A Unified Memory Perspective for Probabilistic Trustworthy AI

Memory, not computation, is the bottleneck for trustworthy AI: The paper identifies that probabilistic approaches to achieving robustness, interpretability, security and privacy in AI systems shift performance constraints from arithmetic units to memory systems, meaning faster processors alone won't solve trustworthy AI challenges.

Stochastic sampling demands reduce efficiency: Increasing reliance on random sampling—a key technique for trustworthy AI—reduces the effective efficiency of data access in memory systems, potentially pushing hardware into "entropy-limited operation" where the system's ability to generate randomness becomes the limiting factor rather than data retrieval speed.

Current hardware architectures weren't designed for this: Conventional computer architectures have fundamental limitations when handling the interleaved deterministic and stochastic (random) operations required by trustworthy AI systems, suggesting that existing hardware may not be optimal for deploying these safety-critical approaches at scale.

New hardware designs are needed: The paper advocates for "probabilistic compute-in-memory" approaches that integrate random sampling directly with memory access, rather than treating them as separate functions, as a pathway toward building hardware specifically optimized for trustworthy AI systems.

A unified framework enables better evaluation: By treating deterministic data access as a special case of stochastic sampling, the paper provides a common analytical framework for evaluating hardware designs on criteria like robustness to hardware imperfections and parallel processing capability.