise fully known, however, in other circumstances, the shape of a tensor can only be determined at graph execution time.

The LSTM model is considered a special case of RNNs. The problems of vanishing gradients and exploding gradients we saw earlier are a disadvantage while using the plain RNN model.

In LSTMs, we add a forget gate, which is basically a memory unit that retains information that is retained across timesteps and discards the other information that is not needed. This also necessitates the need for input and output gates to include the results of the forget gate as well.

As you can see, the LSTM model can become quite complex. In order to still retain the functionality of retaining information across time and yet not make a too complex model, we need GRUs.

Basically, in GRUs, instead of having an additional Forget gate, we combine the input and Forget gates into a single Update Gate :

It is this reduction in the number of gates that makes GRU less complex and faster than LSTM.

An optimization algorithm that is used to minimize some function by repeatedly moving in the direction of steepest descent as specified by the negative of the gradient is known as gradient descent. It's an iteration algorithm, in every iteration algorithm, we compute the gradient of a cost function, concerning each parameter and update the parameter of the function via the following formula:

Where,

In machine learning, it is used to update the parameters of our model. Parameters represent the coefficients in linear regression and weights in neural networks.

Batch Gradient Descent :

Mini-batch Gradient Descent :

TensorFlow has numerous advantages, and some of them are as follows :

A Restricted Boltzmann Machine, or RBM for short, is an undirected graphical model that is popularly used in Deep Learning today. It is an algorithm that is used to perform:

Leaky ReLU, also called LReL, is used to manage a function to allow the passing of small-sized negative values if the input value to the network is less than zero.

With the use of sequential processing, programmers were up against :

The procedure of developing an assumption structure involves three specific actions.

An epoch is a terminology used in deep learning that refers to the number of passes the deep learning algorithm has made across the full training dataset. Batches are commonly used to group data sets (especially when the amount of data is very large). The term "iteration" refers to the process of running one batch through the model.

The number of epochs equals the number of iterations if the batch size is the entire training dataset. This is frequently not the case for practical reasons. Several epochs are used in the creation of many models.

where,

Yes, if the problem is represented by a linear equation, deep networks can be built using a linear function as the activation function for each layer. A problem that is a composition of linear functions, on the other hand, is a linear function, and there is nothing spectacular that can be accomplished by implementing a deep network because adding more nodes to the network will not boost the machine learning model's predictive capacity.

Backpropagation in Recurrent Neural Networks differ from that of Artificial Neural Networks in the sense that each node in Recurrent Neural Networks has an additional loop as shown in the following image:

This loop, in essence, incorporates a temporal component into the network. This allows for the capture of sequential information from data, which is impossible with a generic artificial neural network.

Following are the applications of autoencoders :

Capsules are a vector specifying the features of the object and its likelihood. These features can be any of the instantiation parameters like position, size, orientation, deformation, velocity, hue, texture and much more.

A capsule can also specify its attributes like angle and size so that it can represent the same generic information. Now, just like a neural network has layers of neurons, a capsule network can have layers of capsules.

The layer between the encoder and decoder, ie. the code is also known as Bottleneck. This is a well-designed approach to decide which aspects of observed data are relevant information and what aspects can be discarded.

It does this by balancing two criteria :

Autoencoder is a simple 3-layer neural network where output units are directly connected back to input units. Typically, the number of hidden units is much less than the number of visible ones. The task of training is to minimize an error or reconstruction, i.e. find the most efficient compact representation for input data.

RBM shares a similar idea, but it uses stochastic units with particular distribution instead of deterministic distribution. The task of training is to find out how these two sets of variables are actually connected to each other.

One aspect that distinguishes RBM from other autoencoders is that it has two biases. The hidden bias helps the RBM produce the activations on the forward pass, while The visible layer’s biases help the RBM learn the reconstructions on the backward pass.

Generative adversarial networks are used to achieve generative modeling in Deep Learning. It is an unsupervised task that involves the discovery of patterns in the input data to generate the output.

The generator is used to generate new examples, while the discriminator is used to classify the examples generated by the generator.

Generative adversarial networks are used for a variety of purposes. In the case of working with images, they have a high amount of traction and efficient working.

Image enhancement :

Image translation :

Another key difference is the amount of data required to train a model. Deep learning models require large amounts of data to train effectively, while traditional machine learning models can often be trained with smaller amounts of data.

Advertisement