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.

What difficulty level is this interview question?

This is a hard difficulty Machine Learning question, commonly asked during Onsite rounds at Microsoft.

What role is this question designed for?

This question is commonly asked for Data Scientist candidates at Microsoft during technical interviews.

Compare CNN/RNN/LSTM and implement K-means | Microsoft Interview Question

Quick Overview

This question evaluates understanding of deep learning architectures (CNN, RNN, LSTM) and unsupervised clustering implementation (K-means), covering inductive biases, parameter sharing and receptive fields, LSTM gating and vanishing-gradient behavior, tensor shape reasoning, algorithmic complexity, initialization, and cluster handling.

Deep Learning Concepts and K-means Implementation (Onsite ML Interview)

Part A: CNNs vs RNNs and LSTMs

Contrast CNNs and RNNs for the following modalities:

(i) 224×224 RGB images
(ii) Variable-length text

Explain and compare:

Inductive biases: translation equivariance, locality (spatial/temporal), temporal order
Parameter sharing
Receptive-field growth
Ability to model long-range dependencies
When a 1D CNN can replace an RNN/Transformer (assumptions and caveats)

For LSTMs:

Write the gate equations and define all symbols and shapes
Show mathematically how the cell state helps mitigate vanishing gradients
Compute output shapes for input (batch=32, seq_len=100, feat=64), hidden size=128 for unidirectional vs bidirectional cases (assume a single layer)

Part B: K-means From Scratch

Implement K-means with the following requirements:

k-means++ initialization
Vectorized assignment and update steps
Convergence based on decrease of the objective (sum of squared distances)

Also discuss:

Time complexity O(n·k·d) per iteration and memory trade-offs
Handling empty clusters
Feature scaling and outliers
Choosing k (silhouette, elbow, BIC)
Prove the objective is non-increasing per iteration
Propose a mini-batch variant and when you would use it

Quick Overview

Part A: CNNs vs RNNs and LSTMs

Contrast CNNs and RNNs for the following modalities:

(i) 224×224 RGB images

(ii) Variable-length text

Explain and compare:

Inductive biases: translation equivariance, locality (spatial/temporal), temporal order

Parameter sharing

Receptive-field growth

Ability to model long-range dependencies

When a 1D CNN can replace an RNN/Transformer (assumptions and caveats)

For LSTMs:

Write the gate equations and define all symbols and shapes

Show mathematically how the cell state helps mitigate vanishing gradients

Compute output shapes for input (batch=32, seq_len=100, feat=64), hidden size=128 for unidirectional vs bidirectional cases (assume a single layer)

Part B: K-means From Scratch

Implement K-means with the following requirements:

k-means++ initialization

Vectorized assignment and update steps

Convergence based on decrease of the objective (sum of squared distances)

Also discuss:

Time complexity O(n·k·d) per iteration and memory trade-offs

Handling empty clusters

Feature scaling and outliers

Choosing k (silhouette, elbow, BIC)

Prove the objective is non-increasing per iteration

Propose a mini-batch variant and when you would use it

Compare CNN/RNN/LSTM and implement K-means

Quick Overview

Deep Learning Concepts and K-means Implementation (Onsite ML Interview)

Part A: CNNs vs RNNs and LSTMs

Part B: K-means From Scratch

Solution

Comments (0)

Compare CNN/RNN/LSTM and implement K-means

Quick Overview

Deep Learning Concepts and K-means Implementation (Onsite ML Interview)

Part A: CNNs vs RNNs and LSTMs

Part B: K-means From Scratch

Solution

Comments (0)