Deep Learning Interview Questions

What should you focus on for a deep learning interview?

Which topics came up most in these interviews?

Fundamentals appeared in 27 of 85 interviews, transformers and attention in 24 of 85, and CNNs and computer vision in 19 of 85. Use these topics as a starting point for checking your explanations.

Transformer architecture was the most asked question, appearing in 13 of 85 interviews. But check the basics too:

• Can you explain the difference between AI, machine learning, and deep learning?
• Can you describe a simple neural network and how it works?
• Can you give examples of deep learning applications?
• Can you explain what tensors are and why they are useful in deep learning?

Can you trace how a transformer or CNN works?

Explain the flow, not just the names of the parts. Reviewed transformer answers missed a clear explanation of how token embeddings pass through stacked transformer blocks and how the encoder and decoder connect.

Reviewed CNN answers missed how convolutional filters produce feature maps that capture spatial patterns and help distinguish image classes.

Check your own explanation against those gaps:

• Can you trace token embeddings through the transformer blocks?
• Can you explain the connection between the encoder and decoder?
• Can you explain how filters produce feature maps, then connect those maps to image classification?

Can you explain how a network learns?

Check that you can explain both how backpropagation computes gradients and how gradient descent uses them to update weights. The questions also cover forward propagation, losses, normalization, overfitting, reproducibility, and evaluation.

• Can you describe forward propagation in a simple neural network?
• Can you explain the role of the loss during training?
• Can you explain what overfitting means and how you would evaluate a model?
• Can you explain how you would make a training run reproducible?

Can you explain choices beyond the architecture?

Use the framework, transfer learning, and deployment questions to check practical decisions, not just model definitions.

• Can you explain what PyTorch offers over NumPy for deep learning and what autograd does?
• Can you explain how you would choose a pretrained model and which layers to freeze with limited labeled data?
• Can you explain quantization, pruning, and distillation, and discuss them alongside latency and edge deployment?

Who gets asked Deep Learning

The 20 most asked Deep Learning questions

Counted by interview, not by mention: a question asked twice in one conversation counts once.

1. How does the transformer architecture work, and what are its main building blocks?

   Asked in 13 of 85 interviews · last asked Oct 14 · Concept · Junior

   Show answer

   A transformer processes token embeddings through stacked attention and feed-forward blocks, using position information to track token order. Self-attention lets each token weigh other tokens and mix their information. Each block also uses residual connections and layer normalization. In an encoder-decoder transformer, the decoder uses causal self-attention and cross-attention to encoder outputs to predict the next token.

   • Multi-head attention uses queries, keys, and values to learn different relationships between tokens.
   • Feed-forward networks transform each token’s representation; residual connections and normalization help training.
   • Causal masking prevents attention to future tokens; an output layer maps decoder representations to token scores.
   • Encoder-only and decoder-only transformers also exist; not every transformer has both parts.

   What the interviewer checks: The interviewer checks whether you can trace token information through the model and explain its main components.

   Common mistake (in 4 of 7 reviewed answers): Many answers forget that embeddings pass through stacked blocks and, in encoder-decoder models, decoder cross-attention reads encoder outputs.

   Other ways it was asked

   • “What is transformer architecture, and how does it work?”
   • “What are transformer architectures, and how do they work?”
   • “What is the Transformer architecture?”

2. How does a CNN work, and what are the main parts of its architecture?

   Asked in 7 of 85 interviews · last asked Sep 23 · Concept · Junior

   Show answer

   A CNN learns spatial patterns using small filters that slide across an image. Each filter applies the same learned weights to local regions, producing a feature map. Convolution layers and activations such as ReLU build increasingly complex features; pooling or strided convolutions can reduce spatial size. A final prediction layer uses these features to distinguish image classes.

   • Filters share weights across locations, reducing parameters compared with fully connected layers.
   • Early layers often learn edges and textures; deeper layers combine them into more complex patterns.
   • ReLU adds nonlinearity; pooling or strided convolutions reduce feature map size.
   • Training adjusts filter and prediction-layer weights through backpropagation.

   What the interviewer checks: The interviewer checks whether you can connect convolution, feature maps, and the main layers to image classification.

   Common mistake (in 3 of 5 reviewed answers): Many answers forget that filters produce feature maps of spatial patterns that help distinguish image classes.

   Other ways it was asked

   • “Can you explain CNN architecture?”
   • “Can you explain the CNN architecture?”
   • “Can you explain in more depth what CNNs are and how they work?”

3. What is the difference between AI, machine learning, and deep learning?

   Asked in 6 of 85 interviews · last asked Aug 9 · Concept · Junior

   Show answer

   AI is the broad field of building systems that perform tasks requiring human intelligence. Machine learning is a part of AI where models learn patterns from data. Deep learning is a part of machine learning that uses neural networks with many layers to learn complex patterns.

   • AI can use fixed rules; it does not always involve learning.
   • Deep learning often needs more data and computing power than simpler machine learning methods.
   • Generative AI creates content, such as text or images; modern systems typically use deep learning.

   What the interviewer checks: The interviewer checks whether you understand how these fields fit together and what makes each different.

   Other ways it was asked

   • “What is the difference between machine learning and deep learning?”
   • “What is the difference between AI and deep learning?”
   • “What's the difference between AI, machine learning, deep learning, and generative AI?”

4. What is deep learning, and can you give some examples?

   Asked in 4 of 85 interviews · last asked Sep 19 · Concept · Junior

   Show answer

   Deep learning is a type of machine learning that uses neural networks with many layers to learn patterns from data. These layers learn useful features, such as edges and shapes in images, rather than relying only on features designed by people. Examples include recognizing objects in photos, converting speech to text, and generating text with large language models.

   • Deep learning is part of machine learning, which is part of artificial intelligence.
   • Multiple layers learn increasingly complex patterns from data.
   • Training often needs large amounts of data and computing power.

   What the interviewer checks: The interviewer is checking whether you understand what makes deep learning different and can name practical uses.

   Other ways it was asked

   • “What do you understand by deep learning?”
   • “What is deep learning?”
   • “Can you give me an example of deep learning?”

5. Can you explain the structure of a simple neural network and how it works?

   Asked in 3 of 85 interviews · last asked Jul 31 · Concept · Junior

   Show answer

   A simple neural network has an input layer, one or more hidden layers, and an output layer. Each hidden neuron combines inputs using weights and a bias, then applies an activation function. The output gives a prediction. During training, backpropagation computes gradients of the prediction error, and an optimizer uses them to adjust weights and biases.

   • The input layer receives features; the output layer produces predictions.
   • Nonlinear activation functions help the network learn complex patterns.
   • A loss function measures how wrong predictions are.
   • Training adjusts weights and biases to reduce the loss.

   What the interviewer checks: The interviewer checks whether you can explain the layers, prediction flow, and basic training process.

   Other ways it was asked

   • “Can you explain neural networks in simple terms?”
   • “Can you explain the architecture of a simple neural network?”
   • “How do neural networks work?”

6. How does backpropagation compute gradients, and how does gradient descent use them to update weights?

   Asked in 3 of 85 interviews · last asked Sep 23 · Concept · Junior

   Show answer

   Backpropagation computes the loss gradient for each weight by applying the chain rule backward through the network. Gradient descent then updates each weight by subtracting its gradient multiplied by the learning rate. Backpropagation computes the gradients; gradient descent uses them to reduce the loss.

   • A forward pass computes predictions and the loss.
   • The backward pass reuses intermediate results to compute gradients efficiently.
   • The learning rate controls update size; too large a step can increase the loss.

   CodeCopy

   For each weight w:
   g = gradient of the loss with respect to w
   w = w - learning_rate * g

   What the interviewer checks: The interviewer checks whether you understand the chain rule and distinguish gradient computation from weight updates.

   Other ways it was asked

   • “Explain backpropagation.”
   • “Can you explain the backpropagation method?”
   • “How do backpropagation and gradient descent work together to update a neural network's weights?”

7. What is convolution, and how does a convolutional layer work?

   Asked in 3 of 85 interviews · last asked Oct 10 · Concept · Mid-level

   Show answer

   Convolution combines nearby input values using a small filter that slides across the input. At each position, a convolutional layer multiplies input values by learned filter weights, sums the results, and adds a bias if used. Each filter produces a feature map. Sharing weights across positions helps detect patterns, such as edges, with fewer parameters.

   • Filters usually span all input channels; each filter produces one output channel.
   • Stride controls how far filters move; padding controls how input borders are handled.
   • Training learns filter weights and biases through backpropagation.
   • Most deep learning libraries use cross-correlation: they do not flip the filter as mathematical convolution does.

   What the interviewer checks: The interviewer checks whether you understand local filtering, shared weights, and how filters produce feature maps.

   Other ways it was asked

   • “What is convolution, and how is it used in neural networks?”
   • “What is convolution, and how does it work in a neural network?”
   • “Can you explain how a convolutional layer works in deep learning?”

8. What is PyTorch, and what advantages does it offer over NumPy for deep learning?

   Asked in 3 of 85 interviews · last asked Oct 17 · Concept · Junior

   Show answer

   PyTorch is a library for building and training deep learning models with tensors. Compared with NumPy alone, it adds automatic gradient calculation, GPU support, and tools such as neural network layers and optimizers. These features make training models easier and can speed up large computations.

   • Tensors are similar to NumPy arrays but can also run on supported GPUs.
   • Automatic differentiation computes gradients needed to train models.
   • Built-in layers, loss functions, and optimizers reduce manual coding.
   • NumPy alone provides CPU array operations, not automatic differentiation or deep learning tools.

   What the interviewer checks: The interviewer checks whether you understand why deep learning needs more than basic array operations.

   Other ways it was asked

   • “What is PyTorch, and where is it used in real-world applications?”
   • “What is PyTorch?”
   • “What is PyTorch, and what advantages does it provide for deep learning compared with using NumPy alone?”

9. What is a multilayer perceptron, and how does it work?

   Asked in 2 of 85 interviews · last asked Oct 15 · Concept · Junior

   Show answer

   A multilayer perceptron (MLP) is a feedforward neural network with an input layer, one or more hidden layers, and an output layer. Each neuron combines weighted inputs with a bias, then applies an activation function. Nonlinear hidden activations let it learn complex patterns. Training uses backpropagation and an optimizer to adjust weights and biases to reduce prediction error.

   • Layers are usually fully connected, and information moves from input to output without loops.
   • Without nonlinear hidden activations, stacking layers cannot learn nonlinear relationships.
   • MLPs can solve classification and regression problems, especially with fixed-size feature inputs.
   • Scale input features and evaluate on held-out data to check how well the model generalizes.

   What the interviewer checks: The interviewer checks whether you understand an MLP’s structure, forward computation, training, and basic uses.

   Other ways it was asked

   • “What is a multilayer perceptron, and when would you use one?”
   • “How does a multilayer perceptron work, and how do you solve a problem using one?”

10. What are tensors, and why are they useful in deep learning and PyTorch?

    Asked in 2 of 85 interviews · last asked Oct 17 · Concept · Junior

    Show answer

    Tensors are arrays of numbers with zero or more dimensions. They represent inputs, model weights, and outputs in deep learning. PyTorch tensors support fast operations on CPUs and GPUs, and they can track operations so PyTorch can compute gradients for training.

    • A scalar is zero-dimensional, a vector is one-dimensional, and a matrix is two-dimensional.
    • Tensor shape describes the size of each dimension.
    • Tensor operations let models process whole batches efficiently.
    • PyTorch can track gradients for floating-point and complex tensors when gradient tracking is enabled.

    What the interviewer checks: The interviewer checks whether you understand tensor dimensions, their role in models, and PyTorch’s computation and gradient support.

    Other ways it was asked

    • “What are tensors, and how are they used in machine learning?”
    • “What are the advantages of using tensors?”
    • “What is a PyTorch tensor?”

11. Describe some common machine learning architectures and when you would use them.

    Asked in 2 of 85 interviews · last asked Oct 19 · Concept · Mid-level

    Show answer

    Feedforward networks suit fixed-size inputs, such as tabular features for classification or regression. Convolutional neural networks (CNNs) learn local patterns and work well for images. Recurrent neural networks (RNNs), including LSTMs, process sequences such as time series. Transformers use attention to connect information across inputs and are widely used for text, images, and audio.

    • Feedforward networks do not directly model spatial structure or sequence order.
    • CNNs reuse filters across positions, but capturing distant relationships can require deeper layers.
    • RNNs process sequences step by step, limiting parallel training; LSTMs help learn longer dependencies.
    • Transformers support parallel training, but standard attention has memory and computation costs that grow quadratically with sequence length.

    What the interviewer checks: The interviewer checks whether you can match architectures to data types and explain their main trade-offs.

    Other ways it was asked

    • “Describe common machine learning architectures, their typical use cases, and briefly note any limitations.”

12. What are the main limitations of CNNs, RNNs, and transformers, and when would you choose another architecture?

    Asked in 2 of 85 interviews · last asked Oct 19 · Concept · Senior

    Show answer

    CNNs can struggle with global context unless designed for it. RNNs limit parallel training and can struggle with long dependencies. Standard full-attention transformers become costly as sequences grow. I would consider graph networks for connected data, boosted trees for tables, or state-space models for long sequences.

    • CNNs favor local patterns; larger receptive fields help capture wider context.
    • LSTMs and GRUs improve long-term memory, but their recurrent steps still limit parallel training.
    • Full attention has quadratic computation cost; memory-efficient implementations avoid storing the full attention matrix.
    • Choose based on data structure, training data, accuracy, latency, and memory; compare against a simple baseline.

    What the interviewer checks: The interviewer checks whether you understand architecture trade-offs and can choose a model for practical constraints.

    Other ways it was asked

    • “What are the main limitations of CNNs, RNNs, and transformers?”
    • “What are common limitations of machine learning architectures, and when would you choose a different architecture?”

13. How do you choose a model architecture, especially when data or computing resources are limited?

    Asked in 2 of 85 interviews · last asked Oct 19 · Scenario · Senior

    Show answer

    I choose an architecture based on the task, data type, data size, and compute limits. I start with a simple baseline and compare suitable models on validation data. With little data, I prefer a smaller model or adapt a pretrained model. With limited compute, I measure training cost, prediction speed, and memory use before choosing.

    • Match the architecture to the input: images, text, sequences, or tabular data.
    • Pretrained models help when their training data is relevant; freezing layers can reduce training cost.
    • Use validation results to compare models; keep the test set for the final evaluation.
    • Choose the simplest model that meets accuracy, speed, and memory needs.

    What the interviewer checks: The interviewer checks whether you balance task fit, validation results, overfitting risk, and practical resource limits.

    Other ways it was asked

    • “How do you choose the right architecture for a problem when multiple models could work?”
    • “How do you choose the right architecture for a problem, especially when data or resources are limited?”

14. What are activation functions, and why do neural networks need them?

    Asked in 2 of 85 interviews · last asked Aug 27 · Concept · Junior

    Show answer

    Activation functions transform a neuron’s weighted input into its output. Nonlinear activations let neural networks learn complex patterns. Without them, stacking linear layers is equivalent to one linear layer, even with biases. Common choices include ReLU in hidden layers and sigmoid for binary classification outputs.

    • Nonlinearity allows deep networks to learn more than a single linear layer.
    • Activation choice affects gradient flow and how easily the network trains.
    • Output activations depend on the task; regression often uses a linear output.

    What the interviewer checks: The interviewer is checking whether you understand why depth alone cannot replace nonlinearity.

    Other ways it was asked

    • “Why do we use activation functions in deep learning?”
    • “What are activation functions?”

15. How do YOLO and the R-CNN family differ, and when would you choose each?

    Asked in 2 of 85 interviews · last asked Aug 27 · Concept · Mid-level

    Show answer

    YOLO is a one-stage detector that predicts boxes and classes directly. The R-CNN family first proposes regions, then classifies them and refines their boxes. I would choose YOLO for low-latency video or edge devices, and consider Faster R-CNN when accuracy matters more than speed. Neither always wins; compare models on your data and hardware.

    • Original R-CNN processes each proposed region separately, making it slow.
    • Fast R-CNN shares image features; Faster R-CNN also learns region proposals.
    • YOLO usually offers lower latency, but speed depends on model size, hardware, and image resolution.
    • Compare detection accuracy, especially for small objects, alongside latency and memory use.

    What the interviewer checks: The interviewer checks whether you understand detection stages and can choose a model for practical constraints.

    Other ways it was asked

    • “What is the difference between YOLO and Faster R-CNN?”
    • “What is a use case for YOLO, and under what conditions might it be preferable to Faster R-CNN?”
    • “Can you compare and contrast different deep learning architectures for object detection, such as the R-CNN family versus YOLO? In what specific scenarios would you choose one over the other?”

16. Can you explain what generative AI is?

    Asked in 2 of 85 interviews · last asked Jul 31 · Concept · Junior

    Show answer

    Generative AI is AI that learns patterns from training data and uses them to create content, such as text, images, audio, or code. It often generates this content in response to a prompt. Its output can look convincing but may be incorrect or biased.

    • It generates content rather than only classifying inputs or predicting a value.
    • Training data shapes what it can generate and the biases it may show.
    • Examples include language models for text and diffusion models for images.

    What the interviewer checks: The interviewer is checking whether you understand what generative AI does and its basic limits.

    Other ways it was asked

    • “What is generative AI, and what does it do?”

17. What is a Vision Transformer, and how does it apply transformers to image data?

    Asked in 2 of 85 interviews · last asked Oct 14 · Concept · Mid-level

    Show answer

    A Vision Transformer (ViT) is a neural network that processes an image as a sequence of patches. It maps each patch to a vector and adds position information. A transformer encoder uses self-attention to learn relationships between patches. For classification, a prediction layer uses a special class token or pooled patch outputs.

    • Patch vectors act like word tokens in a text transformer.
    • Position information tells the model where each patch belongs.
    • Self-attention connects patches across the image, not just nearby regions.
    • Standard self-attention cost grows quadratically with the number of patches.

    What the interviewer checks: The interviewer checks whether you understand how images become tokens and how self-attention combines their information.

    Other ways it was asked

    • “How can transformers be applied to image data?”
    • “What is a ViT (Vision Transformer)?”

18. What is an attention mechanism, and how does it work in a transformer?

    Asked in 2 of 85 interviews · last asked Oct 15 · Concept · Mid-level

    Show answer

    Attention lets each token gather information from other tokens based on their relevance. In a transformer, learned projections create query, key, and value vectors. Query–key dot products are scaled, then softmax turns them into attention weights. Each output is a weighted sum of the value vectors.

    • Self-attention gets queries, keys, and values from the same sequence.
    • Dot products are divided by the square root of the key dimension to keep scores manageable.
    • Multi-head attention learns several attention patterns in parallel.
    • Causal masking prevents tokens from attending to future tokens during next-token prediction.

    What the interviewer checks: The interviewer checks whether you understand how attention weights are computed and used to combine token information.

    Other ways it was asked

    • “What is attention in a transformer exactly?”
    • “What is an attention mechanism?”

19. How does a mixture-of-experts model work, and why are only some parameters active at a time?

    Asked in 2 of 85 interviews · last asked Jul 16 · Concept · Senior

    Show answer

    A sparse mixture-of-experts model has several expert networks and a router that picks a few experts for each token. Only the chosen experts and shared layers run for that token. This lets the model have many total parameters while using much less computation than running every expert.

    • The router learns which experts to select for each token.
    • Active parameters include selected experts and shared layers, not every expert.
    • Sparse routing reduces computation, but all expert weights still need storage.
    • Training often encourages balanced expert use so a few experts do not receive nearly all tokens.

    What the interviewer checks: The interviewer is checking whether you can explain routing and distinguish total parameters from parameters used per token.

    Other ways it was asked

    • “What do you know about mixture of experts?”
    • “What is mixture of experts?”
    • “How does a model have 100 billion-plus parameters but only activate a small number at a time?”

20. What is self-attention, and what does it enable in a transformer?

    Asked in 2 of 85 interviews · last asked Oct 14 · Concept · Junior

    Show answer

    Self-attention lets each token combine information from other allowed tokens in the same sequence. It uses query and key vectors to compute weights, then takes a weighted sum of value vectors. This gives transformers context-aware representations and helps them connect distant tokens without processing the sequence step by step.

    • Queries, keys, and values come from the same input sequence.
    • Attention weights depend on token content, so connections change with the input.
    • Causal masking prevents a token from attending to future tokens.
    • Training can process token positions in parallel; standard autoregressive generation still proceeds token by token.

    What the interviewer checks: The interviewer checks whether you understand how self-attention builds context and how masking affects which tokens can interact.

    Other ways it was asked

    • “What is self-attention?”
    • “What feature does self-attention add to Transformer architecture?”

Related interview questions

• Machine Learning Engineer interview questions
• AI Engineer interview questions
• Data Scientist interview questions
• Data Analyst interview questions
• Software Engineer interview questions
• Machine Learning interview questions
• Computer Vision interview questions
• Generative AI and LLM interview questions
• RAG interview questions
• Agentic AI interview questions
• MLOps interview questions

How this list is made

Every question here was asked by a real interviewer in a real job interview that a candidate ran MeetAssist in. After the call, the interviewer's questions are pulled from the transcript. Practice sessions, tests, exams and anything that was not a job interview are dropped.

Questions are rewritten so they make sense on their own: names of people, employers and clients are removed, and speech-to-text mistakes are fixed. Small talk, logistics and questions about one candidate's own résumé are left out. Rephrasings of the same question are merged, and the count is the number of different interviews it came up in.

The interviews come from candidates who use MeetAssist, so they lean towards software, data and AI roles: the shares describe these interviews, not every Deep Learning interview. Answers to the most asked questions are written and checked by an editor; a "common mistake" is shown only when reviews of real answers in these interviews show it, with the count.

85 interviews, 62 candidates, Feb 2026 to Oct 2026. Updated every month with new interviews.