Machine Learning Engineer Concept Interview Questions

Kavy U. - "Basically, normalization layer is a technique use to accelerate and stabilize the training of neural network."See full answer

Alex N. - "Perplexity - the measure LLM is surprised when it predicts next word. For example: I love to eat --- if LLM selects as next word "fruits" it will be less surprising than if LLM selects as next word "metal". It is better to have lower perplexity score. Cross Entropy is the measure how well model match true labels. So if the next word is "cat" and LLM assigns 0.5 to it then cross entropy value is -log(0.5) = 0.69 and if it assigns 0.9 probability to word cat then cross entropy value is - log(0.9"See full answer

Nathan B. - "There are many good answers to this that AI scientists around the world, I and my coworkers have tried over the years. For one, RAG is a great option to fact-check and enforce citation generation, update the data in the knowledge base of the generative AI, etc. "See full answer

Alex N. - "The adjusting context window size in LLM change trade off between reasoning capability, accuracy, computation cost. It influence how attention mechanist allocate resources across the input. Longer context window let it you input greater number of words and have more context to generate proper next token. However llms have lost in the middle issue. They remember the beginning of text and end of text but lost information located in the middle of long input. Another problem is Attention Dilution."See full answer

Mark G. - "I look at AI safety, ethics, guardrails, and governance as layers of one core responsibility: making systems useful without letting them cause harm. The Technical Core: Safety Safety is about managing technical risks like reliability, bias, and unintended behavior. It cannot be a final checklist before launch. It has to be part of the whole cycle, from how data is collected to how the system is monitored after it goes live. The Human Element: Ethics Ethics asks if a system should"See full answer

Anchal V. - "supervised learning: model is trained on the labeled data unsupervised learning: no labels provided - model learns by finding patterns , structure and groupings in the data. Semi-supervised learning: use small set of labels to guide learning for the larger pool of unlabeled data. reinforcement learning: leans by interacting with students the environment, receives reward and penalties based on actions self supervised: no labelled data . The model makes its own practice problems by"See full answer

Hardik saurabh G. - "Hallucinations are evaluated by measuring how often generated outputs contain information that is not supported by trusted sources. what hallucination means in context: Intrinsic hallucination: contradicts provided context Extrinsic hallucination: introduces unsupported facts Fabrication: confidently incorrect answers"See full answer

Dessalew A. - "Gradient Descent = core engine for training most ML models It works by iteratively minimizing loss via gradients Many improvements exist (Adam, RMSProp, etc.) Alternatives exist for: Faster convergence Non-differentiable problems Direct metric optimization"See full answer

Louie Z. - "DNNs can learn hierarchical features, with each layer learning progressively more abstract features, and generalizes better. SNNs are better for simplier problems involving smaller datasets and if low latency is required."See full answer

Anonymous Aardvark - "in simple words, linear regression helps in predicting the value whereas logistics regression helps in predicting the binary classification. But lets talk through some example Linear regression model: E-commerce website pricing recommendation engine is built on linear regression model where we do have some variables such as competitor price, internal economics and consumer demand etc when we put this in a supervised learning model, it helps in predicting prices Logistics regression model"See full answer

Will I. - "Is it bad to get the answer a different way? Will they mark that as not knowing Bayes Theorem or just correct as it is an easier way to get the answer? The way I went is to look at what happens when the factory makes 100 light bulbs. Machine A makes 60 of which 3 are faulty, Machine B makes 40 of which 1.2 are faulty. Therefore the pool of faulty lightbulbs is 3/4.2 = 5/7 from machine A and 1.2/4.2 = 3/7 from Machine B."See full answer

Michael F. - "As an interviewer, I have asked this question to candidates in the past. Here are the major topics I am looking for in an interview The candidate should understand that there are ways of measuring the loss of a particular clustering. For example, we can take the average distance of each point to it's cluster center. The candidate should understand that this loss will always decrease as the number of clusters increases. For that reason, we can't just pick the value of K that minimizes the l"See full answer

Jyoti V. - "Over-fitting of a model occurs when model fails to generalize to any new data and has high variance withing training data whereas in under fitting model isn't able to uncover the underlying pattern in the training data and high bias. Tree based model like decision tree and random forest are likely to overfit whereas linear models like linear regression and logistic regression tends to under fit. There are many reasons why a Random forest can overfits easily 1. Model has grown to its full depth a"See full answer