Back to search

National Technical University of Athens (NTUA)

Design automation of approximate machine learning circuits for ultra resource constrained environments

Abstract

dc:description

The continuous proliferation of connected devices and embedded intelligence has shifted computing from centralized data centers toward distributed, edge-centric processing. Edge computing reduces latency and preserves privacy by bringing analytics closer to the data source. In this spectrum, many IoT devices, microcontrollers, sensors, and other low-power nodes, operate under tight energy and cost constraints, giving rise to TinyML and extreme-edge computing. At the farthest end lie ultra-constrained devices that can sometimes run even without batteries, relying solely on ambient energy harvesting, and requiring extremely low manufacturing cost per unit. Printed and Flexible Electronics (PFEs) have emerged as a promising technology for such applications. By leveraging additive manufacturing, PFEs can produce ultra-low-cost, mechanically flexible, conformal, and even disposable circuits. They can be printed on substrates like paper, textiles, or polymers using non-toxic materials, enabling form factors that rigid silicon chips cannot achieve. These properties make PFEs attractive for smart packaging, wearable health patches, and environmental sensors, domains where conventional silicon devices struggle due to rigidity or cost constraints. However, PFEs face severe technological challenges that hinder their use in intelligent systems. Printed transistors have large feature sizes and high variability, resulting in very low integration density and slow switching speeds. Their immature fabrication processes also lead to high static power consumption and limited reliability. In particular, printed devices struggle to incorporate dense memory or on-chip storage. Even very simple digital circuits in printed form typically require extremely high area and power. As a result, existing printed ML accelerators are restricted to only the most basic classifiers, highlighting a significant gap between the potential of pervasive intelligence and the capabilities of current hardware. Early state-of-the-art efforts have attempted to bridge this gap through bespoke architectures, fully customized printed ML circuits where parameters are hardwired to minimize overhead. While these designs significantly improve efficiency, they remain limited by their rigidity and by the absence of systematic optimization methodologies. Moreover, the hardware overhead of peripheral components such as analog-to-digital converters (ADCs) becomes dominant as the computational core shrinks, revealing a new frontier in printed ML co-design. This dissertation addresses these challenges by applying approximate-computing (AC) principles across the hardware–software stack to make ML inference feasible under extreme constraints. We introduce a cross-layer co-design methodology that minimizes hardware requirements while preserving accuracy. Post-training approximations reduce model complexity through coefficient approximation, netlist pruning, and voltage scaling, achieving up to six-fold area and power reductions versus exact designs. Building on this, a hardware-aware training framework guides model retraining toward weights that are friendly to printed hardware, yielding battery-operable printed neural networks. We also propose on-sensor solutions, including ADC-aware decision trees and co-designed converters, that enable self-powered ML inference in printed systems. Beyond model-specific circuits, the dissertation extends the AC methodology to programmable domain-specific architectures. A mixed-precision RISC-V accelerator is proposed, featuring lightweight instruction-set extensions and multi-pumped SIMD units tailored for deep-learning inference at the extreme edge. This design achieves significant energy savings without sacrificing programmability, demonstrating how AC principles can scale from bespoke printed circuits to general-purpose embedded processors. Collectively, these contributions show that by orchestrating approximations at the algorithmic, architectural, and circuit levels, we can bridge the gap between ML workloads and ultra-low-cost hardware. In our printed implementations, the proposed methods reduce area and power by more than an order of magnitude while keeping accuracy loss below 1%. This paves the way for sustainable, on-device intelligence in disposable or biocompatible systems powered by energy harvesting.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Armeniakos, Georgios
  • Αρμενιάκος, Γεώργιος

Subjects

dc:subject × 22

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62288

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Armeniakos, Georgios; Αρμενιάκος, Γεώργιος. Design automation of approximate machine learning circuits for ultra resource constrained environments. National Technical University of Athens (NTUA), 2026. http://hdl.handle.net/10442/hedi/62288