The Lazy Flipper: MAP Inference in Higher-Order Graphical Models by Depth-limited Exhaustive Search

This article presents a new search algorithm for the NP-hard problem of optimizing functions of binary variables that decompose according to a graphical model. It can be applied to models of any order and structure. The main novelty is a technique to constrain the search space based on the topology of the model. When pursued to the full search depth, the algorithm is guaranteed to converge to a global optimum, passing through a series of monotonously improving local optima that are guaranteed to be optimal within a given and increasing Hamming distance. For a search depth of 1, it specializes to Iterated Conditional Modes. Between these extremes, a useful tradeoff between approximation quality and runtime is established. Experiments on models derived from both illustrative and real problems show that approximations found with limited search depth match or improve those obtained by state-of-the-art methods based on message passing and linear programming.

💡 Research Summary

The paper introduces the “Lazy Flipper,” a deterministic depth‑limited exhaustive search algorithm for MAP inference in binary graphical models of arbitrary order and topology. The core idea is to restrict the search to subsets of variables that are connected through the model’s potentials, thereby eliminating redundant examinations of disconnected or duplicate variable sets. Two specialized data structures enable this restriction: (1) a Connected Subgraph Tree (CS‑tree) that uniquely represents every connected variable subset as a lexicographically smallest CSR‑sequence, guaranteeing that each connected subgraph is visited exactly once; and (2) a Tag List that records variables affected by recent flips, allowing the algorithm to revisit only those subsets whose energy might change after a flip.

The algorithm proceeds iteratively. Starting from an arbitrary initial labeling (often the minimizer of unary potentials), it first attempts single‑variable flips. When no single flip improves the energy, it incrementally increases the subset size n up to a user‑specified maximum n_max. For each n, the CS‑tree is traversed in size‑order, generating all connected subsets of size n. Each subset is evaluated; if flipping the entire subset reduces the total energy, the current labeling is updated greedily and all variables involved are tagged. After the “exploration” phase, a “revisiting” phase uses the Tag List to re‑evaluate all subsets (of size ≤ n) that intersect the recently flipped variables, ensuring that cascading improvements are not missed. The process repeats until no improving flip exists for the current n, then n is increased. If n_max equals the total number of variables, the algorithm exhaustively explores all connected subsets and is guaranteed to find the global optimum. With a finite n_max, the algorithm guarantees that the final labeling is optimal within Hamming distance n_max (i.e., no better labeling exists that differs in ≤ n_max variables).

The Lazy Flipper generalizes Iterated Conditional Modes (ICM): when n_max = 1 the algorithm reduces to ICM. It also extends Block‑ICM, which is limited to regular grid neighborhoods, to arbitrary irregular and higher‑order factor graphs. Compared with randomized subgraph sampling methods (e.g., Jung et al., 2009) and Swendsen‑Wang/Wolff style cluster updates, Lazy Flipper’s deterministic enumeration eliminates the massive redundancy that plagues random approaches (e.g., a 6‑node connected set on a 2‑D grid has 720 permutations, but only 40 unique connected subsets). The trade‑off is higher memory consumption: storing one representative per connected subgraph can require several gigabytes (up to ~8 GB in the authors’ experiments), which is acceptable on modern servers.

Experimental evaluation covers synthetic grid models, irregular graphs, submodular and non‑submodular higher‑order potentials, and real‑world problems such as image segmentation and Bayesian network inference. The authors compare Lazy Flipper (with various n_max values) against state‑of‑the‑art inference methods: standard Belief Propagation (BP), Tree‑Reweighted BP (TRBP), Dual Decomposition with sub‑gradient descent, and ICM. Results show that even with modest depths (n_max = 3–5) Lazy Flipper consistently attains lower energy than BP/TRBP, especially on non‑submodular high‑order models where message‑passing struggles. Runtime scales approximately linearly with the number of variables for a fixed n_max, and the algorithm remains exponential only in the worst‑case depth, offering a practical balance between solution quality and computational effort.

In summary, the Lazy Flipper provides a principled, deterministic framework for MAP inference that leverages graph topology to prune the search space, offers provable guarantees (global optimality at full depth, Hamming‑distance optimality at limited depth), and empirically outperforms or matches leading approximate inference techniques on a broad set of challenging problems. Future work suggested includes extending the method to multi‑label variables, adaptive depth strategies, and parallel/GPU implementations to further accelerate large‑scale applications.