SemDraft
A framework that learns executable pattern-making programs by searching from paired before-and-after sewing patterns. It recovers edit action sequences through a triple-certification mechanism consisting of a deterministic compiler, action-pair order determination, and independent replay.
Timeline (last 90 days)
The paper "Learning to Search for Executable Patternmaking from Certified Programs" proposes the SemDraft framework, which is reportedly the first system that can learn to search for executable patternmaking programs from paired "before-modification/after-modification" sewing patterns.
SemDraft defines the task as follows: the input is a pair of patterns before and after modification, and the output is a program composed of a sequence of edit actions, rather than directly generating the finished pattern.
SemDraft uses a deterministic compiler to verify the recovered edit action sequence, requiring that the action sequence can compile the first pattern into the second pattern.
SemDraft certifies programs through three independent checks: recompiling from scratch, determining the order of action pairs (ordered or commutative), and independent replay.
Performing a commutativity binary classification on each pair of edit actions in the search space can recover the partial order relations among patternmaking operations and use them to prune the search space.
Under a fixed budget of 1000 search executions, the learned ranker certified 63 and 64 cases respectively on two 80-case test sets that are order-sensitive, while the non-learned ranker could certify at most 53 and 52.
Traditional process mining assumes the existence of timestamped event logs, whereas in SemDraft's scenario there is not even a log—only two static snapshots before and after modification remain.
SemDraft's "search plus certification" paradigm is highly isomorphic to "Programming by Example" in the field of program synthesis.
SemDraft transforms an unsupervised search problem into a supervised learning-to-rank problem, using certified programs as the intermediary.
The SemDraft method currently relies on a deterministic compiler and on case designs that are order-sensitive; in scenarios where compilation behavior is unstable or edit actions are ambiguous, the strictness of the certification mechanism may become the ceiling on coverage.
1 more timeline events
All Facts (11)
The paper "Learning to Search for Executable Patternmaking from Certified Programs" proposes the SemDraft framework, which is reportedly the first system that can learn to search for executable patternmaking programs from paired "before-modification/after-modification" sewing patterns.
50%UnverifiedSemDraft defines the task as follows: the input is a pair of patterns before and after modification, and the output is a program composed of a sequence of edit actions, rather than directly generating the finished pattern.
50%UnverifiedSemDraft uses a deterministic compiler to verify the recovered edit action sequence, requiring that the action sequence can compile the first pattern into the second pattern.
50%UnverifiedSemDraft certifies programs through three independent checks: recompiling from scratch, determining the order of action pairs (ordered or commutative), and independent replay.
50%UnverifiedPerforming a commutativity binary classification on each pair of edit actions in the search space can recover the partial order relations among patternmaking operations and use them to prune the search space.
50%UnverifiedUnder a fixed budget of 1000 search executions, the learned ranker certified 63 and 64 cases respectively on two 80-case test sets that are order-sensitive, while the non-learned ranker could certify at most 53 and 52.
50%UnverifiedTraditional process mining assumes the existence of timestamped event logs, whereas in SemDraft's scenario there is not even a log—only two static snapshots before and after modification remain.
50%UnverifiedSemDraft's "search plus certification" paradigm is highly isomorphic to "Programming by Example" in the field of program synthesis.
50%UnverifiedSemDraft transforms an unsupervised search problem into a supervised learning-to-rank problem, using certified programs as the intermediary.
50%UnverifiedThe SemDraft method currently relies on a deterministic compiler and on case designs that are order-sensitive; in scenarios where compilation behavior is unstable or edit actions are ambiguous, the strictness of the certification mechanism may become the ceiling on coverage.
50%UnverifiedFor the apparel industry, SemDraft's method may allow old pattern archives that have only the before- and after-modification versions and whose intermediate process has long been forgotten to have the operation paths of that time reconstructed.
50%