Artificial Intelligence Interview Questions and Answers

• Machine Learning means using computers/software to analyze large amounts of data to find patterns or trends in them.
• Deep Learning uses multiple neural networks stacked on top of each other to perform pattern recognition.
• Artificial Intelligence refers to when machines actually think like humans.

Yes. A strong AI can be defined as an AI that has human-level intelligence. It is also known as AGI (artificial general intelligence), and it’s not possible yet. A weak AI is one that has only narrow intelligence. For example, a robot with limited vision, speech, dexterity, etc.

An AI system is a computer program designed to perform intelligent actions based on information received from its environment and processing inputs from sensors.

• Text translation – This is done by natural language processing, where the software is able to translate text into another language
• Image Recognition – this involves computers being able to identify objects within images
• Speech recognition – allows computers to understand what people say and make decisions accordingly

• Voice recognition
• Chatbots
• Automatic translation
• Video games

• Weak AI or Narrow AI: It provides specific capabilities without having any real understanding of the world around it. It does not have any reasoning ability or goal-directed behavior. Examples of such systems are chatbots, image recognizers, and search engines.
• Strong AI or General AI: It is capable of accomplishing tasks that require thinking, planning, and decision making. It can learn new things and adapt itself to changing environments. Example of such systems are self-driving cars, robots, smart homes, virtual assistants, and augmented reality.
• Expert AI or Super AI: It is a system that combines all three forms of AI i.e. knowledge, reasoning, and goal-driven behaviour. It can solve complex problems, handle uncertainty, deal with ethical issues and work collaboratively with others. Examples of such super AI are IBM Watson, DeepMind AlphaGo, and Google Brain.

• Reactive Machines
• Theory of Mind AI
• Limited Memory AI
• Self Aware AI
• Artificial Superhuman Intelligence (ASI)
• Artificial General Intelligence (AGI)
• Artificial Narrow Intelligence (ANI)

• Vending Machine: A machine that dispenses items based on preprogrammed instructions.
• Cash Register: A machine that accepts money and gives out change.
• ATM: A machine that allows you to withdraw or deposit money using a card.

Limited Memory AI is a type of artificial intelligence that has limited memory capacity. It can only store a certain amount of information at a time. This makes it difficult for the system to learn from past experiences. There are two types of limited memory systems:

• Incremental learning
• Reinforcement Learning

Some applications of Limited Memory AI are:

• Chatbots- Chatbot uses incremental learning because it learns new words without having to be trained each time.
• Machine Translation – Machine translation uses reinforcement learning because it needs to adapt itself to its environment.
• self-driving cars – Self-driving cars need to process data every second to navigate safely. So they use reinforcement learning.

Self -aware AI is an AI which has a high level of awareness and understanding of its own existence and the world around it. It is also able to think about things independently and come up with ideas. An example for Self Aware AI is Amazon Echo which is able to answer questions about anything by just asking.

Artificial superhuman intelligence is an AI that will surpass human intellect. An ASI would be able to achieve superhuman intelligence through neural networks, algorithms, genetic algorithms, quantum computing, and other emerging technologies.

Some examples of Artificial Superhuman Intelligence (ASI) are:

• IBM SyNAPSE
• Intel’s Loihi

AI general intelligence is an AI that is capable of performing complex tasks which require more than simple calculation or pattern recognition. AGI is also referred to as strong AI. An example of AGI is the Pillo robot created for purpose of answering questions related to health.

Artificial narrow intelligence is an AI that does one specific task better than any human being. An ANI is also called a narrow AI. An example of ANI is Siri which is a virtual assistant that helps users in completing various tasks like scheduling meetings, making reservations, finding directions, sending messages, etc.

• Lisp
• Prolog
• Java
• C++
• Python
• SQL
• Matlab
• R
• Ruby
• JavaScript

Machine Learning Application Implementation:

• Implementing Machine Learning Algorithm on your code base.
• Define training set & test set
• Create a model based on your coding
• Test the model
• Deploy the model
• Use the model to do predictions
• Evaluate the performance of your model
• Monitor the results
• Make sure the model works correctly
• Make changes if needed
• Repeat steps 1-8

Choosing an Algorithm for a Problem:

• Understand what you want the solution to look like
• Choose the best algorithm based on the requirement.
• If you have multiple requirements, then try each of them till you get the best result.

Example: You want to predict whether a person will buy product X or Y. Then, based on the features you have chosen, you could go for Logistic Regression, Random Forest, Support Vector Machines. etc.

Deep Learning Frameworks:

• Tensorflow
• Caffe
• Theano
• Keras
• Pytorch
• Torch
• MXNet
• OpenCV
• Sklearn

Tower Of Hanoi is used to build up an AI algorithm because it can help in solving problems of arranging blocks in different positions.

DNN is used because it’s very good at learning patterns from data. It has been found that by using DNN, models can learn much faster. Also, when compared to traditional AI techniques (such as decision trees), DNN models tend to perform better on large datasets.

Advantages of Deep Learning:

• Can handle big data sets easily
• Can extract useful features directly from raw information
• Has the ability to learn new things without being explicitly programmed
• Can work well even if there is no prior knowledge about the data
• More robust and reliable6) Easier to debug

Q-learning is a type of reinforcement learning where the agent tries to maximize its cumulative reward by selecting actions that leads to higher rewards.

Bayes Inference:

• Bayes rule is a probability theorem that says that “If A is true, B must be true with some degree of certainty”.
• Bayes Rule states that the posterior probability of hypothesis A given evidence E is equal to the conditional probability of A given E times the prior probability of A.

Generative Model:

• Generate something out of nothing
• Creates an object from existing objects
• Produces new information or content from existing information

Markov Chain:

• A random process whose future behavior depends only on the present state of the system.
• A mathematical description of a stochastic sequence
• A way of describing systems that operate according to simple rules.
• An example of a Markov chain would be weather forecasting – if the current conditions are known, the forecast can be made more accurately than if the conditions were not known.
• A Markov chain is also called a Markov Process.

Neural Network:

• A network of neurons is connected by synapses.
• A representation of how the brain processes information.
• A computational device that mimics the human brain.
• A machine that learns from experience.

Turing Test:

• A test meant to determine whether machines can think like humans.
• A test is used to determine whether computers have achieved human-level intelligence.
• Alan Turing devised a hypothetical test to determine whether machines could think.

Feature Engineering:

• Feature extraction is a method of transforming raw data into a form suitable for classification.
• Feature Extraction is a technique of creating a set of characteristics or features from input data.
• The main purpose of this step is to preprocess the original dataset so that we can use algorithms such as KNN or SVM to classify the data.

Expert System:

• A computer program that simulates the decision making of an intelligent person
• It is based on artificial intelligence.
• It is a tool that uses human knowledge to make decisions

Characteristics of Expert Systems:

• It is easy to understand and implement
• It has been proven effective in many applications, especially those related to medicine and finance.
• It is user friendly
• It provides recommendations based on experience

An expert system helps us in the following ways:

• It can predict what might happen next
• It can suggest what should be done next
• It can explain why things happened the way they did
• It can suggest alternatives
• It can give advice
• It can evaluate options
• It can compare different courses of action
• It can recommend the best course of action
• It can find patterns
• It can create models and simulate
• It can learn and diagnose
• It can control, assist and teach

A* Search Method:

• It is widely used as a heuristic search method
• It is one of the most commonly used search methods for pathfinding
• It is used extensively in games

Breadth-first search Algorithm:

• It starts with all nodes at the same distance from the root node.
• It then expands only nodes that are adjacent to unexplored nodes.
• If two paths reach the same goal state, the one with the shortest distance to the goal is chosen.

Depth-First Search Algorithm:

• It searches through every possible branch of the tree.
• At each level of the tree, it checks if the current node is already explored.
• If not, it explores it by recursively calling itself on the child nodes.
• When it reaches the end of the tree, it returns the result.

Bidirectional Search Algorithm:

It is also known as BFS(Backward First Search), which works like DFS, but instead of going forward, it goes backward.

Reinforcement learning (RL):

• It is an approach in AI that involves reward feedback
• The learner gets rewarded for performing well
• The goal is to train the agent so that it behaves better over time
• RL algorithms try to maximize rewards

• An agent is playing a game against another agent
• Each turn, both agents have access to information about their opponent’s actions
• Based on this information, the agent chooses whether to act or wait
• If it makes the right choice, it receives a reward
• This process continues until the game ends

Alpha-beta Pruning:

• It reduces the size of the search space.
• It means reducing the number of branches that must be checked.
• It is based on the idea that there are more likely states than non-likely ones.
• It saves computational power by checking fewer possibilities.
• It is used in planning problems such as chess and Goes.

Lookahead search algorithm:

• It keeps track of its own history
• It looks ahead to see what will happen next
• It uses information from the past to make decisions now
• It does not require any changes to the original program

Markov’s decision Process:

• A decision made by a system can depend on either previous decisions or on external factors
• t is a model used to describe systems that change over time
• In a Markov Decision Process, the probability of a transition depends solely on the present state and not on the past
• It is widely used in artificial intelligence because it models human behavior

Example of Markov’s decision process:

• Suppose you are looking at an exam with 100 questions
• You know that 50% of them are easy and 50% of them are hard
• From your experience, you remember that the first 10 are always easy, and the rest are all difficult
• So, you can predict the difficulty of the remaining 90 questions easily
• Now, suppose if you get an answer wrong, you lose one mark for each question you got correct
• Also, if you get an answer right, you gain one mark per question
• So, knowing these two facts, you should be able to guess the average mark you will get.

Stochastic search:

• It randomly selects one action at a time
• It may select the same action multiple times
• It tries all other available options before selecting one
• It starts with a random policy and slowly improves it based on experience.

Fuzzy Logic:

• Fuzzy sets are mathematical theories that represent vague concepts like “small”, “medium”, “large,” etc., and membership functions are values assigned to input variables (for example, the distance between cities).
• Fuzzy logic provides a way to deal with imprecision and uncertainty.

Applications of fuzzy logic:

• It was originally developed for use in linguistic programs
• It has been applied in many areas, including engineering, medicine, electronics, management, and business
• It helps solve complex problems involving judgment
• It allows computers to work with incomplete data and ambiguous situations
• It has also been used in robotics and neural networks

Reward Maximization:

• Reward maximization is an attempt to maximize overall performance, regardless of cost.
• The agent receives a positive reward when it achieves its goal and a negative reward when it fails to do so.
• This is done by adjusting actions in order to achieve the best possible outcome.

FOPL:

First Order Predicate Logic is a type of propositional calculus where every proposition must contain only atomic propositions such as ‘xy’ or ‘p’. It consists of three parts:

• Formal part- It represents all the logical rules of inference
• Material part- It is a set of sentences in the English language
• Semantic part- It represents the meaning of the sentence.

Exploitation and Exploration Trade-Off:

• In reinforcement learning, the agent needs to balance between exploiting what it knows about the environment and exploring new possibilities
• Exploiting means taking the most likely path towards success, while exploring means moving away from the known paths and trying something else
• By balancing between exploitation and exploration, the agent learns faster and avoids getting stuck in local minima.

Inductive machine learning:

• Inductive reasoning is the process of inferring general patterns from specific examples
• It involves identifying useful information from large amounts of unstructured data
• It is used to build systems that learn from their experiences rather than being programmed explicitly
• Examples include natural language processing, speech recognition, and visual object detection

Deductive learning:

• A deduction is a form of reasoning that begins with certain assumptions and then uses them to make conclusions.
• Deductive reasoning is based on previously established facts and principles, which can be expressed using laws, axioms, propositions, rules, algorithms, and other forms of derivation.
• It is often used to develop mathematical models, computer programs, theories, and proofs.
• Examples include logic programming, theorem proving, and automated deduction.

Abductive ML:

Abduction is the process of making inferences to explain why some phenomena occur.

• It is based on prior knowledge and observations, and aims at explaining unknowns through hypotheses.
• An example would be a medical diagnosis where the doctor makes an educated guess about the cause of illness.
• Abductive ML is also related to abduction in epistemology, i.e., the theory of knowledge.

Algorithm Techniques:

• Supervised Algorithms: They are trained by given labeled training data.
• Unsupervised Algorithms: These find patterns without any labels.
• Reinforcement Learning: This is an important category of AI because it deals with decision-making under uncertainty and imperfect information.
• Reinforcement learning combines supervised and reinforcement methods.
• Q-Learning: It is one of the most popular reinforcement learning algorithms.
• Policy Gradient Methods: It is another commonly used reinforcement learning technique.

Parametric Model:

• A parametric model is a function whose output depends on several parameters called “parameters” or “features”. For example, we could say that the height of a person depends on his/her age (that’s a parametric model).
• Parametric models are used in applications like regression analysis, pattern recognition, and time series prediction.

Non-parametric model:

• A non-parametric model has no predefined structure for describing its input-output relationship. The number of free parameters in a model may change during the course of the learning process. In contrast, a parametric model assumes a fixed structure for representing the relation between inputs and outputs.
• Non-parametric models are used when there is little or no underlying structure in the problem domain. For example, they are widely used in machine translation or text summarization tasks.

Model Parameters:

• In statistics, a parameter refers to a constant value that affects the shape of distribution or curve.
• In AI, a model parameter is usually a constant value that affects how the model behaves.

For example, consider a linear regression model. We can think of this as having two parameters: slope and intercept. If these values were changed, then the resulting line would be different.

Hyperparameter Parameters:

• Hyperparameters are variables that determine the behavior of a model. Examples include the size of your dataset, the type of neural network architecture, and the learning rate.
• In general, hyperparameters are not considered part of the model itself; rather, they control the model’s behavior.

Hyperparameters in DNN:

• In deep learning, hyperparameters control the complexity of the model. For instance, if you have more layers, you will need less neurons per layer.
• When designing a deep learning system, you must choose appropriate hyperparameters to ensure that the learning converges quickly and that the final solution does not overfit the training set.

Algorithms for Hyperparameter Optimization:

• There are three main categories of algorithms for optimizing hyperparameters: grid search, random search, and Bayesian optimization. Each category has its own strengths and weaknesses.
• Grid Search: This algorithm evaluates all possible combinations of hyperparameter settings and selects those with the best performance. However, it takes a lot of computing power because it requires evaluating every combination of hyperparameters.
• Random Search: This algorithm randomly samples from some space of valid hyperparameter values and uses them to train the model. It is fast but often produces suboptimal results.
• Bayesian Optimization: This algorithm iterates through a sequence of hyperparameter settings, updating the posterior probability of each setting as data arrives. It is able to sample effectively even though we don’t know exactly where to find good solutions.

Naive Bayes Algorithm:

• The Naive Bayes classifier learns by building statistical relationships between features. For example, it might learn that a feature X is always associated with a certain output Y.
• To make predictions, the Naive Bayes classifier applies a simple mathematical formula that combines information about a given feature into an overall estimate.

Natural Language Processing (NLP):

• Natural language processing is an umbrella term for any computer science techniques that involve analyzing natural languages such as English.
• One of the main reasons why text analysis became so popular was due to advances in machine learning and artificial intelligence. These technologies allowed researchers to analyze large datasets and extract useful insights.

Text Mining:

• Text mining refers to the process of extracting meaningful patterns out of unstructured or semi-structured documents.
• For instance, text mining can be used to identify key phrases in scientific articles.

Components of NLP:

• Natural Language Understanding: This component analyzes user input and extracts meaning from text. For example, this component allows computers to understand human speech.
• Natural Language Generation: This component generates text based on existing knowledge and context. For example, this could include generating captions for images.

Data Overfitting:

• Data overfitting occurs when models fit training data well but fail to generalize to unseen data.
• In Machine Learning, overfitting happens when models perform better on training data than they do on test data.

Fixing data overfitting:

• Regularization: This method modifies the parameters of the model such that the model fits less well with training data.
• Feature Selection: This technique removes unnecessary features from the dataset before modeling.

Stemming:

• Stemming involves transforming words into their base forms. It is usually performed using the Porter stemmer algorithm.
• A stemmer transforms words like “go” and “went” into “goe” and “gwnt”, respectively.

Lemmatization:

• Lemmatization is a process that converts inflected words into lemmas. For example, converting “walked” to “walking”.
• When dealing with texts containing many inflections, lemmatizing becomes necessary.

Extracting Named Entities:

• Named entity recognition (NER) refers to the task of identifying named entities in a piece of text.
• Named entities include people, places, organizations, dates, times, numbers, currencies, units of measurement, etc.

Backpropagation Algorithm:

• Backpropagation algorithms are supervised learning methods where we have access to labeled output values and use those labels to update weights in our network layer by layer until we reach the last hidden layer.
• The backpropagation procedure generally works through three phases. First of all, it calculates the gradient of the loss function. Then it updates each weight according to how much its gradient has changed. Finally, it uses the new weights to calculate the error on the outputs.

Gradient Descent:

• Gradient descent is an optimization algorithm that iteratively moves towards a local minimum of a cost function. It uses the steepest descent to move downhill from one point to another.
• It has been used extensively in solving problems involving non-convex functions.

Hebbian Rule:

• Hebbian Rule states that if two neurons are connected together, the information will flow between them.
• It comes from a famous psychologist Donald Hebb who proposed it as a learning rule in 1949.

Feature Selection Techniques:

• Selecting variables based on statistical significance can result in selecting redundant or irrelevant variables.
• Using an automated variable selection technique reduces the number of features and enhances the accuracy of prediction.
• Correlation analysis selects among correlated features.

Minimax algorithm is an algorithm for finding the best strategy in games such as chess or checkers. In this case, the goal is to minimize the maximum possible score while maximizing the lowest possible score.

Key points about minimax algorithm

• It should be used when you want to find the best solution without having any idea about what the actual best solution is.
• It requires a very large memory space compared to other search algorithms.
• The problem must be defined in terms of a game tree.

Minimax algorithm finds optimal strategies by searching the entire game tree. This approach can be computationally expensive but often yields good results.

Dimensionality Reduction Methods: Dimensionality reduction refers to the process of mapping data points into lower-dimensional spaces so that they can be visualized more easily. Some common dimensionality reduction techniques include principal component analysis (PCA), linear discriminant analysis (LDA), kernel PCA, ISOMAP, Locally Linear Embedding (LLE), etc.

Sequential Supervised Learning:

Sequential supervised learning uses training examples one after another without any preprocessing. The classifier learns from the previous example by making predictions about the current example. Examples may also be presented in batches. Batch processing means presenting all training examples at once instead of one at a time.

In sliding-window methods, we slide a window over the input sequence. The window slides over the input in steps. At each step, we make a prediction based on the training data available up to that point. After observing the test example, if it matches our prediction then we update the model parameters accordingly. We repeat these two steps until all the training examples have been considered.

We use recurrent sliding windows because we want to model the temporal dependencies among features. That is, we would like to learn how the features change over time. For example, we might expect that the feature “age” will increase over time. To accomplish this, we need to keep track of the history of observations. We do this by keeping a running window over the input sequence and updating the model parameters based on the new observation.

Hidden Markov Models are probabilistic models that describe sequences of events. They assume that there exists some underlying state machine and that when an event occurs, the system changes its internal state according to some transition probability distribution. An HMM consists of three components: states, transitions, and emissions. A state encodes the possible values of a variable; a transition specifies the probability of changing state; an emission associates a probability distribution with each state.

Maximum Entropy Markov models are similar to Hidden Markov Models, but they differ in their assumptions about the structure of the underlying state space. In particular, HMMs assume that the states are discrete while maximum entropy Markov models assume that the states are continuous.

Conditional Random Fields (CRFs) are probabilistic graphical models for structured output labeling problems. Like other statistical models, CRFs represent the conditional independence relationships between variables given observed data. However, unlike traditional statistical models, CRFs can capture nonlinearities in these relationships. This allows them to perform better than linear statistical models in many domains, including speech recognition, image segmentation, and protein folding.

Graph Transformer Networks are a class of deep neural network architectures that transform graphs into another graph representation. These transformations are designed such that information flows from one layer to another. The first layer acts as an attention mechanism, selecting nodes or edges of interest. The second layer performs a transformation operation upon those selected nodes, e.g., edge contraction, node addition, etc. The third layer then combines all transformed nodes together using a learned function to produce a final graph.

Activation functions are mathematical functions used within artificial intelligence algorithms to apply non-linearity to the inputs. Common examples include sigmoid (a smooth approximation of the step function), tanh (hyperbolic tangent), and rectified linear units (ReLU).

A vanishing gradient refers to the situation where the gradients become very small over time. This causes training to slow down or even stop altogether. One way to prevent this is by normalizing the weights so that the magnitude of the weight vector remains constant.

An example of a vanishing gradient would be if you have two layers of perceptrons and both have a large number of hidden units (e.g. 100,000 units per layer). Then the weights on the first layer will tend to go to zero because the input to the second layer has become too large. If we normalize the weights then the second layer will still work fine.

Long Short Term Memory (LSTM) is a recurrent neural network architecture used for tasks like machine translation and handwriting recognition. It’s made up of three parts: memory cells, gates, and connections. The memory cell contains a set of internal values that are updated based on new incoming information. Each gate controls how much information passes through to the next layer. Finally, there is a connection matrix that connects each cell to every neuron in the previous layer.

The key components of a Long Short Term Memory (LTSM) model are Gates, Tanh(x), and Sigmoid(x).

Long short term memories operate in three stages. First, they read the last value of the cell. Second, they decide whether to update themselves or not. Third, they pass their decision along to the next unit.

A recurrent Neural Network (RNN) is a type of neural network algorithm capable of learning long-term dependencies automatically. A typical feedforward neural network consists of multiple layers of simple processing elements. In contrast, an RNN can consist of many more levels of processing elements arranged in a feedback loop.

CNN stands for Convolutional Neural Network whereas RNN stands for Recurrent Neural Network. Both of them are types of neural networks but use different approaches to learn features. CNN uses convolutions while RNN uses gated recurrence.

Word Embeddings is a technique used to represent words as vectors in a high dimensional space. Word embeddings are trained using unsupervised methods, meaning they do not require annotated data.

Convolutional Neural Network (CNN) is one of the most powerful techniques for image classification. As the name suggests it applies convolutions to images. There are several variants of CNN including AlexNet, VGGNet, etc.

There are three main categories of RNN namely; Simple RNN, Gated RNN, and LSTM.

Autoencoders are a class of deep neural network algorithms that seek to reconstruct their inputs from outputs. Autoencoders are often used in unsupervised learning where no labels are available.

• Data denoising
• Dimensionality reduction
• Image reconstruction
• Image colorization

Tensorflow is an open-source software library for Machine Learning developed by Google. It provides a programming interface that enables users to build, train, deploy, and manage models quickly.

Advantages of Tensorflow include easy training, portability, and good performance.

Cost functions measure the error between predicted and actual output. Cost functions help us to minimize errors such as cross-entropy loss.

Activation functions are non-linear transformations applied to the input before passing it to the next layer. There are two major types of activation functions: Sigmoid and Tanh.

Dense layers connect all units to the same number of units. They have no hidden layers and therefore are computationally efficient.

Dropout is a method to prevent overfitting. The idea behind this is to randomly set a particular neuron to zero at each iteration during training. This causes the network to become less sensitive to individual weights and forces the model to discover general patterns instead of memorizing individual weights.

Batch Normalisation is a regularization technique used to improve the convergence speed of training neural nets. It helps reduce internal covariate shift and thus makes training more stable.

Dropout layers can be thought of as a special case of the batch norm where we only drop random values within a mini-batch. However, unlike batch norm, dropout allows us to control how much we want to lose information with respect to our target.

• Rectified linear unit (ReLU) is a type of activation function invented by Dr. Geoffrey Hinton in 2012. ReLUs are commonly used because they do not require any additional parameters compared to other activation functions.
• In simple terms, ReLU works by adding 1 to the input and then applying the sigmoid function to the result. This transformation is called rectification.