2020/07/04 by K. Darshana Abeyrathna, Abeyrathna, K. Darshana, Ole‐Christoffer Granmo +11
Computer Science · Engineering · #Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Ferroelectric and Negative Capacitance Devices #Machine Learning (cs.LG) #Optimization and Search Problems #Scheduling and Optimization Algorithms
paper · pdf · doi:10.48550/arxiv.2007.02114
openalex publication_date 2020/07/04 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Due to the high energy consumption and scalability challenges of deep\nlearning, there is a critical need to shift research focus towards dealing with\nenergy consumption constraints. Tsetlin Machines (TMs) are a recent approach to\nmachine learning that has demonstrated significantly reduced energy usage\ncompared to neural networks alike, while performing competitively accuracy-wise\non several benchmarks. However, TMs rely heavily on energy-costly random number\ngeneration to stochastically guide a team of Tsetlin Automata to a Nash\nEquilibrium of the TM game. In this paper, we propose a novel finite-state\nlearning automaton that can replace the Tsetlin Automata in TM learning, for\nincreased determinism. The new automaton uses multi-step deterministic state\njumps to reinforce sub-patterns. Simultaneously, flipping a coin to skip every\nd'th state update ensures diversification by randomization. The d-parameter\nthus allows the degree of randomization to be finely controlled. E.g., d=1\nmakes every update random and d=\∞ makes the automaton completely\ndeterministic. Our empirical results show that, overall, only substantial\ndegrees of determinism reduces accuracy. Energy-wise, random number generation\nconstitutes switching energy consumption of the TM, saving up to 11 mW power\nfor larger datasets with high d values. We can thus use the new d-parameter\nto trade off accuracy against energy consumption, to facilitate low-energy\nmachine learning.\n