Top 30 Advanced AI & ML Concepts Quiz
30 advanced multiple-choice AI & ML concept questions, the corners that separate having used it from understanding it: internals, edge cases, and the reasons behind the design. They come from 30 bites in the AI & ML library, the hardest slice of the 606 AI & ML concept questions in the library. Answer them here or read straight down. Every question carries the correct option, why it is correct, and a link to the bite it came from.
Artificial intelligence, machine learning, and data science
30 questions. Pick an answer, or open “Show the answer” to read it.
Answers are graded in your browser. Nothing is saved, and no XP or streak is earned here. The app keeps score.
Question 1 of 30
What is the most significant consequence of the vanishing gradient problem in deep neural networks?
Show the answer
Answer: b · Early layers of the network fail to learn effective features from the input data.
The card states that vanishing gradients cause 'the gradients for the earliest layers become so small they effectively vanish, and those layers stop learning,' meaning they cannot learn effective features. Option D is incorrect because vanishing gradients lead to a failure to learn, not typically overfitting. Options C and D are general problems but not the specific, direct consequence on the learning of early layers.
Question 2 of 30
For which scenario would an LSTM be preferred over a traditional RNN?
Show the answer
Answer: c · Tasks requiring memory of context from distant points in a long sequence.
LSTMs are specifically designed to overcome the vanishing gradient problem in traditional RNNs, enabling them to maintain and utilize information from far back in a sequence. For short-term dependencies, a simpler RNN might be more efficient, and for large-scale parallel tasks, Transformers are often preferred.
Read the full bite: LSTMs: Giving Neural Networks a Longer Memory
Question 3 of 30
What is the primary limitation of a basic Seq2Seq model when dealing with very long input sequences?
Show the answer
Answer: a · The fixed-size context vector acts as an information bottleneck, leading to the forgetting of early input details.
The card states that the "primary weakness of basic Seq2Seq is its reliance on a single, fixed-size context vector" which "becomes an information bottleneck" for "very long inputs," causing the model to "forget details from the beginning of the input." Option D describes the problem that Seq2Seq was designed to solve, not its limitation.
Read the full bite: Seq2Seq: Turning One Sequence Into Another
Question 4 of 30
Which core concept allows causal inference to distinguish cause-and-effect from mere association?
Show the answer
Answer: c · The estimation of a counterfactual outcome
The card explicitly states that causal inference's mental model is to create a 'counterfactual'—what would have happened if the cause had not been introduced—to determine the causal effect. While identifying confounders and using RCTs are crucial methods in causal inference, the counterfactual is the underlying conceptual tool for distinguishing causation from association. Analyzing correlations is what causal inference aims to move beyond.
Read the full bite: Causal Inference: Proving Cause, Not Just Correlation
Question 5 of 30
Which scenario most strongly indicates the necessity of implementing a Continuous Training (CT) pipeline for an ML model?
Show the answer
Answer: c · The model's performance metrics are steadily degrading in production due to shifts in the input data distribution.
Continuous Training (CT) is specifically designed to address model performance decay caused by 'data drift' or non-stationary data patterns in production. Option C directly describes this core problem. Option D refers to traditional CI/CD for code or infrastructure, not CT for model retraining due to data changes.
Read the full bite: Continuous Training: CI/CD for ML Models
Question 6 of 30
What is the primary advantage of using uplift modeling over a traditional conversion prediction model?
Show the answer
Answer: b · It enables targeting of individuals whose conversion probability is significantly increased by a specific intervention.
Uplift modeling's core purpose is to identify 'Persuadables'—those whose behavior will change due to an intervention, maximizing the incremental impact of costly resources. Option C is incorrect because uplift's advantage is not just general prediction accuracy, but specifically predicting causal impact.
Read the full bite: Uplift Modeling: Who to Target, Not Just Who Will Convert
Question 7 of 30
A spam filter trained on 2020 email data struggles in 2023 because spammers now use keywords previously common in legitimate emails. What type of drift is this?
Show the answer
Answer: b · Concept Drift, because the relationship between keywords and whether an email is spam has fundamentally altered.
This is Concept Drift because the fundamental relationship between the input features (keywords) and the target variable (spam) has changed; what once indicated legitimate email now indicates spam. While the distribution of keywords has shifted (Data Drift), the core issue is the altered meaning of those keywords in predicting spam, which is characteristic of Concept Drift.
Read the full bite: Data Drift vs. Concept Drift: When Models Go Stale
Question 8 of 30
Which statement best explains why radiometry is crucial for a self-driving car's camera system?
Show the answer
Answer: d · It quantifies the physical energy of light hitting the sensor, enabling objective machine interpretation.
Radiometry measures the actual physical energy of light (in watts), which is critical for machines like self-driving cars to objectively interpret their environment, independent of human perception. Option B is incorrect because radiometry specifically avoids human perception weighting, which is the domain of photometry.
Read the full bite: Radiometry: Measuring Light as Physics, Not Perception
Question 9 of 30
When a production model performs poorly, how does ML Metadata primarily facilitate debugging?
Show the answer
Answer: b · It enables tracing the model's complete lineage, including the exact data, code, and hyperparameters used for its training.
Option B accurately describes ML Metadata's core debugging function: tracing a model's lineage back to its training inputs and processes. Option D is tempting due to the 'git blame' analogy, but ML Metadata tracks artifacts and executions, allowing identification of the code version used, not automatic flagging of specific code changes.
Read the full bite: ML Metadata: The Logging Layer for ML Pipelines
Question 10 of 30
Which optical phenomenon is NOT directly accounted for by a standard Bidirectional Reflectance Distribution Function (BRDF)?
Show the answer
Answer: a · The way light penetrates and exits a marble sculpture.
The card explicitly states that standard BRDF models are for opaque surfaces only and do not account for light transmitted through a material (translucency) or scattered underneath the surface, which describes light interacting with a marble sculpture. The other options are all forms of surface reflection that BRDFs are designed to model.
Read the full bite: BRDF: Modeling How Surfaces Reflect Light
Question 11 of 30
Which scenario best highlights the necessity of cross-attention in a neural network?
Show the answer
Answer: d · Translating a sentence where the decoder needs to align specific output words with relevant input words.
Cross-attention is crucial in sequence-to-sequence tasks like translation, enabling the decoder to dynamically focus on relevant parts of the encoder's output. Option A describes the information bottleneck that cross-attention was designed to overcome, not its necessity.
Read the full bite: Cross-Attention: How Models Connect Two Ideas
Question 12 of 30
What is the fundamental consequence of a digital camera utilizing a Bayer filter for its image sensor?
Show the answer
Answer: c · Each photosite records only one color component, requiring software interpolation for full-color pixels.
The card states that a Bayer filter allows each pixel to see only one color, and the full-color image is then reconstructed in software through a process called demosaicing. Option D is incorrect because the card explicitly states a Bayer filter does not provide 24 million full-color pixels directly.
Read the full bite: Bayer Filter: Faking Full Color with a Filter Grid
Question 13 of 30
What fundamental limitation necessitates the tokenization step for feeding text into a Transformer model?
Show the answer
Answer: b · Transformer models are mathematical functions that exclusively process numerical tensors.
The card explicitly states that "Neural networks, including Transformers, are mathematical functions that operate on numbers, not raw text strings." Thus, text must be converted into a numerical tensor format. While tokenization helps manage vocabulary and prepares for embeddings, the core reason is the model's numerical input requirement.
Read the full bite: Transformer Preprocessing: From Text to Tensors
Question 14 of 30
Which scenario best highlights the core advantage of implementing a data lake?
Show the answer
Answer: c · A research team collecting vast amounts of raw, multi-format sensor data for exploratory machine learning.
The correct answer (C) aligns with the data lake's purpose: storing diverse, raw data for exploratory analysis and machine learning without upfront schema. Options A, B, and D describe use cases better suited for transactional databases, data warehouses, or indexed systems, which data lakes are explicitly not designed for.
Read the full bite: Data Lake: A Single Repository for Raw Data
Question 15 of 30
What is the primary advantage of the Canny edge detector over simpler methods like the Sobel filter?
Show the answer
Answer: b · It produces clean, thin, and continuous edge maps by effectively managing noise and connecting weak edges.
The card highlights that Canny's design provides a more robust method that finds continuous, well-defined edges, which is achieved through its multi-stage process of noise reduction, thinning (non-maximum suppression), and connecting weak edges (hysteresis). Option A is incorrect because the card states simpler filters might be preferred if raw speed is the only priority, implying Canny is not faster.
Question 16 of 30
An engineer fits a distribution to server latency using MLE and reports tight confidence intervals. If the chosen distribution family does not match the true data-generating process, what best describes the result?
Show the answer
Answer: b · The estimates maximize likelihood within the wrong family and can be precisely misleading
C is correct because MLE finds the parameters that make the observed data most probable within the assumed model, so a wrong family yields a precise but misleading fit. D is tempting but wrong because misspecification does not automatically inflate uncertainty; the method can be confidently wrong.
Read the full bite: MLE: Find the Parameters That Make Data Likely
Question 17 of 30
Which visual effect cannot be achieved using only an affine transformation?
Show the answer
Answer: d · Making an object appear to shrink as it moves further away
Affine transformations preserve parallelism, meaning parallel lines remain parallel, which prevents them from modeling perspective effects like objects appearing smaller with distance. This requires a more general projective transformation. Rotating an image around its center point, while involving both rotation and translation, is fully supported by affine transformations.
Read the full bite: Affine Transformations: Move, Scale, and Rotate Without Bending
Question 18 of 30
According to the card, what is the primary reason SVD can decompose any real or complex matrix while eigendecomposition cannot?
Show the answer
Answer: c · SVD represents any matrix as rotation, scaling, and rotation, whereas eigendecomposition requires a square normal matrix with an orthonormal eigenbasis.
The card explicitly states that eigendecomposition is restricted to square normal matrices with orthonormal eigenbases, while SVD factors any matrix as rotation, scaling, rotation. Option A describes an external computational fact not mentioned in the card, and option D confuses normal matrices with length-preserving transformations.
Question 19 of 30
According to the card, what is the most accurate mental model for understanding how a Large Language Model (LLM) processes information?
Show the answer
Answer: b · A sophisticated pattern-matching engine that predicts word sequences based on learned statistical relationships.
The card describes an LLM as a "sophisticated pattern-matching engine for words and ideas" that learns "statistical relationships between words" to predict text. Option C is incorrect because the card explicitly states an LLM is "not a database of facts".
Question 20 of 30
How does the Laplacian of Gaussian (LoG) primarily identify the center of a blob in an image?
Show the answer
Answer: a · By finding local extrema (peaks or valleys) after applying a Gaussian blur and then the Laplacian operator.
The card states that after applying the Gaussian blur and Laplacian, 'the local extrema (peaks and valleys) mark their centers.' Option B is incorrect because zero-crossings indicate the edges of blobs, not their centers.
Read the full bite: Laplacian of Gaussian (LoG) for Blob Detection
Question 21 of 30
Which statement accurately describes the primary role of the Jacobian matrix for a differentiable multi-variable function at a specific point?
Show the answer
Answer: d · It provides the best linear approximation of the function's local behavior.
The card states that "The Jacobian matrix is the best linear approximation of a function at a specific point," capturing its local, first-order behavior. Option B is incorrect because the card explicitly mentions that the Jacobian does not provide second-order derivative information like curvature or concavity.
Read the full bite: The Jacobian Matrix: A Derivative for Multiple Dimensions
Question 22 of 30
Under a fixed FLOP budget, a team must choose between a 100B-parameter model on 200B tokens and a 50B-parameter model on 400B tokens. What should they expect?
Show the answer
Answer: b · The 50B model will likely match or outperform the 100B model while costing less to serve
The card states that for a fixed FLOP budget, scaling parameters and tokens equally is optimal, so the smaller model on more data can match or beat the larger one while being cheaper to serve. Option D reflects the outdated assumption that parameters alone drive performance, while D incorrectly assumes overfitting rather than undertraining is the risk.
Read the full bite: LLM Scaling Laws: Match Parameters to Tokens
Question 23 of 30
What is the primary advantage of employing pipeline parallelism in deep learning model training?
Show the answer
Answer: d · It enables the training of models whose total parameter count exceeds the memory capacity of a single GPU.
B is correct as the card states pipeline parallelism is for models 'too large to fit into the memory of a single accelerator.' A is incorrect because, despite using parallelism, the card notes that communication overhead and pipeline bubbles can make it slower for models that fit on a single GPU, so its primary benefit isn't universal speedup.
Read the full bite: Pipeline Parallelism: An Assembly Line for Your Model
Question 24 of 30
What is the primary motivation for employing Tensor Parallelism in large language models?
Show the answer
Answer: c · To process a single, excessively large model layer by distributing its internal components across multiple GPUs.
Tensor Parallelism is specifically designed to enable the execution of individual model layers that are too large to fit into a single GPU's memory by splitting the layer's components. Option A describes pipeline parallelism, which distributes entire layers, not parts of a single layer.
Read the full bite: Tensor Parallelism: Split Layers, Not Just Models
Question 25 of 30
What is the primary problem a point-in-time correct join solves when preparing data for machine learning models?
Show the answer
Answer: b · Data leakage, by ensuring that only historically available feature values are used for each training example.
The card explicitly states that a point-in-time correct join is vital "to prevent data leakage" by ensuring you only retrieve feature values that were available at a specific point in time. Option D is incorrect because the card notes it is "not to use it" for real-time online inference.
Read the full bite: Point-in-Time Correctness: Avoiding Data Leakage in ML
Question 26 of 30
What is the primary challenge in ML workflows that Feature Governance, typically implemented via a feature store, aims to resolve?
Show the answer
Answer: b · The risk of inconsistent feature logic leading to performance degradation between training and serving.
The card explicitly states that Feature Governance exists to prevent "training-serving skew," which is a mismatch in feature logic between training and serving, causing "silent model failures." Option B directly addresses this core problem. Option A describes a technical challenge that the online store component helps mitigate, but it's not the fundamental problem of *governance* itself, which is about consistency of definitions and logic.
Read the full bite: Feature Governance: Taming Your ML Inputs
Question 27 of 30
How do entity embeddings primarily represent and relate high-cardinality categorical data in a neural network?
Show the answer
Answer: d · By learning dense, low-dimensional vectors that position functionally similar categories closer in a continuous space.
Entity embeddings learn dense, low-dimensional vectors for each category, adjusting them during training so that categories with similar functional roles in the prediction task are positioned closer together in the embedding space. Option B describes one-hot encoding, which creates sparse, high-dimensional vectors and does not inherently capture relationships between categories.
Read the full bite: Entity Embeddings: Smart Maps for Categorical Data
Question 28 of 30
After chaining several Dask DataFrame operations, why might a user observe no immediate data processing or results?
Show the answer
Answer: b · Dask operations are lazily evaluated, building a task graph that requires an explicit .compute() call to execute.
Dask uses lazy evaluation, meaning operations only build a task graph without immediate execution. The actual computation is triggered by an explicit call to methods like .compute(). Option D is incorrect because while data loading can take time, the lack of immediate processing after operations is due to lazy evaluation, not just loading.
Read the full bite: Dask: Parallel Computing with Familiar APIs
Question 29 of 30
What is the primary non-technical factor that might prevent a commercial project from utilizing SURF?
Show the answer
Answer: c · The underlying algorithms are patented, necessitating licensing for commercial deployment.
The card explicitly states that the primary reason to avoid SURF in commercial products is its patented algorithms, requiring licenses. The other options describe technical limitations that SURF is designed to overcome or perform well against, making them incorrect.
Read the full bite: SURF: A Faster, Patented Alternative to SIFT
Question 30 of 30
What is ORB's primary mechanism for making its feature descriptors robust to in-plane rotation?
Show the answer
Answer: c · It computes an orientation for each keypoint and rotates the BRIEF sampling pattern accordingly.
ORB achieves rotation invariance by first computing an orientation for each keypoint and then rotating, or 'steering,' the BRIEF descriptor's sampling pattern to match that orientation. Option D is incorrect because FAST detects keypoints, but the rotation invariance for the descriptor is added in the subsequent BRIEF modification.
Read the full bite: ORB: Fast, Free Feature Detection for Computer Vision
Could you explain these out loud?
That is what an interview actually tests. Tezvyn gives you questions like these with what the interviewer is really checking, the answer that lands, and the mistake that ends the conversation, in the four minutes before your next meeting.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.