A framework for designing power-efficient inference accelerators in tree-based learning applications. (March 2022)

• Record Type: Journal Article
• Authors: Abreu, Brunno; Grellert, Mateus; Bampi, Sergio
• Is Part Of: Engineering applications of artificial intelligence. Volume 109(2022)
• Publication Date: 2022-03
• Subjects: Machine Learning -- Decision trees -- Random forests -- Inference -- VLSI circuits -- Power–accuracy trade-off; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2021.104638
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: Machine Learning techniques (ML) are being widely adopted in embedded devices due to their efficiency and flexibility. However, the strict power limitations in such devices, combined with the variable resource requirements of ML models, require further understanding of how model complexity affects power and performance. This paper proposes a framework that facilitates the design space exploration of dedicated decision trees (DT) and random forests (RF) accelerators by enabling a joint assessment of power dissipation and prediction accuracy. The proposed framework translates tree-based structures to hardware description languages (HDL). The HDL modules are submitted to logic and physically-aware hardware synthesis flows, allowing a detailed power-performance analysis of VLSI DTs and RFs. Using four data sets of embedded applications as case studies, we found that quantizing the input features leads to accuracy gains of up to 6.3% compared with the precise versions. We also show that using shallower trees may lead to small prediction loss with significant reductions in power, which is favorable for power-constrained applications. Our translator achieves better results in terms of energy/inference w.r.t. prior related works under comparison, one of which employed standard methods for hardware translation such as High-Level Synthesis. The proposed solution presents a power reduction of 10 times or more for the same inference throughput reported in prior works.