Artificial Intelligence Interview Question Part 1

What do you understand by Artificial Intelligence?

Artificial intelligence is computer science technology that emphasizes creating intelligent machine that can mimic human behavior. Here Intelligent machines can be defined as the machine that can behave like a human, think like a human, and also capable of decision making. It is made up of two words, “Artificial” and “Intelligence,” which means the “man-made thinking ability.”

With artificial intelligence, we do not need to pre-program the machine to perform a task; instead, we can create a machine with the programmed algorithms, and it can work on its own.

Why do we need Artificial Intelligence?

The goal of Artificial intelligence is to create intelligent machines that can mimic human behavior. We need AI for today’s world to solve complex problems, make our lives more smoothly by automating the routine work, saving the manpower, and to perform many more other tasks.

Give some real-world applications of AI.

There are various real-world applications of AI, and some of them are given below:

• Google Search Engine: When we start writing something on the google search engine, we immediately get the relevant recommendations from google, and this is because of different AI technologies.
• Ridesharing Applications: Different ride-sharing applications such as Uber uses AI and machine learning to determine the type of ride, minimize the time once the car is hailed by the user, price of the ride, etc.
• Spam Filters in Email: The AI is also used for email spam filtering so that you can get the important and relevant emails only in your inbox. As per the studies, Gmail successfully filters 99.9% of spam mails.
• Social Networking: Different social networking sites such as Facebook, Instagram, Pinterest, etc., use the AI technology for different purposes such as face recognition and friend suggestions, when you upload a photograph on Facebook, understanding the contextual meaning of an emoji in Instagram, and so on.
• Product recommendations: When we search for a product on Amazon, we get the recommendation for similar products, and this is because of different ML algorithms. Similarly, on Netflix, we get personalized recommendations for movies and web series.

Artificial Intelligence Interview Question

How Artificial intelligence, Machine Learning, and Deep Learning differ from each other?

The difference between AI, ML, and Deep Learning is given in the below table:

Artificial Intelligence	Machine Learning	Deep Learning
The term Artificial intelligence was first coined in the year 1956 by John McCarthy.	The term ML was first coined in the year 1959 by Arthur Samuel.	The term DL was first coined in the year 2000 Igor Aizenberg.
It is a technology that is used to create intelligent machines that can mimic human behavior.	It is a subset of AI that learns from past data and experiences.	It is the subset of machine learning and AI that is inspired by the human brain cells, called neurons, and imitates the working of the human brain.
AI completely deals with structured, semi-structured data.	ML deals with structured and semi-structured data.	Deep learning deals with structured and unstructured data.
It requires a huge amount of data to work.	It can work with less amount of data compared to deep learning and AI.	It requires a huge amount of the data compared to the ML.
The goal of AI is to enable the machine to think without any human intervention.	The goal of ML is to enable the machine to learn from past experiences.	The goal of deep learning is to solve the complex problems as the human brain does, using various algorithms.

What are the types of AI?

Artificial intelligence can be divided into different types on the basis of capabilities and functionalities.

Based on Capabilities:

• Weak AI or Narrow AI: Weak AI is capable of performing some dedicated tasks with intelligence. Siri is an example of Weak AI.
• General AI: The intelligent machines that can perform any intellectual task with efficiency as a human.
• Strong AI: It is the hypothetical concept that involves the machine that will be better than humans and will surpass human intelligence.

Based on Functionalities:

• Reactive Machines: Purely reactive machines are the basic types of AI. These focus on the present actions and cannot store the previous actions. Example: Deep Blue.
• Limited Memory: As its name suggests, it can store the past data or experience for the limited duration. The self-driving car is an example of such AI types.
• Theory of Mind: It is the advanced AI that is capable of understanding human emotions, people, etc., in the real world.
• Self-Awareness: Self Awareness AI is the future of Artificial Intelligence that will have their own consciousness, emotions, similar to humans.

What are the types of Machine Learning?

Machine Learning can be mainly divided into three types:

1. Supervised Learning: Supervised learning is a type of Machine learning in which the machine needs external supervision to learn from data. The supervised learning models are trained using the labeled dataset. Regression and Classification are the two main problems that can be solved with Supervised Machine Learning.
2. Unsupervised Learning: It is a type of machine learning in which the machine does not need any external supervision to learn from the data, hence called unsupervised learning. The unsupervised models can be trained using the unlabeled dataset. These are used to solve the Association and Clustering problems.
3. Reinforcement Learning: In Reinforcement learning, an agent interacts with its environment by producing actions, and learn with the help of feedback. The feedback is given to the agent in the form of rewards, such as for each good action, he gets a positive reward, and for each bad action, he gets a negative reward. There is no supervision provided to the agent. Q-Learning algorithm is used in reinforcement learning.

Advance Artificial Intelligence Interview Question

Explain the term “Q-Learning.”

Q-learning is a popular algorithm used in reinforcement learning. It is based on the Bellman equation. In this algorithm, the agent tries to learn the policies that can provide the best actions to perform for maximining the rewards under particular circumstances. The agent learns these optimal policies from past experiences.

In Q-learning, the Q is used to represent the quality of the actions at each state, and the goal of the agent is to maximize the value of Q.

What is Deep Learning, and how is it used in real-world?

Deep learning is a subset of Machine learning that mimics the working of the human brain. It is inspired by the human brain cells, called neurons, and works on the concept of neural networks to solve complex real-world problems. It is also known as the deep neural network or deep neural learning.

Some real-world applications of deep learning are:

• Adding different colors to the black & white images
• Computer vision
• Text generation
• Deep-Learning Robots, etc.

Which programming language is used for AI?

Below are the top five programming languages that are widely used for the development of Artificial Intelligence:

• Python
• Java
• Lisp
• R
• Prolog

Among the above five languages, Python is the most used language for AI development due to its simplicity and availability of lots of libraries, such as Numpy, Pandas, etc.

Artificial Intelligence Interview Question

What is the intelligent agent in AI, and where are they used?

The intelligent agent can be any autonomous entity that perceives its environment through the sensors and act on it using the actuators for achieving its goal.

These Intelligent agents in AI are used in the following applications:

• Information Access and Navigations such as Search Engine
• Repetitive Activities
• Domain Experts
• Chatbots, etc.

How is machine learning related to AI?

Machine learning is a subset or subfield of Artificial intelligence. It is a way of achieving AI. As both are the two different concepts and the relation between both can be understood as “AI uses different Machine learning algorithms and concepts to solve the complex problems.”  

What is Markov’s Decision process?

The solution for a reinforcement learning problem can be achieved using the Markov decision process or MDP. Hence, MDP is used to formalize the RL problem. It can be said as the mathematical approach to solve a reinforcement learning problem. The main aim of this process is to gain maximum positive rewards by choosing the optimum policy.

MDP has four elements, which are:

• A set of finite states S
• A set of finite actions A
• Rewards
• Policy P_a

In this process, the agent performs an action A to take a transition from state S1 to S2 or from the start state to the end state, and while doing these actions, the agent gets some rewards. The series of actions taken by the agent can be defined as the policy.

Advance Artificial Intelligence Interview Question

What do you understand by the reward maximization?

Reward maximization term is used in reinforcement learning, and which is a goal of the reinforcement learning agent. In RL, a reward is a positive feedback by taking action for a transition from one state to another. If the agent performs a good action by applying optimal policies, he gets a reward, and if he performs a bad action, one reward is subtracted. The goal of the agent is to maximize these rewards by applying optimal policies, which is termed as reward maximization.

What are parametric and non-parametric model?

In machine learning, there are mainly two types of models, Parametric and Non-parametric. Here parameters are the predictor variables that are used to build the machine learning model. The explanation of these models is given below:

Parametric Model: The parametric models use a fixed number of the parameters to create the ML model. It considers strong assumptions about the data. The examples of the parametric models are Linear regression, Logistic Regression, Naïve Bayes, Perceptron, etc.

Non-Parametric Model: The non-parametric model uses flexible numbers of parameters. It considers a few assumptions about the data. These models are good for higher data and no prior knowledge. The examples of the non-parametric models are Decision Tree, K-Nearest Neighbour, SVM with Gaussian kernels, etc.

What do you understand by the hyperparameter?

In machine learning, hyperparameter is the parameters that determine and control the complete training process. The examples of these parameters are Learning rate, Hidden Layers, Hidden units, Activation functions, etc. These parameters are external from the model. The selection of good hyperparameters makes a better algorithm.

The HMM is used in various applications such as reinforcement learning, temporal pattern recognition, etc.

Artificial Intelligence Interview Question

What is Strong AI, and how is it different from the Weak AI?

Strong AI: Strong AI is about creating real intelligence artificially, which means a human-made intelligence that has sentiments, self-awareness, and emotions similar to humans. It is still an assumption that has a concept of building AI agents with thinking, reasoning, and decision-making capabilities similar to humans.

Weak AI: Weak AI is the current development stage of artificial intelligence that deals with the creation of intelligent agents and machines that can help humans and solve real-world complex problems. Siri and Alexa are examples of Weak AI programs.

Give a brief introduction to the Turing test in AI?

Turing test is one of the popular intelligence tests in Artificial intelligence. The Turing test was introduced by Alan Turing in the year 1950. It is a test to determine that if a machine can think like a human or not. According to this test, a computer can only be said to be intelligent if it can mimic human responses under some particular conditions.

In this test, three players are involved, the first player is a computer, the second player is a human responder, and the third player is the human interrogator, and the interrogator needs to find which response is from the machine on the basis of questions and answers.

What is overfitting? How can it be overcome in Machine Learning?

When the machine learning algorithm tries to capture all the data points, and hence, as a result, captures noise also, then overfitting occurs in the model. Due to this overfitting issue, the algorithm shows the low bias, but the high variance in the output. Overfitting is one of the main issues in machine learning.

Methods to avoid Overfitting in ML:

• Cross-Validation
• Training With more data
• Regularization
• Removing Unnecessary Features
• Early Stopping the training.

Advance Artificial Intelligence Interview Question

Tell one technique to avoid overfitting in neural networks?

Dropout Technique: The dropout technique is one of the popular techniques to avoid overfitting in the neural network models. It is the regularization technique, in which the randomly selected neurons are dropped during training.

What is NLP? What are the various components of NLP?

NLP stands for Natural Language Processing, which is a branch of artificial intelligence. It enables machines to understand, interpret, and manipulate the human language.

Components of NLP:

There are mainly two components of Natural Language processing, which are given below:

1. Natural Language Understanding (NLU):
   It involves the below tasks:
   • To map the input to useful representations.
   • To analyze the different aspects of the language.
2. Natural Language Generation (NLG)
   • Text Planning
   • Sentence Planning
   • Text Realization

What are the different components of the Expert System?

An expert system mainly contains three components:

1. User Interface: It enables a user to interact or communicate with the expert system to find the solution for a problem.
2. Inference Engine: It is called the main processing unit or brain of the expert system. It applies different inference rules to the knowledge base to draw a conclusion from it. The system extracts the information from the KB with the help of an inference engine.
3. Knowledge Base: The knowledge base is a type of storage area that stores the domain-specific and high-quality knowledge.

Artificial Intelligence Interview Question

What is the use of computer vision in AI?

Computer vision is a field of Artificial Intelligence that is used to train the computers so that they can interpret and obtain information from the visual world such as images. Hence, computer vision uses AI technology to solve complex problems such as image processing, object detections, etc.

Explain the minimax algorithm along with the different terms.

Minimax algorithm is a backtracking algorithm used for decision making in game theory. This algorithm provides the optimal moves for a player by assuming that another player is also playing optimally.

This algorithm is based on two players, one is called MAX, and the other is called the MIN.

Following terminologies that are used in the Minimax Algorithm:

• Game tree: A tree structure with all possible moves.
• Initial State: The initial state of the board.
• Terminal State: Position of the board where the game finishes.
• Utility Function: The function that assigns a numeric value for the outcome of the game.

What is game theory? How is it important in AI?

Game theory is the logical and scientific study that forms a model of the possible interactions between two or more rational players. Here rational means that each player thinks that others are just as rational and have the same level of knowledge and understanding. In the game theory, players deal with the given set of options in a multi-agent situation, it means the choice of one player affects the choice of the other or opponent players.

Game theory and AI are much related and useful to each other. In AI, the game theory is widely used to enable some of the key capabilities required in the multi-agent environment, in which multiple agents try to interact with each other to achieve a goal.

Different popular games such as Poker, Chess, etc., are the logical games with the specified rules. To play these games online or digitally, such as on Mobile, laptop, etc., one has to create algorithms for such games. And these algorithms are applied with the help of artificial intelligence.

Advance Artificial Intelligence Interview Question

What are some misconceptions about AI?

There are lots of misconceptions about artificial intelligence since starting its evolution. Some of these misconceptions are given below:

• AI does not require humans: The first misconception about AI is that it does not require human. But in reality, each AI-based system is somewhere dependent on humans and will remain. Such as it requires human gathered data to learn about the data.
• AI is dangerous for humans: AI is not inherently dangerous for humans, and still, it has not reached the super AI or strong AI, which is more intelligent than humans. Any powerful technology cannot be harmful if it is not misused.
• AI has reached its peak stage: Still, we are so far away from the peak stage of the AI. It will take a very long journey to reach its peak.
• AI will take your job: It is one of the biggest confusions that AI will take most of the jobs, but in reality, it is giving us more opportunities for new jobs.
• AI is new technology: Although some people think that it is a new technology, this technology actually first thought in the year 1840 through an English newspaper.

What are the eigenvalues and eigenvectors?

Eigenvectors and eigenvalues are the two main concepts of Linear algebra.

Eigenvectors are unit vectors that have a magnitude equal to 1.0.

Eigenvalues are the coefficients that are applied to the eigenvectors, or these are the magnitude by which the eigenvector is scaled.

What is an Artificial neural network? Name some commonly used Artificial Neural networks.

Artificial neural networks are the statistical model inspired by the functioning of human brain cells called neurons. These neural networks include various AI technologies such as deep learning and machine learning.

An Artificial neural network or ANN consists of multiple layers, including the Input layer, Output Layer, and hidden layers.

ANN, with the help of various deep learning techniques, is the AI tools to solve various complex problems like pattern recognition, facial recognition, and so on.

Some commonly used Artificial neural networks:

• Feedforward Neural Network
• Convolutional Neural Network
• Recurrent Neural Network
• Autoencoders

Artificial Intelligence Interview Question

Give a brief introduction of partial, alternate, artificial, and compound keys?

Partial Keys: A set of attributes that uniquely identifies weak entities, which are related to the same owner entity.

Alternate Keys: All candidate keys except the primary key are known as alternate keys.

Compound Key: It has multiple fields that enable the user to uniquely recognize a specific record.

Artificial Key: It is the extra attribute added to the table when there are no stands alone or compounds key is available. It is created by assigning a number to each record in the table.

What is a Chatbot?

A chatbot is Artificial intelligence software or agent that can simulate a conversation with humans or users using Natural language processing. The conversation can be achieved through an application, website, or messaging apps. These chatbots are also called as the digital assistants and can interact with humans in the form of text or through voice.

The AI chatbots are broadly used in most businesses to provide 24*7 virtual customer support to their customers, such as HDFC Eva chatbot, Vainubot, etc.

What is knowledge representation in AI?

Knowledge representation is the part of AI, which is concerned with the thinking of AI agents. It is used to represent the knowledge about the real world to the AI agents so that they can understand and utilize this information for solving the complex problems in AI.

Following elements of Knowledge that are represented to the agent in the AI system:

• Objects
• Events
• Performance
• Meta-Knowledge
• Facts
• Knowledge-base

Advance Artificial Intelligence Interview Question

What are the various techniques of knowledge representation in AI?

Knowledge representation techniques are given below:

• Logical Representation
• Semantic Network Representation
• Frame Representation
• Production Rules

Which programming language is not generally used in AI, and why?

Perl Programming language is not commonly used language for AI, as it is the scripting language.

How does RL work?

The working of reinforcement learning can be understood by the below diagram:

The RL-based system mainly consists of the following components:

• Environment: The environment is the surrounding of the agent, where he needs to explore and act upon.
• Agent: The agent is the AI program that has sensors and actuators and the ability to perceive the environment.
• State: It is the situation that is returned by the environment to the agent.
• Reward: The feedback received to the agent after doing each action.

In RL, the agent interacts with the environment in order to explore it by doing some actions. On each action, the state of agent gets changed or sometimes remains the same, and based on the type of action, and he gets a reward. The reward is feedback, which may be negative or positive based on the action.

The goal of the agent is to maximize the positive reward and to achieve the goal of the problem.

Artificial Intelligence Interview Question

What are the different areas where AI has a great impact?

Following are some areas where AI has a great impact:

• Autonomous Transportation
• Education-system powered by AI.
• Healthcare
• Predictive Policing
• Space Exploration
• Entertainment, etc.

What are the different software platforms for AI development?

1. Google Cloud AI platform
2. Microsoft Azure AI platform
3. IBM Watson
4. TensorFlow
5. Infosys Nia
6. Rainbird
7. Dialogflow

Kindly explain different ways to evaluate the performance of the ML model.

Some popular ways to evaluate the performance of the ML model are:

• Confusion Matrix: It is N*N table with different sets of value that is used to determine the performance of the classification model in machine learning.
• F1 score: It is the harmonic mean of precision and recall, which is used as one of the best metrics to evaluate the ML model.
• Gain and lift charts: Gain & Lift charts are used to determine the rank ordering of the probabilities.
• AUC-ROC curve: The AUC-ROC is another performance metric. The ROC is the plot between the sensitivity.
• Gini Coefficient: It is used in the classification problems, also known as the Gini Index. It determines the inequality between the values of variables. The high value of the Gini represents a good model.
• Root mean squared error: It is one of the most popular metrics used for the evaluation of the regression model. It works by assuming that errors are unbiased and have a normal distribution.
• Cross-Validation: It is another popular technique for evaluating the performance of the machine learning model. In this, the models are trained on subsets of the input data and evaluated on the complementary subset of the data.

Advance Artificial Intelligence Interview Question

Explain rational agents and rationality?

A rational agent is an agent that has clear preferences, model uncertainty, and that performs the right actions always. A rational agent is able to take the best possible action in any situation.

Rationality is a status of being reasonable and sensible with a good sense of judgment.

What is tensor flow, and how it is used in AI?

Tensor flow is the open-source library platform developed by the Google Brain team. It is a math library used for several machine learning applications. With the help of tensor flow, we can easily train and deploy the machine learning models in the cloud.

Which algorithm is used by Facebook for face recognition? Explain its working.

Facebook uses the Deep Face tool that uses the deep learning algorithms for the face verification that allows the photo tag suggestions to you when you upload a photo on Facebook. The deep face identifies the faces in the digital images using neural network models. The working of Deep Face is given in below steps:

• It first scans the uploaded images. It makes the 3-D model of the image, and then rotate that image into different angles.
• After that, it starts matching. To match that image, it uses a neural network model to determine the high-level similarities between other photos of a person. It checks for the different features such as the distance between the eyes, the shape of the nose, eyes color, etc.
• Then it does the recursive checking for 68 landmark testing, as each human face consists of 68 specific facial points.
• After mapping, it encodes the image and searches for the information of that person.

Artificial Intelligence Interview Question

What is a market-basket analysis?

The market-basket analysis is a popular technique to find the associations between the items. It is frequently used by big retailers in order to get maximum profit. In this approach, we need to find combinations of items that are frequently bought together.

For example, if a person buys bread, there are most of the chances that he will buy butter also. Hence, understanding such correlations can help retailers to grow their business by providing relevant offers to their customers.

How can AI be used in fraud detection?

The artificial intelligence can be broadly helpful in fraud detection using different machine learning algorithms, such as supervised and unsupervised learning algorithms. The rule-based algorithms of Machine learning helps to analyze the patterns for any transaction and block the fraudulent transactions.

Below are the steps used in fraud detection using machine learning:

• Data extraction: The first step is data extraction. Data is gathered through a survey or with the help of web scraping tools. The data collection depends on the type of model, and we want to create. It generally includes the transaction details, personal details, shopping, etc.
• Data Cleaning: The irrelevant or redundant data is removed in this step. The inconsistency present in the data may lead to wrong predictions.
• Data exploration & analysis: This is one of the most crucial steps in which we need to find out the relation between different predictor variables.
• Building Models: Now, the final step is to build the model using different machine learning algorithms depending on the business requirement. Such as Regression or classification.

Give the steps for A* algorithm?

A* algorithm is the popular form of the Best first search. It tries to find the shortest path using the heuristic function with the cost function to reach the end node. The steps for A* algorithms are given below:

Step 1: Put the first node in the OPEN list.

Step 2: Check if the OPEN list is empty or not; if the list is empty, then return failure and stops.

Step 3: Select the node from the OPEN list which has the smallest value of evaluation function (g+h), if node n is goal node then return success and stop, otherwise

Step 4: Expand node n and generate all of its successors, and put n into the closed list. For each successor n’, check whether n’ is already in the OPEN or CLOSED list; if not, then compute evaluation function for n’ and place into Open list.

Step 5: Else if node n’ is already in OPEN and CLOSED list, then it should be attached to the back pointer, which reflects the lowest g(n’) value.

Step 6: Return to Step 2.

Advance Artificial Intelligence Interview Question

What is the inference engine, and why it is used in AI?

In artificial intelligence, the inference engine is the part of an intelligent system that derives new information from the knowledge base by applying some logical rules.

It mainly works in two modes:

• Backward Chaining: It begins with the goal and proceeds backward to deduce the facts that support the goal.
• Forward Chaining: It starts with known facts, and asserts new facts.

What do you understand by the fuzzy logic?

Fuzzy logic is a method of reasoning applied to the AI, which resembles human reasoning. Here the word “fuzzy” defines things that are not clear, it means the situations where it is difficult to decide if the state is True or False. It involves all the possibilities that occur between Yes and NO.

The below diagram shows the difference between fuzzy logic and Boolean logic

Since it resembles human reasoning, hence it can be used in neural networks.

What is a Bayesian network, and why is it important in AI?

Bayesian networks are the graphical models that are used to show the probabilistic relationship between a set of variables. It is a directed cycle graph that contains multiple edges, and each edge represents a conditional dependency.

Bayesian networks are probabilistic, because these networks are built from a probability distribution, and also use probability theory for prediction and anomaly detection. It is important in AI as it is based on Bayes theorem and can be used to answer the probabilistic questions.

Artificial Intelligence Interview Question

What is a heuristic function, and where is it used?

The heuristic function is used in Informed Search, and it finds the most promising path. It takes the current state of the agent as its input and produces the estimation of how close the agent is from the goal. The heuristic method, however, might not always give the best solution, but it guaranteed to find a good solution in a reasonable time. Heuristic function estimates how close a state is to the goal. It is represented by h(n), and it calculates the cost of an optimal path between the pair of states. The value of the heuristic function is always positive.

Admissibility of the heuristic function is given as:h(n) <= h*(n)

Here h(n) is heuristic cost, and h*(n) is the estimated cost. Hence heuristic cost should be less than or equal to the estimated cost.

Explain the different domains of Artificial Intelligence.

Machine Learning: It’s the science of getting computers to act by feeding them data so that they can learn a few tricks on their own, without being explicitly programmed to do so.
Neural Networks: They are a set of algorithms and techniques, modeled in accordance with the human brain. Neural Networks are designed to solve complex and advanced machine learning problems.
Robotics: Robotics is a subset of AI, which includes different branches and application of robots. These Robots are artificial agents acting in a real-world environment. An AI Robot works by manipulating the objects in it’s surrounding, by perceiving, moving and taking relevant actions.
Expert Systems: An expert system is a computer system that mimics the decision-making ability of a human. It is a computer program that uses artificial intelligence (AI) technologies to simulate the judgment and behavior of a human or an organization that has expert knowledge and experience in a particular field.
Fuzzy Logic Systems: Fuzzy logic is an approach to computing based on “degrees of truth” rather than the usual “true or false” (1 or 0) Boolean logic on which the modern computer is based. Fuzzy logic Systems can take imprecise, distorted, noisy input information.
Natural Language Processing: Natural Language Processing (NLP) refers to the Artificial Intelligence method that analyses natural human language to derive useful insights in order to solve problems.

What is an API? How do we deploy our own API to production Alize an ML Model?

Most Data Scientists know how to run python code on a Jupyter Notebook. We run the codes, do data analysis, come up with the final model result and stop there. How do machine learning systems in the real world interface with the rest of the systems in place? Even though python is great to experiment and perform machine learning, in terms of deployment, there are two main options:

• Rewrite the logic in a language that fits into the technology stack being leveraged.
  Imagine going through the entire machine learning workflow only to have to do it again. This is an option we would not recommend as the return on investment is not worth it.
• API-first approach
  Okay, so how do applications that use different frameworks or languages cross-communicate? APIs! APIs help cross-language applications function efficiently. So, what if you made the final machine learning model that makes sense, and you create a sort of endpoint that other developers or the front-end team can connect to. All they would need is the URL endpoint from where the API is being served.

Application Programming Interfaces help applications talk to each other. The person who needs to connect to the API needs to perform a simple REST call to the API using a Software Development Kit (SDK).

The two more popular approaches to productional zing our machine learning models are using Flask or Fast API. So, we serialize/desterilize the machine learning model objects into a pickle (.pkl) or an h5py file, letting users query it.

Think about it this way, let’s say you are deploying a model that predicts churn probability for a customer. So, if the query were a new customer, then the returned results would be the probability, which we could configure in the front end to say churn or no-churn based on the threshold.

Advance Artificial Intelligence Interview Question

What are the popular Cloud Services that support AI implementations? Explain!

AI Implementation is a fancy way of saying, are you able to handle an end-to-end implementation of AI / ML models on the cloud? To understand further, it is good to get familiar with the term MLaaS, which is Machine Learning as a Service. MLaaS offers multiple ML-related services as an additional component to cloud computing. There are multiple options in the market, so it would depend on the use-cases being tackled. Let’s discuss a few of the popular ones:

• Google’s Vertex AI
  
  Google recently unveiled their Unified MLOps & AI Platform – Vertex AI to help Data Scientists / ML Engineers increase experimentation, deploy faster and manage models better.
• AWS Machine Learning
  
  AWS is the most significant player in the cloud computing market. Their Machine Learning offerings have matured over time to cater to all machine learning needs from speech recognition, computer vision, AI, etc.
• Azure Machine Learning
  
  Azure has been grabbing more market share over the years because it has done a lot right in the machine learning space. Azure Machine Learning is a no-code, drag and drop interface to perform machine learning at Scale along with model management.

What are some scenarios where AI Cloud Service could be used to increase ROI?

Some scenarios where you can leverage AI cloud services include – 

• Customer Service Automation: Integrating Machine Learning and speech recognition into customer service can help save a lot of overhead costs
• Personalization and Engagement: Tailoring personalized customer experiences can lead to better sales and happier customers, which in turn leads to a higher revenue
• Fraud Detection: Multi-touch fraud detection algorithms can help isolate and pick out points where gaps can be plugged
• Scalable Media: Scalable architecture can be leveraged based on the number of concurrent users and workload 
• Forecasting: Multiple decisions can be optimized by leveraging accurate forecasting in terms of sales, supply-demand, costing, dynamic pricing etc. 

What is NLP? Explain the components of NLP?

Human language is spread with intricacies that make it complicated to decipher. Obtaining intended meaning from speech or text is even more difficult, especially when software is used. Natural Language Processing (NLP), a part of Information Retrieval, breaks down speech and text to help software understand what is happening.

The components of NLP are as follows:

• Speech Recognition
  Recognizing speech and being able to convert it reliably to text. There are various components of speech itself, along with multiple accents globally, which makes this a challenge.
• PoS Tagging
  Differentiation of the parts of a sentence to help tag the parts of speech such as noun, verb, adjective, etc.
• Word sense disambiguation
  Semantic analysis to determine the correct meaning of the word being used
• Naming Entity Recognition
  Identify proper nouns as entities as a part of the sentence, such as name, location, etc.
• Co-reference resolution
  Identify if pronouns refer to the same object. For instance, the question “How old is Obama?” And the follow-up question of how tall is he? should provide the correct answer if the co-reference resolution is configured correctly 
• Sentiment Analysis
  Understanding the tone of the speech or text based on subjective qualities or words used. This is used in social media to better gauge sentiments towards an event or a product.
• Natural Language Generation
  This gives power to the algorithm to make sentences that would make sense to humans. This is currently a hot area of research along with NLU (Natural Language Understanding)

Artificial Intelligence Interview Question

Now that large companies are leveraging AI to make decisions that daily affect human lives, it is essential to critique the implemented algorithms. On that note, what is AI governance?

There are still people who still believe AI is the future. The truth is the AI age is fully here and is becoming more prominent every year. A good point of view to discuss for this question is a recent article by Bloomberg Fired by Bot: Amazon Turns to Machine Managers, And Workers Are Losing Out – Bloomberg where algorithms terminate contract drivers when they are not even at fault.

AI / ML algorithms analyze the delivery drivers’ path, the time taken to deliver, customer satisfaction scores, and a host of other factors and rate the driver’s performance. Amazon is a company that operates on optimization, so they have replaced an entire HR Team with algorithms. This trend is worrying, especially if transparent policies to give visibility into end-to-end reasoning are not provided. AI Governance is a necessary evil. Without it, the very pillars of society can break down.

The four guideposts of AI Governance are as follows:

Integrity
The integrity and validity of the algorithm have to be justified by analyzing decision lineage and justification of what micro-decisions lead to the final output.

Explain ability
End-to-end transparency of the process can help stakeholders understand the decision and the reasoning behind it.

Fairness
Bias is an upcoming issue in most machine learning systems. Bias can be due to the data or algorithmic as well. It is on AI Governance to ensure that AI Systems are ethical, non-prejudiced, and protected against bias.

Resilience
Agile systems that take into consideration technical robustness and compliance can help protect the system against bad actors.

What are the components of AI Governance?

AI Governance can be measured as the following components:

• Data
  Tracks the data flow from start to end to ensure that the data lineage and provenance is validated to ensure there are no loopholes
• Security
  If someone in the AI System can manipulate the model’s results by tampering, this can lead to severe issues. You can tackle this in the future by using blockchain to imprint AI Systems.
• Cost and value of data
  Key performance indicators to track the cost of the data and the value obtained from the algorithm help measure effectiveness continuously.
• Bias
  Exposing selection and measurement bias with continuous automated tracking can help understand when a model drifts from its initial purpose (through self-learning). You should monitor this constantly to ensure that AI Ethics are maintained.
• Accountability
  Clarity on the individuals responsible for the system and accountable for its decisions is part of AI Governance of the future. All the way from security loopholes, maintenance, and monitoring
• Audit
  Audit trails and third party reviews can ensure that systems that affect human life are held accountable. 
• Time
  Model drift and impact over time should be captured to ensure that the model is more efficient than the traditional implementation.

Good AI exists. Looking at the top companies globally, we know that a robust framework exists where AI can scale. How do the best companies in the world build Scalable AI?

There are five principles using which we can Scale AI with AI Governance in place.

1. Algorithms as micro-services
Instead of making the entire AI workflow about the algorithm, consider algorithms as micro-services that can be subscribed to and managed independently. This leads to the following advantages that can be prescribed at scale.

• Reusability and lose coupling.
• Auto-scaling and scheduling
• Portability and virtualization

2. Six Sigma, factory-like approach to manufacturing and managing algorithm

Considering algorithms as part of the entire flow instead of the whole process means that we can focus more on manufacturing algorithms and reducing errors. Monitoring KPIs and version control can help to manage multiple algorithms at Scale. This will also help to keep track of performance along with who worked on the algorithm last.

• Standardized and automated workflows
• Performance monitoring 
• Audit Trace

3. Data Integration at Scale
Most data architectures rely on a single source of truth. In real-world enterprises, companies rely on multiple sources of truth rather than trying to make a single route alone. Having multiple data integration routes helps optimize the operational as well as analytical use of data.

• Experimentation in production
• Big Data
• Data Warehouse for core ETL tasks
• Direct data pipelines
• Tiered Data Lake

4. Move from Batch to event-based triggers
For more significant streams of real-time data, batch processing is preferred as there is greater control over the end-to-end architectural needs. However, going ahead, it is advisable to move from Batch to event-based triggers as this will help reduce infrastructural costs with a just-in-time application.

5. Leverage the cloud components to perform agile development
To scale AI, we have to move past legacy infrastructure and leverage cloud components such as MLaaS, PaaS, SaaS, and IaaS. This will give us access to the latest security, technology, reduced cost, and even world-class APIs.

Advance Artificial Intelligence Interview Question

What are the difficulties when it comes to scaling implementations of AI?

If leveraging and scaling AI was so easy, every company would have done it. It is difficult to Scale AI. Here are the top 5 difficulties that companies face when they attempt to Scale AI:

1. Technical Performance
When AI models are moved from development/testing to production, many new issues become unstable. When AI is scaled, technical problems are imminent.

2. Data Volumes and Veracity
Data volume and quality decide how fast the AI System is ready to scale. The larger the set of predictions and usage, the larger is the implications of Data in the workflow.

1. Complex Technology Implications at Scale
2. Onerous Data Cleansing & Preparation Tasks

3. Business Processes & People

People are the biggest surprise element in any AI Implementation. Companies want to be AI-first, but as they realise, AI is not just about training users but rather about amending processes, updating policies and putting in the right kind of business support.

1. Internal, company-level changes
2. Customer-facing changes

4. Unexpected Behavior
When dealing with a machine with multiple moving parts and complexities, it is challenging to pinpoint precisely where the error has occurred if the machine stops. Similarly, testing AI implementations is not just about unit testing with the appropriate data, but rather designing for all use-cases to deal better when there are unexpected errors.

5. Data Security and Governance
These vulnerabilities can make or break AI Systems at Scale. As businesses grow to rely on an AI-first approach, complete transparency and control over the system are critical. One breach in Data Security can break the reputation of the stakeholder. AI Governance is the most critical component in this entire piece.

State out the Enterprise AI Journey?

Being mindful of where an organization is in the AI journey can help make better decisions to ensure the company is headed in the expected direction. Following are some of the milestones that can help understand where the enterprise is at an organizational level:

• BI and Apps
  Every company starts with automation first. We need to understand and monitor the current state of data evolution at the enterprise level. This happens with the help of Business Intelligence Tools, analytics, and reporting.
  Tools: Power BI, Tableau, Qlik Sense 
• SaaS with AI
  Now, it is time to shoot for the low-hanging fruit. Multiple hypotheses and use-cases are put forth that are attempted to solve with the Data Science. Enterprises in this stage shoot for immediately actionable insights.
• AI Accelerators
  In this stage, companies are fully ready to leverage AI solutions. Solutions where speech, text, and other structures, as well as unstructured data, can be used to make better decisions
• Custom AI
  The final stage in the AI Journey is when a Custom AI solution to solve business problems can be made. It can be said that companies like Google are at this stage.

What are some of the common problems companies face when it comes to interpreting AI / ML? 

There are multiple use-cases where companies have an in-house or hire an analytics vendor to perform analytics on a dataset and get the output. Even though there is a machine learning model in place and the recommended solution, companies refuse to implement the proposed solutions. 

When AI / ML solutions are proposed, it is generally used to replace a heuristic (business-driven) model with a data-driven approach. However, the stakeholders view machine learning as a black box with an input and an output. Just showing them the output without a complete understanding of the model does not inspire as much confidence. Businesses tend to trust the model more if they understand the micro-decisions that lead to the final output.

For instance, SHAP and LIME are some of the popular approaches towards interpretable machine learning. Let’s take an example of a problem where we are trying to predict if a customer will churn (1) or not churn (0). 

Here is the implementation of LIME showcasing the result and the reasoning behind the output that the customer will not churn. Below is the implementation using SHAP.

The diagrams above will help explain the reasoning and the logic behind the decisions made in a model-agnostic approach.

Artificial Intelligence Interview Question

Explain the concept of XAI and how it can be used in real-world implementations?

Explainable AI is the implementation of AI / ML models in such a manner that human experts can understand them. Simpler models are interpretable. More complex models are not very interpretable. As models move from a simple regression model to deep learning to solve problems to improve results, there is a new observation that can be made.

Never models have low interpretability and high predictability. The latest area of research is XAI or Explainable AI on how we can maintain the strides we have made in predictability while making interpretability better.

Following are some of the popular approaches in machine learning interpretability : 

• SHAP takes a global view of machine learning models and uses Shapely values to leverage a global approach to interpreting results. 
• LIME stands for Locally Interpretable Model-Agnostic Explanations. LIME focuses on local interpretability and can be used for all approaches (model-agnostic).
• EBM or Explainable Boosting Machines uses the GA2M framework, which leverages a generalized additive model with an interaction model added to it (this is still a relatively new framework) 

AI/ML Models are black boxes that give us the desired output. What are some of the ways we can help the business feel comfortable with the models being implemented?

Some of the factors to bring explain ability back to AI/ML Models are as follows:

• Intention: Understanding the reasoning behind a decision can help both identify ways to correct and improve the reasoning, thus leading to better results
• Impact: If the impact is a human’s livelihood, the model is much more critical than a model that is trying to target the right demographic in a marketing campaign
• Control: Is the system allowed to make decisions autonomously, or does the system propose recommendations that are reviewed downstream by human operators before the final decision is made
• Rate: The number of decisions that the system makes and the audit and review mechanism
• Rigour: The system’s robustness will reduce unpredictable behaviour for scenarios it has not been trained on, like unseen data or abnormal values
• Law & Regulation: The legal and regulatory framework that the system operates within, along with those accountable for the system’s behaviour
• Reputation: The respect that people have for a system or organization based on previous performance
• Risk: The classic question of risk versus reward. What are the outcomes of a wrong decision? Death or merely a few cents lost on the dollar.

All of these factors together combine to form better systems for Interpretable AI.

Explain some examples of tools and technologies that can help implement interpretable machine learning or XAI.

• SHAP: It is a popular implementation that uses Shapely Additive explanations (SHAP) to explain the output of any model globally. It is a model agnostic framework.
• LIME: Locally Interpretable Model-agnostic Explanations is another popular model-agnostic framework that uses local interpretation to showcase the micro-decisions that the model makes before it comes up with the final output
• Shapash: This implementation takes the visualizations from SHAP and LIME explanations and forms it as an easy-to-use web application. It is essentially a forked version of SHAP and LIME to make it easier to showcase it to business users
• Explainer Dashboard: Think of the best of SHAP and LIME. Now, add many more machine learning interpretability / XAI elements and make a great web-based dashboard out of it for your use case. This is what the Explainer dashboard is
• DALEX: Dalex provides wrappers around ML frameworks. DALEX (model Agnostic Language for Exploration and explanation. All of DALEX’s plots are interactive as well
• EBM: Explainable Boosting Machines was created by Interpret ML that leverages glass-box modules
• ELI5: This is an explain ability package by MIT that provides local as well as global explanations

Advance Artificial Intelligence Interview Question

All of the conversations about XAI seem to be problems of the future. If you were to design an AI / ML model for a business now and production Alize a model, what are some of the considerations you would have as a leader of the project?

One of the common problems faced in the real world is that clients view the Data Science workflow as yielding results. Business teams try to show linear progress Sprint on Sprint to showcase that incremental work, however minimal, is being performed.

If Data Science were a linear activity, it would be called Data Engineering. Data Science Teams work in an extremely non-linear & complicated fashion but fail to educate the business about how Science is actually happening within the team. Rather than educating the business that Data Science is a set of experimental, non-linear attempts at obtaining something useful from the data, teams force-fit the results into the perception that the Data Science Team is moving towards the final goal.

As time passes and experiments are failing, the expectations of the business are on the rise. When they see the final result, there is a general disappointment as expectation generally does not match reality. In fact, I am so right about this that it surprises me.

87% of Machine Learning models do not go into production. –Venture Beat

 So, what are some of the ways that we can production Alize interpretable models for the business?

• Be real and transparent with the business: Instead of force-fitting results linearly, adopt a truly Agile approach to Data Science. Essentially, lay down the hypothesis you are dealing with, and Sprint by Sprint deal with each one and report back.
• Failures are good: Don’t be afraid to showcase failures to your stakeholders. This helps build trust and showcases that some experiments work and some don’t.
• Logging: Make a dashboard early on in the project so that the business can keep track along with your team of the successful and failed attempts. This might help bring new perspectives to the table later on. Do not just showcase the highlights of the week once a week in a PowerPoint to the meeting. Leverage the dashboard to showcase the overall progress – do not be afraid to iterate!
• Automate success: For the experiments that work, make production-ready code and keep the blocks ready so that when the time comes, the team is set to implement along with the basic test cases.
• Be critical of the data: One of the most common misses across some of the largest teams is that if we solve problem X in development, the ask becomes problem Y in production. Set up basic test cases for the data as well so that the system is robust!

What are some of the ways to push an AI/ML model to production? 

When a model is in the prototype and development phase, many teams use IDEs such as Jupyter Notebook. While this might seem like the best option during development as we can instantly see the results, it is advisable to write code worthy of being moved into production. The entire cycle can be visualized as shown below.

Now, let’s discuss how AI / ML models can be deployed into production. The ideology behind all of the steps is pipelines! Experiment to see what works best for your data, automate it using pipelines, and then monitor the performance of the workflow. 

• Data: Data Engineering Pipelines
  Data is everything. Make sure that the quality of data works for your use case.
  • Data Ingestion
  • Exploratory Data Analysis (using RAD Tools)
  • Validation
  • Data Wrangling
  • Data Splitting
• Model: Machine Learning Pipelines
  Every team enjoys experimentation with data. Therefore, it is essential to also be transparent with the business during the documentation phase and give equal focus to logging all experiments & learnings.
  • Model Training
  • Model Evaluation
  • Model Testing
  • Model Packaging
• Code: Deployment Pipelines
  Once the final ML model has been decided, it is critical to push it into production. Deploying pipelines would include the following phases:
  • Model Serving
  • Model Performance Monitoring
  • Model Performance Logging

Based on demand, the model serving requirements can be further analyzed.

How can Ethical AI be implemented at an enterprise level? What are its implications for Enterprise AI use-cases and Governance?

Ethical AI is the practice of leveraging AI with good intentions to empower employees and businesses. Ethical AI lets companies Scale AI with confidence. Companies are currently in the development phase of what AI is going t be in the future. There is no single source of truth. Each company has their own Ethical AI Framework:

Google

Microsoft

Facebook

So, there is currently no one-stop solution as to what Ethical AI should look like at an enterprise level as it is a work in progress. However, some of the pillars of Ethical AI are as follows:

Privacy & Security

Fairness & Inclusion

Robustness & Safety

Transparency & Control

Accountability & Governance

Artificial Intelligence Interview Question

With privacy focus coming to the forefront, a lot of attention has been brought to data collected. A new branch of data collection and processing for ai / ml is federated learning. Explain further.

Federated machine learning is Google’s proposed way to approach advertisements to replace the current cookie-based tracking approach. The Secure Aggregation Principle is a step-by-step procedure as to how privacy-preserving principle can be implemented:

1. ML Model on-device: With chipsets getting more advanced and having dedicated hardware for AI / ML, a local version of the model will be deployed on the device 

2. ML model becomes smarter: The model on the device will get trained and become more innovative based on the user’s usage

3. Transfer of model results (only): The results of the model are then transferred from the local device onto the centralized server

4. Results Aggregated on centralized server: The results from several devices are then aggregated on the centralized server. This will not contain any user data, which is the privacy-preserving information

5. Update machine learning model: The centralized machine learning model is now implemented based on the new data. This will help give the user better experiences.

6. The better model implemented with Federated Learning: Once the model is updated, it will be deployed to have a better and wiser model.

This leads to benefits such as hyper-personalized, low cloud infra overheads, minimum latencies, privacy-preserving machine learning models!

Explain how a model can be monitored after it is moved to production. Break down the approach.

Functional monitoring for models in production can happen under the following brackets:

• Data (Input)
  • Data Quality Issues
  • Data / Feature Drift
  • Outliers
• Model
  • Model Drift
  • Model Configuration
  • Model Version
  • Concerted Adversaries (Hackers or unwanted agents)
• Predictions (output)
  • Model Evaluation metrics
  • Prediction Drift

Some of the more general production challenges are as follows:

• Changes in Data Distribution
• Ownership of the model in production
• Training-Serving skews
• Model or Concept Drift
• Black Box Models
• Issues with pipelines
• Outliers in the Data
• Issues with Data Quality

Mention some use-cases as to how AI / ML can be used within Cyber Security use-cases.

Instead of looking at machine learning and figuring out how to apply it to Cyber Security use cases, let’s look at Cyber Security applications and see how Machine Learning can help.

• Network Protection
  This refers to Intrusion Detection Systems (IDS), where machine learning can help using Network Traffic Analysis (NTA).
• Endpoint Protection
  When we download an executable file and run it, the risk of malware is much higher than normal. Using machine learning to isolate and classify such risks can help secure systems.
• Application Security
  Machine Learning can help with Enterprise security or Web Application Firewalls (WAF) applications.
• User Behavior
  User Behavior began as Security Information and Event Management (SIEM). Gartner has a specialised framework to deal with User and Event Behavior Analytics (UEBA).

• Process Behavior
  Security risks for use-case driven business risks. This is a customs problem and will depend on the domain, the horizontals and verticals of the business.

Advance Artificial Intelligence Interview Question

Explain the flow of Cyber Security attacks and how AI / ML models can help plug the gaps.

Cyber Security is critical at each part of the pipeline. There is no single part where Machine Learning can aid in keeping systems safe. At every step, there is scope for automation and machine learning to improve the quality of cybersecurity at a reduced cost. There are a few main reasons why Machine Learning Systems beat out more traditional approaches:

• Machine Learning can recognize abnormal patterns flag them easily. This can be set to as sensitive as required, depending on the use-case
• Machine Learning systems keep improving as they keep learning
• Since Machine Learning systems are automated, they require minimal intervention and are cost-efficient. 

How can Cyber Security be implemented across a PoC Team or at an enterprise level with the help of AI / ML?

Cybersecurity is a critical part of both small and large organizations. Following Information security practices that enterprises have set up is a good place to start. The three components that can be catered to in cybersecurity at a broad level are people, process, and technology.

• People
  Companies do not realize the value of cuber security professionals until it is too late. Similar to a lot of other scenarios, prevention is better than cure.
  • Lack of skilled professionals
  • Data across defined boundaries
  • Social Engineering attacks
• Process
  Business processes vary immensely even within the same enterprise. Catering to cybersecurity compliance along with balancing cost pressure can lead to better systems.
  • Cost pressure
  • Regulatory compliance
• Technology
  This is the most expensive and critical element in the cybersecurity workflow. We can scale enterprise-wide and secure most applications with fundamental analysis on best practices org-wide.
  • Boundary-less enterprise
  • One-size fits all security technology
  • Speed of technology adoption

What is the difference between AI and RPA?

Technologies such as RPA can be considered as a subset of Artificial Intelligence. AI and RPA work hand in hand to automate complicated tasks that require more thinking along with doing.

In case you are still looking for it, there is no clear answer. It is like asking for the difference between your brain and your hand when performing a task. The muscle memory from your hand works for many repetitive use-cases, and when there is a change or anything new, your brain needs to come into play and decide on new processes.

Similarly, by combining RPA and AI, enterprises increase capabilities and make their processes more efficient. 

Focusing on RPA helps you initially clearly define tasks, sub-tasks and focus on automation. Once the stakeholders are able to see the effectiveness of RPA at Scale, the eventual logical step is to pull in AI to perform more complicated tasks. AI and RPA implementation start out with a Proof of Concept (POC) that can later be scaled, depending on the use case. Some of the cases that can be deemed AI-worthy are:

• Workflows that have uncertain outcomes that require interpretation
• Highly variable processes that do not depend on a rule-based system entirely might evolve over time
• Processes that rely on unstructured data or AI Enhanced APIs

Artificial Intelligence Interview Question

Explain the landscape of how RPA and AI can be used in tandem to increase revenue at an enterprise level? Discuss a few use cases.

Depending on where the enterprise is in its AI journey, there are different ways AI and RPA can be used to increase the bottom line. Just to recall, the following are a few examples of RPA:

• Moving files/folders
• Read/write to databases
• Scrape data
• Log into apps
• Extract structured data
• Fill forms
• Email events

Following are the use-cases with AI and RPA working in tandem:

• Understanding Documents: Rather than performing similar tasks, entity extraction and intent classification may also be performed
• Classification of Emails: Emails can be put into different buckets for further processing downstream
• Semi-structured data: Data such as images, text, speech, and video data have nuances that will need to be analyzed and understood on a case to case basis
• Speech to text: Speech data has multiple components that make it especially difficult to interpret
• Chatbots: Chatbots can help automate multiple elements of a workflow with the help of RPA and AI

Why is RPA preferred?

For people who are more familiar with automating processes, there might be less respect for RPA teams as it seems like there is not much getting done. Don’t let Data Scientists ruin RPA efforts. RPA can save millions of dollars annually by automating simple tasks. To being with, that may be where the real savings for organizations may be.

• Accuracy: RPA bots do not make mistakes because they do not think; they just do
• Low Technical Barrier: There are no programming skills required to configure RPA bots
• Compliance: This is one of the main advantages of RPA. Compliance across various geographies can be adhered to perfectly
• Non-Invasive Technology: Systems do not need to be upgraded; no major initial investment as such is needed
• Improved Employee Morale: Everyone likes being valued. Automating rudimentary and repetitive tasks that add no value leaves the employee to find more engaging & interesting work
• Productivity: Process cycles are much faster as compared to the manual effort as bots don’t take breaks or make mistakes
• Reliability: Processes can run 24/7 with no interruptions
• Consistency: Routine tasks are performed the exact same each and every time 

Do you think machines will surpass humans? What will happen if this becomes true in the future?

Yes, machines can surpass humans if they can think as humans, which looks like a reality considering the advancements in technology. If this happens, a lot of tasks can be performed without errors that humans make otherwise. On the other hand, if machines start making all the decisions, there could be chaos in the existing systems. However, since humans are the ones controlling AI, machines can only work in tandem and never surpass humans.

Advance Artificial Intelligence Interview Question

What are the AI technologies that you use every day?

There are many AI technologies that we use knowingly or unknowingly, for example, we use Alexa/Siri which uses Natural Language Processing. Robots are used in many places to perform repetitive tasks. Drones are used to capture images from any height and direction. Facebook performs face and image recognition.

Recently, there was news that UK researchers developed an AI tool (still in the research phase) to detect various types of brain injuries based on 600 different CT scans.

Do you believe that self-driving cars will work like human driving someday?

With the current advancements in AI, it is only a matter of time before this happens – this could be one of the answers as true, there have been tremendous improvements in computer vision and image classification technologies. However, you could also argue that it is a far-fetched dream and can’t be as dynamic or accurate as human driving.

If you were to build an AI system, what would you like to build?

Your answer can be based on what is the one job that you probably would want to automate. For example, a system that would enable people to maintain social distancing norms wherever they are and alert the police if someone is violating the same. Drones are already doing that, but someone needs to monitor it. This could be a system such that users can appreciate the benefits of maintaining social distancing.

Or you could ask for a home management system that can alert if there is a shortage of groceries, vegetables, and other essential items at home, set up appointments, read out daily news, etc.…

A robot that can clean the house, chop vegetables, wash and fold clothes, etc.… does seem like a reasonable system too!

Artificial Intelligence Interview Question

Explain the layers of a neural network.

There are three layers in a neural network –

• Input: This layer consists of input values. Input can be loaded from different data sources like file, databases, web services, etc…
• Hidden: There can be many hidden layers that transform the input data received. For most problems, a single hidden layer is sufficient. The input data is propagated through the hidden layers in the forward direction using the activation function.
• Output: This layer gives the final output. It computes the output through the information received from the last hidden layer. The output layer has only one node.

Explain the different branches (domains) of AI.

There are 6 branches or capabilities of AI –

1. Machine learning – ML helps machines to understand large input datasets, apply models to the same using various algorithms, and get insights and predictions about the data.
2. Neural network – NN tries to mimic a human brain and perform tasks using cognitive science and machines. Neural networks have three layers – input, hidden, and output. The neural network uses the concept of neurology.
3. Robotics – robotics deals with the designing, creation, operation, and usage of robots. Robots can perform tasks that are otherwise difficult, dangerous, or monotonous for humans.
4. Expert systems – expert systems can mimic the intelligence of a human brain to make decisions. Expert systems are purely based on reasoning and logic created using if-then-else statements and rules.
5. Fuzzy logic – Fuzzy logic is used to modify uncertain information by measuring the degree to which a hypothesis is valid. The degree of truth can lie between 0 and 1, and there are more states than just 0 and 1.
6. Natural language processing (NLP) – NLP enables humans to communicate with the machine using human language (natural language). Through NLP, the computer can also identify how a user would behave in certain scenarios, for example, analyze the tone of the text, recognize speech, analyze sentiments, etc…

Is there a way in which a program can improve itself? How?

Yes, programs can evolve through machine learning. Such algorithms are called genetic algorithms, which can improve by selecting the best possible solution, with a set of constraints. The program keeps repeating until it reaches a certain limit. Genetic algorithms are said to be the programmatic implementation of the survival of the fittest.

Advance Artificial Intelligence Interview Question

Which programming language is best suited for robotics and why?

There are many languages but C++ and Python are the most preferred languages. Both are used in the Robot Operating system. C++ has many features like embedded programming, navigation, image processing, motor control, etc.… that cater to robotics. Python comes with loads of libraries for machine learning and deep learning. Other than that, MATLAB is a good language to learn for robotics because it supports embedded programming, rapid prototyping, and image processing tools.

What do you think are the areas where AI has a great impact?

AI has a great influence on numerous areas. At present it is –

• Computing field
• Speech Recognition
• Bioinformatics
• Humanoid robots
• Space and Aeronautics
• Weather forecasting

Explain Tree Topology?

As the name suggests “Tree” topology has several connected elements arranges like the branches of a tree. The structure has at least three specific levels in the hierarchy. These are scalable and accessible while troubleshooting and are so preferred. A common drawback in this topology is the hindrance or malfunctioning of the primary node.

Artificial Intelligence Interview Question

Narrate some of the branches of AI?

There are some branches of AI are as follow:

Automatic Programming
Constraint Satisfaction
Bayesian Networks
Knowledge representations
Machine Learning
Natural Language Processing (NLP)
Neural Networks
Robotics
Speech recognition

How to select the best hyperparameters in a tree-based model?

There are two best Hyperparameter in a tree-based model

• Measure the performance over training data
• Measure the performance over validation data

We have to consider the validation result while comparing with the test results, so the answer is B

Explain the objective and related terminology used in the search algorithms of AI?

This is the most popular Artificial Intelligence Interview Questions asked in an interview. Searching is the universal techniques used in AI problem techniques. This algorithm is used to search a particular position. Every search terminology has some components.

• Problem space: this is the environment in which the search takes place.
• Problem Instance: it’s a result of the Initial State + Goal state.
• Problem Space Graph: This is used to represent a problem state.
• The depth of a problem: Here we can define the length of the shortest path.
• Space Complexity: We can calculate this by the maximum number of nodes that are stored in memory.
• Time Complexity: It is defined as the maximum number of nodes that are created.
• Admissibility: This is the property of the algorithms that are used to find the optimal solutions.
• Branching Factors: This can be calculated by the average number of child nodes in the problem space graph.
• Depth: it is the length of the shortest path from inception to the goal state.

Here are some of the search algorithms

Breadth-first search
Depth-first search
Bidirectional search
Uniform cost search

Advance Artificial Intelligence Interview Question

List down some of the best AI software platforms?

Following are the best AI software platforms:

• Tensor Flow
• Azure Machine Learning
• Payment
• Salesforce Einstein
• Cloud Machine Learning

List three techniques for selecting features in machine learning.

• The Wrapper approach
• The Filter approach
• The Embedded approach

Describe the frames and scripts?

Frames are a semantic network, which is a usual form of expressing expert system non-procedural knowledge. Frame, a structure of artificial data, divides information by reflecting stereotyped situations into the substructure. Scripts are like frames, except that the values must be ordered in the slots. In natural language understanding, script(s) are used to organize a knowledge base to understand the situation.

Artificial Intelligence Interview Question

What is Hidden Markov Model (HMMs) is used?

Hidden Markov Models are a ubiquitous tool for modeling time series data or to model sequence behavior.  They are used in almost all current speech recognition systems.

In Hidden Markov Model, how does the state of the process is described?

The state of the process in HMM’s model is described by a ‘Single Discrete Random Variable’.

In HMM’s, what are the possible values of the variable?

‘Possible States of the World’ is the possible values of the variable in HMM’s.

Advance Artificial Intelligence Interview Question

	Artificial Intelligence Part 2	Artificial Intelligence Part 3