Notes

• Proposes a generic zkSNARK framework modeling neural networks as matrix-computation DAGs.
• Introduces a dual-layer design: LiteBullet ensures linear proving, PoP compresses proofs and hides structure.
• LiteBullet uses folding-based inner product arguments, avoiding polynomial expansion and sumcheck.
• The Evalyn interface unifies diverse matrix and nonlinear operations, supporting heterogeneous NNs.
• Batching and lookup tables improve efficiency for nonlinear functions and matrix commitments.
• Achieves O(Mn²) prover and O(log(Mn)) proof/verification complexity, with an open-source implementation.

• 提出通用 zkSNARK 框架，将神经网络建模为矩阵计算 DAG。
• 设计双层结构：LiteBullet 保持线性 prover，PoP 压缩证明并隐藏架构。
• LiteBullet 基于 folding 内积证明，避免多项式扩展与 sumcheck。
• Evalyn 接口统一不同矩阵与非线性操作，支持异构 NN。
• 批处理与查表提升非线性函数和矩阵承诺效率。
• 实现 O(Mn²) prover 与 O(log(Mn)) proof/验证复杂度，并开源实现。