SYSTEM AND METHOD FOR SHARING AND OVERCOMING CONTROL POINTS
Patent Information
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- NAT UNIV OF IRELAND MAYNOOTH
- Filing Date
- 2017-09-13
- Publication Date
- 2026-06-17
AI Technical Summary
Existing reverse-mode automatic differentiation (AD) methods face high memory burden and inefficiency when applied to complex control architectures, as they require storing intermediate values during computation, which becomes infeasible for long-running computations.
A system and method for divide-and-conquer checkpointing that transforms programs into a formalism allowing convenient manipulation and introduces operators to decompose computations, enabling efficient storage reduction by pausing and resuming computations at arbitrary points, supporting binary and n-ary checkpoint trees with various termination criteria.
Reduces storage requirements and computation time for reverse-mode AD by allowing decomposition of complex programs into manageable segments, achieving sublinear space and time complexity without modifying the program structure.