M6-071M6: Multimodal & Generative ModelsFlow Matching & Rectified FlowHard
⚡Flashcard Drill➔Mastery:
Path Geometries and Optimal Transport Interpolants in Flow Matching
🎯 Active Recall Self-Check (Draft First, Verify After)
Login-free practice tracker: recall core answer, audit pitfalls, and log review status.
Step 1 · Draft your concise answer & assumptions
Step 2 · Check Pitfalls & Follow-up Inquiries
📐 Mathematical Definition
⚡ One-Sentence Core Answer
Optimal Transport displacement interpolants in Flow Matching construct straight non-crossing probability trajectories that minimize kinetic transport cost, unlocking high-quality generative sampling in minimal integration steps.
📌 Key Takeaways
- •Path geometry spectrum: Independent Linear paths (simple straight lines with random pairings), Variance-Preserving Diffusion paths (curved trajectories), and Optimal Transport paths (cost-minimized non-crossing trajectories)
- •Minibatch Optimal Transport (m-OT): computes pairwise cost matrices within training mini-batches using the Hungarian or Sinkhorn algorithm, pairing noise with geometrically proximate data points
- •Kinetic energy minimization: straightening probability trajectories minimizes total kinetic energy , reducing ODE solver truncation errors
📐 Mathematical Derivations
Mathematical Mechanism: 1. General Probability Interpolants (Albergo & Vanden-Eijnden, 2022): Define interpolant path between base noise and target data : The conditional velocity is: . (a) Diffusion Schedule: (curved trigonometric path). (b) Standard Straight Path: . 2. Optimal Transport (OT) Displacement Formulation: The Monge-Kantorovich 2-Wasserstein transport problem solves: Along the optimal transport plan , the trajectories provably never intersect in space-time. The kinetic energy is minimized: 3. Minibatch OT Matching Protocol: For batch size with noise samples and data samples : Compute pairwise Euclidean distance matrix . Find optimal permutation via linear sum assignment: Train CFM strictly on matched pairs .
🏭 Production Trade-offs
Deep Dive & Engineering Trade-offs: ① The Independent vs OT Pairing Trade-off: Independent random pairing is trivial to implement, but pairs noise vectors on the left of the distribution with data points on the right, forcing millions of trajectories to cross like tangled spaghetti. At every intersection, the learned velocity must predict the average velocity, curving the flow. Minibatch OT sorts pairs so that each noise vector flows toward the closest data point, dramatically reducing path crossing and straightening trajectories. ② Minibatch OT Approximation Limits: Minibatch OT ( or ) solves transport strictly within the local batch; it is not the global Optimal Transport map across the entire infinite dataset. However, empirical studies show that minibatch OT achieves 90% of the trajectory straightening benefits of exact global OT at trivial computational cost ( on CPU using Sinkhorn/Hungarian algorithms). ③ Reflow as Continuous OT Alignment: When minibatch OT is combined with 1-Reflow (retraining on model-generated ODE endpoints), the trajectories become virtually straight lines, enabling 2-4 step generation with high visual fidelity. ⑤ Interview Strategy: Formulate general interpolants , define the 2-Wasserstein kinetic energy minimization objective, describe the Minibatch OT cost assignment matrix , and explain why reducing trajectory crossings allows fewer ODE sampling steps.
⚠️ Common Interview Pitfalls
- ✕Attempting exact global Optimal Transport across millions of dataset samples; exact OT is computationally intractable, making Minibatch OT the required approximation
- ✕Assuming curved diffusion schedules are mathematically required for generative modeling; straight OT paths achieve superior few-step performance
- ✕Confusing independent linear Flow Matching with Optimal Transport Flow Matching; independent linear paths still suffer from severe path intersections
🎯 Interviewer Follow-up Inquiries
- ?Why does Minibatch Optimal Transport (m-OT) within small batches () effectively straighten trajectory flows across the broader dataset?
- ?What is the mathematical relationship between the 2-Wasserstein distance and the kinetic energy of the probability flow vector field?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
📚