🎉 75% of content is free forever — Unlock Premium from $10/mo →
CW
Search courses…
💼 Servicesℹ️ About✉️ ContactView Pricing Plansfrom $10

Optimization Theory for Machine Learning: Convex & Non-Convex Methods

AI/ML PremiumOptimization Theory for Machine Learning🟢 Free Lesson

Advertisement

Optimization Theory for Machine Learning: Convex & Non-Convex Methods

Module: AI/ML Premium | Difficulty: PhD Level | Prerequisites: Linear Algebra, Real Analysis


1. Convex Optimization Foundations

1.1 Convex Sets and Functions

Convex vs Non-Convex FunctionsConvex FunctionUnique global minimumLine segment above curveNon-Convex FunctionMultiple local minimaNP-hard in generalConvex Optimization Properties1. Any local minimum is global minimum2. Set of minima is convex3. First-order condition sufficient: ∇f(x*) = 0 ⟺ x* is optimal4. Strong convexity: f(y) ≥ f(x) + ∇f(x)ᵀ(y-x) + (μ/2)‖y-x‖²

2. Gradient Descent Analysis

2.1 Basic Gradient Descent

2.2 Convergence Analysis

Gradient Descent Trajectoryx* (global min)x₀SGD pathGD (deterministic)SGD (stochastic)Convergence RatesGD: O(1/T) for convex, O((1-μ/L)ᵀ) for strongly convexSGD: O(1/√T) with η = O(1/√T)SVRG: O(1/T) variance-reduced

3. Stochastic Gradient Descent

3.1 SGD with Momentum

3.2 Adaptive Learning Rate Methods

Optimizer ComparisonAdamm_t = β₁m + (1-β₁)gv_t = β₂v + (1-β₂)g²θ -= η·m̂/(√v̂+ε)Fast convergenceAdaptive per-param ηMay not generalizeβ₁=0.9, β₂=0.999SGD+Momentumv_t = βv + gθ -= η·vNesterov: v = βv + ∇f(x-ηβv)Better generalizationRequires tuning ηβ=0.9, η=0.01-0.1AdaGradG_t = G_{t-1} + g_t²θ -= η·g/(√G+ε)Learning rate decaysGood for sparse dataAggressive decayη=0.01RMSPropv = βv+(1-β)g²θ -= ηg/√vExponential moving avgFixes AdaGrad decayConvergence RatesIterations TLossAdam: Fast initiallySGD+M: Better finalAdaGrad: Sparse tasksKey insight: Adam convergesfaster but SGD+M oftenfinds better solutions

4. Second-Order Methods

4.1 Newton's Method

4.2 Quasi-Newton Methods

Optimization Methods ComparisonGradient DescentO(n) per iterationO(1/T) convergenceSimple, scalableNo Hessian neededNewton's MethodO(n³) per iterationO(1/2ᵀ) convergenceExpensive HessianFastest near optimumBFGS / L-BFGSO(n²) per iterationSuperlinear convergenceNo explicit HessianGood for moderate nL-BFGSO(mn)m vectorsScalableMemory: O(mn)Trust Region MethodsInstead of line search, solve: min m_t(p) s.t. ‖p‖ ≤ Δ_tm_t(p) = f(x_t) + ∇f(x_t)ᵀp + ½pᵀ∇²f(x_t)pAdjust trust radius Δ based on actual vs predicted reductionMore robust than line search for non-convex problems

5. Non-Convex Optimization

5.1 Landscape Analysis

5.2 Escaping Saddle Points


6. Learning Rate Schedules

Learning Rate SchedulesConstantη_t = η₀Step Decayη_t = η₀ · γ^⌊t/s⌋Exponentialη_t = η₀ · γᵗCosine Annealingη_t = η_min + ½(η_max-η_min)(1+cos(πt/T))Modern Schedule: Warmup + Cosine DecayPhase 1 (Warmup): Linear increase from 0 to η_max over T_w stepsPhase 2 (Decay): Cosine annealing from η_max to η_min over T - T_w stepsWhy warmup? Early gradients are noisy; large η causes instabilityUsed in: Transformer training (Attention Is All You Need), large batch trainingη_t = η_min + ½(η_max - η_min)(1 + cos(π · max(0, t-T_w)/(T-T_w)))

7. Distributed Optimization

7.1 Data Parallelism

7.2 Communication Efficiency

MethodCommunicationCompression
AllReduce per workerNone
Gradient CompressionTop-k, Random-k
Local SGDPeriodic syncNone
Federated Averaging per roundDifferential Privacy

8. Convergence Proofs

8.1 Proof Sketch for SGD


Summary

MethodPer-iteration CostConvergenceBest For
Gradient DescentConvex
SGDLarge-scale
SGD+MomentumDeep learning
AdamSparse/NLP
NewtonSmall-scale
L-BFGSSuperlinearMedium-scale

Research Frontier: Adaptive methods with provable generalization guarantees remain an open problem. Recent work on Sharpness-Aware Minimization (SAM) connects optimization geometry to generalization.

Need Expert AI/ML Premium Help?

Get personalized tutoring, project support, or professional consulting.

Advertisement