Explain CNN shapes, params, and trade-offs

Q: Explain CNN shapes, params, and trade-offs

This is a Machine Learning interview question from Apple for Data Scientist roles. View the full question and solution on PracHub.

Q: How do I approach Machine Learning interview questions?

Machine Learning questions require understanding of core concepts and practice. PracHub provides solutions with explanations to help you master machine learning interviews.

Question

CNN Shapes, Compute, and Design Trade-offs

Context

You are given an input tensor X with shape H×W×C = 64×64×3. Consider the following convolutional neural network (CNN):

L1: Conv 3×3, stride=1, padding=1, 32 filters
L2: MaxPool 2×2, stride=2
L3: Conv 3×3, dilation=2, stride=1, padding=2, 64 filters
L4: Depthwise separable conv: depthwise 3×3 (stride=1, padding=1) + pointwise 1×1 to 128 channels

Tasks

(a) Compute the output shape H×W×C after each layer.

(b) Compute parameter counts and MACs for L1, L3, and L4. Compare L4’s MACs to a standard 3×3 conv with 64→128 channels (same input size as L4). Show formulas and numbers.

(c) Compute the receptive field size (in input pixels) of a single activation after L4.

(d) Where would you place BatchNorm relative to activation for stable training (pre‑act vs post‑act) and why?

(e) When would you prefer stride vs pooling vs dilation to preserve information while controlling compute?

(f) Explain the bias–variance and optimization trade‑offs of using depthwise separable convolutions in tiny‑model regimes.

Explain CNN shapes, params, and trade-offs

CNN Shapes, Compute, and Design Trade-offs

Context

Tasks

Solution (Locked)

Comments (0)