ML Depth

Fundamentals

Metrics

  • Regression

  • Classification

  • Ranking

    • Online ranking: ROC-AUC, PR-AUC

    • Offline selection: ROC-AUC

  • Generative Models

    • BLUE, ROUGE

    • GPT Labeling: Accuracy, Human-like

Projects

  • Motivation

  • Approach: Metrics, features, architecture

  • Challenges

  • Outcome

Transformer-Based Models

Initialization

  • Xavier/He initialization

Tokenizers

  • Byte Pair Encoding (BPE)

  • WordPiece

  • SentencePiece

Activations

  • Sigmoid, Tanh, ReLU, LeakyReLU

  • SiLU, GELU, Swish

  • GLU, GeGLU, ReGLU, SwiGLU

Normalization

  • Internal covariate shift

  • Batch normalization

  • Layer normalization (LayerNorm)

  • RMS normalization (RMSNorm)

  • Pre-layer normalization (Pre-LN)

Position Embedding Schemes

  • Learned positional encoding

  • Sinusoidal positional encoding

  • Rotary positional encoding

Attention Mechanisms

  • Functions: Additive, dot product, scaled dot product

  • Architectures:

    • Full (e.g., Flash, Paged, Ring)

    • Sparse

    • Multi-query attention

    • Grouped-query attention

  • Techniques:

    • Self-attention (SHA)

    • Multi-head attention (MHA)

    • Encoder attention

    • Decoder attention with KV-cache

Decoding Techniques

  • Beam search

  • Top-p, Top-k sampling

  • Temperature scaling

  • Speculative decoding

Transformer Architectures

  • Models: PyTorch implementation

    • BERT: MLM

    • GPT: CLM

    • T5: Seq2Seq

  • Papers - Encoder: BERT, RoBERTa, XLM-R - Decoder: GPT, LLama3, DeepSeek - Seq2Seq: T5 Learnings, mT5 - Autoencoder: BART - PLM: XLNet - RTD: Elextra - MoE: Mixtral - State-Space: Mamba - Retriever Embeddings: Generalizable T5-based Retriever (GTR)

Hardware and Optimization

  • Flash attention

  • Quantization (e.g., INT8 LLM)

  • Paged attention

  • Ring attention

LLM Techniques

  • Prompt engineering

  • Prompt tuning

  • Retrieval-Augmented Generation (RAG)

  • LoRA, QLoRA

  • Supervised Fine-Tuning (SFT)

  • Reinforcement Learning with Human Feedback (RLHF)

  • Proximal Policy Optimization (PPO)

  • Direct Preference Optimization (DPO)

Convolution-Based Models

  • Convolution-based models: ConvNet, ResNet, Graph CN, LightGCN, Graph Transformers.

Generative Models

  • Generative models on latent variable space: VAE, VQVAE, GAN, diffusion models, diffusion transformers.

Multimodal Models

TODO

Good to Know

  • Popular NL tasks and remember the dataset names.

  • Kernel methods: kernel meaning embedding, MMD, other IPMs – read every inch of our paper.

  • Probability and statistics: parametric and non-parametric methods for inference, CI, and hypothesis testing framework.

  • Bayes Net (representation, inference, learning).

  • Causality – how to think systematically about finding the root cause of a problem; Bing search causality paper.

  • Latent variable models: K-means, mixture of Gaussians, PCA, kernel PCA, ICA.

  • Clustering: convex, non-convex, evaluation of clustering performance.

  • Regression and discriminative classification: model assumption, interpretation, evaluation – collinearity, the other stuff.

  • Theory as applied problem statement.

  • Code kernel methods, tree methods, regression, VAE, GAN, diffusion models.

Sample Questions

Question sections moved to the question bank in docs/source/gen/interviews/qb.rst.