Adobe¶
Table of Contents
ML Practicals¶
Experience with LLMs
ML Ops
Experience with different parts of the stack - analysis, feature design, feature pipelines, experimentation, registry/versioning, deployment, feature logging, monitoring
How would you measure latency increase?
Project discussion - Model boundary
Model evaluation
Metrics
Explain ROC-AUC and PR-AUC in detail. In what scenarios, one is better than the other?
Does a high ROC-AUC imply overfitting? Why/why not? How would you detect overfitting?
Slices
How to choose slices for evaluation? Numerical, categorical, temporal, edge cases.
Feature design
How would you design features with numerical and categorical signals
Productionizing
How would you select the criteria for marking a model productionizable?
Is it metric based? How would you select the threshold?
How would you evaluate your model in presence of feedback loops from production models?
Golden dataset creation
What are the criteria for selecting a golden set for offline evaluation?
Would the criteria for automatic productionizing be any different?
How often would you update it?
How would you ensure that it has balanced representation for rare but important cases?
Systems
Experience with map reduce
ML Theory/Design¶
Master’s Thesis
Explain submodular optimisation with examples.
Experience with LLMs
Explain different types of finetuning tasks. Is knowledge distillation the same as finetuning?
LoRA
Explain how LoRA works.
Explain rank. Why is row rank = column rank?
Suppose you change k entries in a matrix. How much can the rank change?
What is the maximum rank of sum of k rank 1 matrices?
Why cannot we use LoRA from the beginning of the training?
Sampling
Given a dataset of 1M examples, how would you design a sampling mechanism to select 10k samples?
What is the probability that \(x_i\) and \(x_j\) are both part of that selected set?
How can you convince a non-technical person that each example is equally likely to be selected?
Transformers
Explain the difference between encoder, decoder and encoder-decoder architecture.
What is the context width of BERT?
How would you solve the SQuAD task using a transformer architecture? How would you model the problem and which architecture would you choose?
Initialisation
How would you initialise the weights? Which technique to use when?
What impact would a specific type of initialisation technique have during model training? What is it aiming to solve?
Batch Norm
Explain batch norm and internal covariate shift.
How exactly does the mena and variance calculation work for a given batch?
Instead of sample mean, can we use any other convex combination of samples to estimate expectation?
What is the difference between biased and unbiased estimators of the variance? Explain bias of an estimator.
Contextual bandits
How would you approach contextual bandits to tackle assets generated online?