AP(TM): Associative processor based Tsetlin machine

The Tsetlin machine (TM) is a logic-based learning model that performs classification using conjunctive clauses over propositional literals, offering interpretable decision logic and efficient bit-level computation. Associative Processor (AP) is a massively parallel processing in memory (PiM) architecture that executes data-parallel Boolean operations directly within associative memory arrays. This paper presents AP(TM), an AP -based implementation of the TM that supports in-memory training and inference. AP(TM)maps clauses to AP rows and literals to column fragments, enabling parallel clause evaluation, class vote accumulation, and automata updates using standard AP primitives such as masked compare (lookup), parallel write, and reduction (popcount). To efficiently realize the stochastic learning rules of the TM , AP(TM)introduces AP -native mechanisms for probabilistic clause selection and per-literal update gating based on distributed stochastic masks. The proposed mapping enables constant-latency training and inference that are independent of the number of clauses. Functional evaluation on MNIST, Fashion-MNIST, IMDb and UCI Adult datasets demonstrates that the proposed AP -based implementation closely matches the learning behavior and accuracy of a software baseline (within 0.32 pp). This work establishes associative in-memory processing as a viable and scalable platform for TM training and inference, independent of the underlying memory technology.