Variational Inference Expert

Reference Materials

• flow_matching_patterns.md: Flow Matching principles, ODE integration, and DDVI objectives.

Workflows

1. Implementing the Flow Matching Sampler

1. Encode the sensory context ($o$) using a permutation-invariant network (System 1 / LLM).
2. Draw a noise sample $x_0 \sim N(0, 1)$.
3. Use a Neanderthal-optimized ODE solver to integrate the vector field $v_t$ from $t=0$ to $t=1$.
4. Return the resulting state $x_1$ as a sample from the posterior $q(s|o)$.

2. Monitoring Inference Quality (ELBO)

1. Calculate the Complexity term: $KL(q(s) | p(s))$ (Ensuring the posterior is not "too complex").
2. Calculate the Accuracy term: $E_q[\ln p(o|s)]$ (Ensuring the posterior "explains the data").
3. Combine them to monitor the Variational Free Energy ($F = Complexity - Accuracy$).

3. Implementing the Wake-Sleep Objective

1. Wake Phase: Minimize $KL(q(s|o) | p(s,o))$ by updating the recognition network parameters.
2. Sleep Phase: Minimize $KL(p(s,o) | q(s|o))$ by updating the generative model parameters (priors and likelihoods).
3. Use the nREPL Server: Leverage the clojure_eval tool to verify ODE integration, ELBO calculations, and Flow Matching vector fields in real-time. The server is typically available on port 7888. Use it to check convergence of the ODE solver and monitor Free Energy components during belief updating.