Every Problem, Every Step, All in Focus: Learning to Solve Vision-Language Problems With Integrated Attention

Chen, Xianyu; Yang, Jinhui; Chen, Shi; Wang, Louis; Jiang, Ming; Zhao, Qi

doi:10.1109/TPAMI.2024.3357631

Abstract An ordering constraint satisfaction problem (OCSP) is defined by a family$$\mathcal F$$

F

of predicates mapping permutations on$$\{1,\ldots,k\}$$

{1, \dots, k}

to$$\{0,1\}$$

{0, 1}

. An instance of ($$\mathcal F$$

F

) onnvariables consists of a list of constraints, each consisting of a predicate from$$\mathcal F$$

F

applied onkdistinct variables. The goal is to find an ordering of thenvariables that maximizes the number of constraints for which the induced ordering on thekvariables satisfies the predicate. OCSPs capture well-studied problems including ‘maximum acyclic subgraph’ () and “maximum betweenness”. In this work, we consider the task of approximating the maximum number of satisfiable constraints in the (single-pass) streaming setting, when an instance is presented as a stream of constraints. We show that for every$$\mathcal F$$

F

, ($$\mathcal F$$

F

) is approximation-resistant to o(n)-space streaming algorithms, i.e., algorithms using o(n) space cannot distinguish streams where almost every constraint is satisfiable from streams where no ordering beats the random ordering by a noticeable amount. This space bound is tight up to polylogarithmic factors. In the case of , our result shows that for every$$\epsilon>0$$

ϵ > 0

, is not$$(1/2+\epsilon)$$

(1 / 2 + ϵ)

-approximable in o(n) space. The previous best inapproximability result, due to Guruswami & Tao (2019), only ruled out 3/4-approximations in$$o(\sqrt n)$$

o (\sqrt{n})

space. Our results build on recent works of Chou et al. (2022b, 2024) who provide a tight, linear-space inapproximability theorem for a broad class of “standard” (i.e., non-ordering) constraint satisfaction problems (CSPs) over arbitrary (finite) alphabets. Our results are obtained by building a family of appropriate standard CSPs (one for every alphabet sizeq) from any given OCSP and applying their theorem to this family of CSPs. To convert the resulting hardness results for standard CSPs back to our OCSP, we show that the hard instances from this earlier theorem have the following “partition expansion” property with high probability: For every partition of thenvariables into small blocks, for most of the constraints, all variables are in distinct blocks.

More Like this