1 · Machine–Learning Model Taxonomy 🔗
1.1 Discriminative vs Generative 🔗
-
Discriminative models learn the conditional distribution $P(y\mid x)$ to separate classes.
- Primary use · classification/regression.
- Typical nature · often deterministic at inference.
- Examples · Logistic Regression, SVM, plain Feed‑forward NNs, ResNet‑like CNNs.
-
Generative models learn the joint distribution $P(x,y)=P(x\mid y)P(y)$ and can synthesize new x.
- Primary use · both classification and data generation/simulation.
- Typical nature · probabilistic/stochastic.
- Examples · Naïve Bayes, Hidden Markov Models, GANs, VAEs, Diffusion models.
| Aspect | Discriminative | Generative |
|---|---|---|
| Learns | $P(y\mid x)$ | $P(x,y)$ or $P(x\mid y)P(y)$ |
| Produces | Class label / score | Sample, likelihood or label |
| Strength | Decision boundaries | Full data distribution |
| Weakness | Needs labelled data | Harder to train/scale |
1.2 Deterministic vs Probabilistic 🔗
| Aspect | Deterministic | Probabilistic |
|---|---|---|
| Output | Same value for same input | Distribution or random sample |
| Pros | Fast, predictable | Express uncertainty, robust to missing data |
| Cons | Ignores uncertainty | More compute, stochastic results |
| Examples | Decision‑Tree inference, SVM, frozen CNN | Naïve Bayes, Gaussian Mixture, Bayesian Network, VAE decoder |
2 · Vectors & Modalities 🔗
2.1 Vectors (Embeddings) 🔗
Definition · Numeric representations that encode the semantics of entities (words, images, users …).
- Enable similarity search (
cosine,dot), clustering, recommendation, retrieval‑augmented generation. - Learned via Word2Vec, GloVe, fastText, BERT, CLIP, etc.
2.2 Modal vs Multimodal 🔗
-
Modal = one input type (text or image or audio).
-
Multimodal = multiple types jointly (e.g., image + caption).
- Rising trend: vision–language models (e.g., GPT‑4V, Gemini), audio‑text (e.g., Whisper), video‑text.
2.3 Neural Language Models 🔗
Large neural nets pretrained on text with self‑supervised objectives.
- Families: GPT‑n (decoder‑only), BERT (encoder‑only), T5 / BART (encoder‑decoder), LLaMA series.
- Tasks: text generation, summarization, translation, code completion, search ranking.
3 · Transformer Family 🔗
3.1 Core Architecture (“Attention is All You Need”, 2017) 🔗
-
Input embeddings + positional encodings.
-
Repeated N × blocks:
- Multi‑head self‑attention → Add & LayerNorm.
- Position‑wise Feed‑Forward NN → Add & LayerNorm.
-
(Decoder adds masked self‑attention + cross‑attention to encoder outputs.)
3.2 Self‑Attention (Quick Intuition) 🔗
Each token forms a Query (Q) vector that is matched against Keys (K) of every token; the similarities weight the Values (V) to build a context‑aware representation.
3.3 Transformer Variants 🔗
| Variant | Architecture | Flagship Models | Typical Tasks |
|---|---|---|---|
| Encoder‑only | Auto‑encoding | BERT, RoBERTa | Classification, QA, embeddings |
| Decoder‑only | Auto‑regressive | GPT, LLaMA‑2 Chat | Text/code generation |
| Encoder‑Decoder | Seq‑to‑Seq | T5, BART, Pegasus | Translation, summarization |
3.4 Language‑Modeling Objectives 🔗
| Objective | Context Used | Predicts | Archetype |
|---|---|---|---|
| Masked (MLM) | Bidirectional | Masked tokens | BERT |
| Autoregressive (AR) | Left‑to‑right (or right‑to‑left) | Next token | GPT |
| Prefix LM | Past tokens (unmasked) + full prefix | Next token | T5 pre‑training |
4 · Learning Paradigms 🔗
| Paradigm | Labels | Model Learns | Canonical Example |
|---|---|---|---|
| Supervised | Explicit | Map x → y | ImageNet classification |
| Unsupervised | None | Structure of x | Word2Vec, PCA, clustering |
| Self‑Supervised | Labels from data | Predict masked / future parts | GPT pre‑training, BERT MLM |
5 · Model Adaptation & Fine‑Tuning 🔗
| Technique | Data Need | Compute | Brief |
|---|---|---|---|
| Prompt Engineering | None (zero‑shot) | – | Steer behavior via instructions/examples |
| Supervised Fine‑Tuning (SFT) | Labelled pairs | High | Adjust all weights to task domain |
| LoRA / Adapters | Labelled pairs | Low–Med | Train tiny rank‑update layers; mergeable |
| RLHF | Human preference scores | Very High | Align model to helpful/safe outputs |
6 · Prompt Engineering 🔗
6.1 Anatomy of a Good Prompt 🔗
Instruction → Context → Input → Output‑format → Tone/Role
6.2 Prompting Techniques 🔗
| Technique | Best For | Key Idea |
|---|---|---|
| Zero‑Shot | Simple, common tasks | Ask directly |
| Few‑Shot | Pattern imitation | Give 2–5 exemplars |
| Chain‑of‑Thought | Reasoning/maths | “Let’s think step by step” |
| Self‑Consistency | Reliable CoT answer | Sample K reasoning paths, majority vote |
| ReAct | Tool‑using agents | Interleave reasoning & external actions |
| Tree‑of‑Thought | Complex planning | Explore multiple branches, backtrack |
| Retrieval‑Augmented (RAG) | Factual or domain answers | Retrieve docs → feed as context |
6.3 Security Concerns in Prompting 🔗
- Prompt injection / jailbreaks
- Data leakage (keys, PII)
- Prompt leakage (system prompt exposure)
- Malicious content generation (spam, phishing, code exploits)
- Token flooding / prompt DOS
Mitigations · input sanitation, guardrail LLMs, content filters, max‑token limits, rate‑limits, red‑teaming.
7 · Foundation Models 🔗
Large, self‑supervised, general‑purpose models adaptable to many downstream tasks.
- Examples : GPT‑4 (text), CLIP (image+text), DALL·E (image), SAM (vision segmentation).
- Benefits : reuse, performance, economy of scale.
- Risks : bias, compute cost, ecological footprint, opacity.
8 · Scaling Laws & Emergent Abilities 🔗
- Empirical power‑law links between loss ↔ parameters, data, compute (OpenAI, DeepMind).
- Emergence : qualitative jumps (few‑shot learning, tool use) above certain scale (≈10 B+, 100 B+ params).
- Implications : unpredictable behaviours, but strong generalization—drives interest in alignment & evals.
9 · Variational Autoencoder (VAE) — Quick Recap 🔗
- Encoder → μ, σ² (latent distribution).
- Reparameterization trick → sample z.
- Decoder → reconstruct/generate x̂.
- Loss = Reconstruction Loss + KL‑Divergence( q(z|x) ‖ N(0,1) ).
| Strength | Why It Matters |
|---|---|
| Generative | New images/text variants |
| Smooth latent space | Interpolation, arithmetic |
| Structured | Semi‑supervised & controllable |
| Stable training | No adversarial collapse (vs GAN) |
10 · Quick Cheat‑Sheet — Which Technique When? 🔗
- Need a fast tweak? → Prompt Engineering.
- Domain‑specific answers? → SFT or LoRA.
- Politeness / helpfulness? → RLHF.
- Up‑to‑date factuality? → RAG.
- Creative image/text synthesis? → Diffusion, VAE, GAN.