Printed Ternary Neural Networks

• Printed Ternary Neural Networks are neural networks with discrete {-1, 0, +1} weights designed for printed electronics, balancing efficiency and accuracy.
• They employ evolutionary circuit optimization and approximate arithmetic to achieve significant reductions in area and power consumption.
• Applications include flexible sensors, wearables, and environmental monitors where ultra-low power, cost-effectiveness, and robustness are essential.

Printed ternary neural networks (TNNs) are neural network architectures in which synaptic weights, and often activations, are restricted to the discrete set $\{-1, 0, +1\}$ and are physically realized using printed electronics. These systems are designed to maximize area and power efficiency—crucial for flexible, disposable, or ultra-low-cost hardware—by exploiting the simplicity of discrete arithmetic and new opportunities for extreme hardware specialization. Recent developments encompass a holistic design approach from the analog-to-digital interface to the bespoke classifier, co-optimized for arbitrary input precision, and typically rely on evolutionary circuit optimization to bridge the gap between theoretical accuracy and stringent hardware constraints. Printed TNNs now form a distinct and practically deployable class of machine learning hardware for resource-constrained applications.

1. Foundations and Motivations

Printed electronics enable circuit fabrication using additive, mask-less techniques (e.g., inkjet printing), which support flexible, stretchable, and conformal substrates at ultra-low manufacturing cost. However, these processes introduce severe constraints: large device feature sizes, limited integration density, and high variability, resulting in area- and power-intensive implementations for conventional digital circuits. TNNs, with weights in $\{-1, 0, +1\}$, match this resource profile. They entirely remove the need for expensive multipliers, as operations reduce to sign inversion, addition, subtraction, or simple wiring changes, and the ternary nature allows dense “zeroing,” enhancing both power savings and functional sparsity (Mrazek et al., 30 Jul 2024, Mrazek et al., 27 Aug 2025).

Compared to binary neural networks, which only allow two discrete levels, ternary networks offer improved expressivity. The "zero" weight increases model flexibility and enables structured sparsity, yielding significant area and power reductions versus both binary and full-precision networks when mapped onto printed circuits. These characteristics make printed TNNs especially attractive for power-constrained, disposable, or on-sensor processing scenarios—such as environmental monitors, wearables, or smart packaging—where silicon VLSI is either infeasible or prohibitively expensive.

2. Printed Hardware Architecture and System-Level Co-Design

The complete processing chain in a printed TNN covers three primary domains: sensor analog interface, digital classifier hardware, and multi-objective circuit co-optimization.

Sensor Interface: Analog-to-Digital Conversion

In printed electronics, the analog-to-digital converter (ADC) or analog-to-binary converter (ABC) is typically an area and power bottleneck. Rather than deploying high-cost flash or SAR ADCs, printed TNNs leverage circuit-simplified ABCs, consisting of only two resistors and one comparator per input. Thresholds can be tuned via the resistor ratio, yielding highly compact interfaces (e.g., $0.07\, \text{mm}^2$ area vs. $12\,\text{mm}^2$ for a 4-bit ADC, and $0.03\,\text{mW}$ power vs. $1\,\text{mW}$) (Mrazek et al., 30 Jul 2024, Mrazek et al., 27 Aug 2025). This architectural strategy achieves up to $8\times$–$15\times$ reductions in component count relative to classical approaches.

Ternary Neuron Circuitry: Core Computing Units

The primary arithmetic operation is a thresholded sum:

$$y = \mathrm{sign}\left( \sum_{i=1}^{n} w_i x_i \right)$$

where $w_i \in \{-1, 0, +1\}$ and $x_i$ is the (quantized or binary) input. Hidden-layer neurons typically decompose the sum into two popcounts—one for positive weights, one for negatives—followed by a comparator:

$$\sum_{\forall i: w_i=1} x_i \geq \sum_{\forall i: w_i=-1} x_i$$

The hardware instantiations of these "popcount" and "popcount-compare" (PCC) units are evolved for area and power efficiency using techniques such as Cartesian Genetic Programming (CGP). For output layers, popcount units aggregate neuron activations.

Printed Circuit Evolution and System Integration

The end-to-end design flow involves:

• Phase 1: Evolutionary optimization of approximate elementary circuits (popcounts, PCCs, linear threshold gates (LTGs)), using estimated area and carefully chosen error metrics (mean arithmetic error, worst-case arithmetic error, distance-to-threshold error).
• Phase 2: NSGA-II multi-objective search, which integrates components into a full TNN, optimizing for area and accuracy trade-offs, and directly accounting for analog interface cost. The result is a per-classifier, per-dataset, and per-precision tailored layout, dramatically reducing unnecessary generality and overhead (Mrazek et al., 30 Jul 2024, Mrazek et al., 27 Aug 2025).

3. Component-Level Approximation and Circuit Techniques

In contrast to classical "bitwise" approximation of arithmetic units, printed TNNs employ holistic circuit-level approximation, replacing full adders and comparators with co-evolved (potentially error-tolerant) blocks.

Approximate Ternary Arithmetic

• Approximate popcount circuits are mutated from full implementations, targeting minimized area for a fixed tolerable error—typically evaluated via arithmetic mean error ($\epsilon_\text{mae}$) and worst-case error ($\epsilon_\text{wcae}$).
• PCCs, which combine two popcounts and a comparator, are evaluated using custom error metrics reflecting functional implications of misclassification rather than pure Hamming distance.
• For output neurons, the sign-based aggregation is corrected via right-shifting and constants to maintain unbiased computation given omitted zero-weight connections.

Input Precision Adaptation

Printed TNN frameworks (e.g., (Mrazek et al., 27 Aug 2025)) support arbitrary input quantization (1–4 bits), enabling fine-grained trade-offs between ADC cost, classifier accuracy, and circuit area. Experimental analysis demonstrates that while 1-bit inputs are optimal for resource-constrained datasets, higher precision is crucial to preserve accuracy on more complex tasks (e.g., Pendigits). The design flow automatically selects the most area/accuracy-efficient configuration for each classifier and application.

4. Area, Power, and Performance Characteristics

Empirical evaluation on UCI machine learning datasets consistently demonstrates:

Metric	Classical MLP	Prior Approximate Printed NN	Printed TNNs (Mrazek et al., 27 Aug 2025)
Area reduction	—	Up to $6\times$	Up to $88\times$
Power (with interface)	—	Baseline	$17\times$ (average) to $59\times$ (max) lower
Accuracy loss (typical)	—	2–10%	Under $5\%$ (with Pareto optimal trade-off)

These figures account for total system cost, including the analog interface. Notably, for many datasets (Breast Cancer, Cardio, RedWine), the total classifier circuit—including interface—consumes less than $2\,\text{mW}$, compatible with existing printed energy harvesters. Even for more demanding datasets, energy requirements remain within the range of printed batteries (Mrazek et al., 30 Jul 2024, Mrazek et al., 27 Aug 2025).

5. Error, Robustness, and Dataset Suitability

Printed TNNs exhibit high robustness to approximation-induced errors:

• Holistic circuit approximation, guided by tailored error metrics and BDD-based (binary decision diagram) analysis, ensures a controlled trade-off between hardware savings and accuracy loss.
• Even with severe area optimization, empirical experiments show accuracy declines are contained ($<$5%), and for moderate reductions (41% area, 42% power), accuracy drops are often negligible (Mrazek et al., 30 Jul 2024).
• The "zero" weight increases resilience to both process variations and input noise, as non-critical connections are explicitly removed.

Arbitrary input precision is vital for generalization: in datasets where a 1-bit input completely fails to preserve accuracy (e.g., more complex pattern recognition), the framework's optimizer selects the lowest workable precision to ensure hardware cost is minimized without sacrificing functional utility (Mrazek et al., 27 Aug 2025).

6. Practicality and Applications

• Open-source printed TNN classifiers are available (e.g., Icarus Verilog netlists, synthesis for Cadence and Synopsys tools), supporting transparent benchmarking and practical deployment (Mrazek et al., 30 Jul 2024).
• Pareto-optimal TNNs support area/power scaling for disposable or always-on, on-sensor embedded ML, such as wearables, healthcare sensors, environmental monitors, and smart packaging—all domains where the physical flexibility and low cost of printed electronics are indispensable.
• The methodology allows fully digital TNNs that bypass the analog noise sensitivity and interfacing issues common in analog neuromorphic printed designs, while directly matching the energy budgets of printed energy-harvesters or batteries.

7. Future Directions and Open Challenges

Future work may focus on:

• Extending circuit-level approximation to larger networks (e.g., via modular or hierarchical LTG decomposition) to enable higher-dimensional input handling (Mrazek et al., 27 Aug 2025).
• Incorporating circuit pruning to further reduce input fan-in per neuron, relieving approximation complexity.
• Enhancing process-variation tolerance through variation-aware training, Monte Carlo design, or error-immune circuit architectures.
• Exploring error-resilient training methodologies (“immunizing” the TNN to hardware-induced noise), and studying timing/reliability co-optimization within the printed electronics paradigm.
• Integrating on-chip timing optimization, reliability analysis, and dynamic adaptation to printed process/temperature-induced drift—given the inherent variability of printed FET and analog blocks (Mrazek et al., 27 Aug 2025).
• Expanding battery-powered operation and further reducing hardware cost for autonomous, maintenance-free intelligence in emerging use-cases.

---

Printed ternary neural networks represent the intersection of algorithmic quantization and printed circuit optimization. Leveraging ternary quantization, circuit-level evolutionary approximation, arbitrary front-end adaptation, and full-system co-design, these systems achieve hardware efficiency that meets or exceeds the limitations of printed electronics while providing practical accuracy for embedded ML applications. As area and energy demands are driven ever lower and new fabrication techniques mature, printed TNNs are likely to form the computational backbone of pervasive, flexible, and disposable intelligence.