BLADE: ReliaBle Dynamic Hardware-Aware SNN-ANN Boundary SeLection for Event-BAseD Object DEtection

Mahdi Taheri; Alwin Paul

Source-linked AI summary

BLADE: ReliaBle Dynamic Hardware-Aware SNN-ANN Boundary SeLection for Event-BAseD Object DEtection

Mahdi Taheri, Alwin Paul

arXiv:2609.17562v1Jul 22, 2026cs.NEcs.AIcs.ARcs.DCcs.ETcs.LG

TL;DR

Existing hybrid SNN–ANN detectors use static inference and select boundaries mainly by accuracy and energy, leaving dynamic inference and reliability insufficiently integrated. BLADE jointly optimizes the boundary and ANN early-exit configuration using reliability, accuracy, execution time, and energy, with hierarchical fault injection. It achieves competitive detection and energy results while identifying and mitigating a dominant catastrophic fault source.

Problem
Existing hybrid SNN–ANN networks use static inference and largely ignore reliability when selecting boundaries, despite deployment constraints on accuracy, energy, execution time, and dependable operation.
Method
BLADE jointly optimizes the SNN–ANN boundary and ANN early-exit configuration across reliability, detection accuracy, execution time, and energy using hierarchical statistical fault injection.
Results
0.691 mAP@0.5 is achieved with 15.82 mJ inference compute energy when the ANN early exit fires; protecting the dominant exponent bit eliminates catastrophic failures across evaluated realistic fault rates.
Takeaways & Limitations
Reliability-aware boundary selection supports more dependable dynamic hybrid SNN–ANN deployments while retaining competitive detection accuracy and energy efficiency.

Abstract

from arXiv · show

Hybrid Spiking Neural Network (SNN)-Artificial Neural Network (ANN) architectures combine the energy efficiency of SNNs with the superior detection accuracy of ANNs for event-based object detection. Existing hybrid SNN--ANN networks, however, employ static inference and select the SNN-ANN boundary primarily according to accuracy and energy consumption, without considering dynamic inference or reliability. This paper presents BLADE, the first reliability-aware boundary selection methodology for dynamic hybrid SNN-ANN networks with ANN early exit. The proposed framework jointly optimizes the SNN-ANN boundary and ANN early-exit configuration according to reliability, detection accuracy, execution time, and energy consumption, while incorporating reliability through hierarchical statistical fault injection during design-space exploration. Experimental evaluation on an event-based object detector achieves an mAP 0.5 of 0.691 while reducing the inference compute energy to 15.82~mJ when the ANN early exit fires. Reliability analysis identifies the most significant floating-point exponent bit as the dominant source of catastrophic failures, producing significant-or-worse accuracy degradation in 58.8% of its fault injections. Protecting this single bit with approximately 3% storage overhead eliminates catastrophic failures across the evaluated realistic technology fault rates. Furthermore, increasing the proportion of SNN computation improves fault tolerance, with the fully SNN configuration achieving a reliability retention of 0.965 under aggressive fault conditions. The results demonstrate that jointly optimizing reliability, accuracy, execution time, and energy consumption enables more dependable deployment of dynamic hybrid SNN--ANN systems for safety-critical edge AI applications.

I. INTRODUCTION

BLADE addresses the limitations of static hybrid SNN–ANN detection by jointly selecting the SNN–ANN boundary and ANN early-exit configuration with reliability, accuracy, execution time, and energy in view.

I. INTRODUCTION: The framework targets dependable edge-AI deployment under strict execution-time, energy, and reliability constraints, especially for safety-critical object detection.The paper validates the approach on an event-based hybrid object detector and reports competitive accuracy and energy efficiency.
I. INTRODUCTION: BLADE jointly optimizes the SNN–ANN boundary and ANN early-exit configuration for reliability, accuracy, execution time, and energy.Reliability is incorporated through hierarchical statistical fault injection during design-space exploration.
I. INTRODUCTION: Dynamic hybrid detection combines SNN energy efficiency with ANN detection accuracy and adapts executed computation to input complexity through early exit.The architecture uses SNN-based early feature extraction and ANN-based detection stages.
I. INTRODUCTION: Existing hybrid methods mainly optimize accuracy and energy with static inference, while dynamic early exit and reliability-aware boundary selection remain largely unexplored.Prior boundary methods may be manual, limited, or retraining-intensive, and hybrid networks typically execute fixed depth for every input.

III. PROPOSED METHODOLOGY

The BLADE methodology enumerates contiguous SNN–ANN boundaries and ANN early-exit configurations, evaluates each candidate across four objectives, and selects deployment solutions through multi-objective optimization.

A. Hybrid SNN–ANN Search Space: BLADE evaluates boundary and early-exit candidates using accuracy, energy consumption, execution time, and reliability before selecting deployment configurations.The workflow constructs valid combinations from a trained detector and applies multi-objective optimization to the measured candidates.
A. Hybrid SNN–ANN Search Space: The boundary K assigns the first K backbone stages to SNN processing and the remaining stages to ANN processing in a contiguous partition.The SNN processes event streams, while the ANN refines features and performs object detection after spiking-to-dense conversion.
A. Hybrid SNN–ANN Search Space: The compact search space includes both substrate endpoints and intermediate hybrid realizations to expose boundary-dependent accuracy, energy, and reliability trade-offs.Enumerating every integer boundary would multiply training and fault-injection characterization cost without changing observed monotonic trends.
A. Hybrid SNN–ANN Search Space: Fault analysis distinguishes fp32 analog and int8 spiking substrates, identifying the fp32 exponent MSB as the protectable catastrophic site.The architecture places fault-injection sites across the heterogeneous components, and selective protection is integrated into candidate evaluation.
A. Hybrid SNN–ANN Search Space: Dynamic execution occurs only after the SNN–ANN boundary: the complete SNN backbone runs for every input, while ANN confidence determines whether inference exits early or continues.The design space is parameterized by boundary K and selected ANN exit location E.

B. Candidate Evaluation

BLADE evaluates every hybrid SNN–ANN candidate using energy, execution time, detection accuracy, and reliability, with hardware-dependent modeling of SNN and ANN computation. Compute-only estimates support relative candidate ranking across heterogeneous accelerator configurations.

B. Candidate Evaluation: Each candidate is evaluated on energy consumption, execution time, detection accuracy, and reliability before final optimization.Accuracy is measured on validation data, while energy and execution time are modeled for the candidate configuration.
B. Candidate Evaluation: SNN energy uses spike activity, timesteps, MAC counts, and neuron updates, whereas ANN energy uses MAC counts and per-operation costs.The model assigns cAC and cn to neuromorphic computation and cMAC to conventional accelerator computation.
B. Candidate Evaluation: The ASAB bridge and detection head execute on the ANN processor, with the detection head contributing constant energy independent of the selected boundary.The selected ANN exit adds an exit-dependent computational cost.
B. Candidate Evaluation: Compute-only energy excludes memory accesses, static power, and communication overhead, so reported values compare candidates rather than represent absolute system power.The omitted components are assumed identical across candidate configurations.
B. Candidate Evaluation: A hardware-dependent crossover condition determines whether each layer is more efficiently implemented as an SNN or ANN layer.The energy constants instantiate one representative heterogeneous accelerator pairing, while the selection procedure is intended to remain unchanged for alternative pairings.

2) ANN Early-Exit Model:

BLADE introduces dynamic execution only after the SNN–ANN boundary by inserting ANN early exits, while always executing the complete SNN backbone. Exit probabilities determine expected energy and latency, which vary jointly with the boundary and exit location.

2) ANN Early-Exit Model:: The complete SNN backbone runs for every input, while ANN early exits allow intermediate predictions to terminate inference before the final detection head.Dynamic execution begins only after the selected SNN–ANN boundary.
2) ANN Early-Exit Model:: Exit confidence exceeding a predefined threshold terminates inference; otherwise, the sample continues through deeper ANN stages or the final detector.The fixed-threshold rule defines the baseline operating point.
2) ANN Early-Exit Model:: Expected energy and execution time are computed from the probabilities and cumulative costs of terminating at each candidate exit.Each exit cost includes the SNN backbone, ANN layers through that exit, and its detection head; latency is represented by Te(K).
2) ANN Early-Exit Model:: Because the SNN computation is fixed, boundary and exit choices jointly control the amount of ANN computation and therefore candidate energy and execution time.Different exit locations modify only post-boundary ANN computation.
2) ANN Early-Exit Model:: Candidate detection accuracy is measured directly on the validation dataset and combined with energy, execution time, and reliability during optimization.

3) Reliability Characterization:

BLADE characterizes candidate reliability through hierarchical statistical fault injection across numerical representations, bit positions, layers, substrates, and complete hybrid configurations. It combines isolated-bit and BER-based faults to estimate accuracy retention under controlled and realistic conditions.

3) Reliability Characterization:: Fault injection evaluates native SNN signed 8-bit integer parameters and neuron states alongside ANN single-precision floating-point parameters without retraining.This separates reliability contributions from the two computational domains.
3) Reliability Characterization:: Single-bit injections isolate vulnerable bit positions, layers, and representations, while BER-based injections model simultaneous faults under technology-dependent soft-error rates.
3) Reliability Characterization:: The candidate reliability score is mean accuracy retention across the evaluated BER ladder relative to fault-free accuracy.
3) Reliability Characterization:: Reliability is estimated by statistical sampling with Wilson confidence intervals and iterative convergence to maintain manageable simulation time and statistical confidence.
3) Reliability Characterization:: Every candidate’s reliability reflects both its SNN–ANN boundary vulnerability and the ANN computation executed before its selected early exit.The hierarchical campaign spans numerical representation, bit position, network layer, SNN and ANN components, and complete configurations.

4) Bit-Level Protection:

BLADE localizes dominant bit vulnerabilities, evaluates selective protection, and selects a feasible candidate through weighted multi-objective optimization. Feasibility requires accuracy, reliability, and catastrophe-probability constraints before normalized objectives are combined.

4) Bit-Level Protection:: Catastrophic failures concentrate in one ANN floating-point exponent bit, while other floating-point bits and SNN integer representations are substantially less vulnerable.Selective protection of this exponent bit improves reliability with negligible storage overhead.
4) Bit-Level Protection:: All candidates are reevaluated after excluding the protected bit from fault injection, isolating boundary and early-exit effects from avoidable catastrophic failures.The protected reliability score then enters optimization with measured accuracy and modeled energy and execution time.
4) Bit-Level Protection:: The selected configuration maximizes a weighted combination of normalized accuracy, reliability, energy, and execution-time metrics over the feasible set.Weights can prioritize reliability, energy, latency, or accuracy according to deployment requirements.
4) Bit-Level Protection:: Algorithm 1 enumerates boundary, exit, and protection choices, characterizes each candidate, applies Pareto filtering and feasibility constraints, and computes the final objective.
4) Bit-Level Protection:: BLADE forms a Pareto set over reliability, accuracy, energy, and execution time, then retains candidates meeting all feasibility thresholds.The catastrophe probability is bounded by the Wilson 95% upper limit for faulted accuracy below mAP < 0.10, with Pmax = 0.05.

1) Optimization Procedure:

BLADE evaluates all valid boundary and ANN early-exit configurations without repeated retraining, then filters and selects candidates using measured multi-objective characteristics.

The workflow generates valid boundary and early-exit configurations, evaluates their energy, execution time, accuracy, and reliability, and retains non-dominated candidates through Pareto filtering.The supplied workflow passage describes candidate generation, evaluation, and Pareto retention before constraint-based selection.
BLADE evaluates candidate configurations within one framework without repeated retraining for each candidate.The candidates are assessed using design-space characteristics rather than retraining-intensive exhaustive exploration.

D. Deployment Configuration

Deployment instantiates the selected boundary and ANN early-exit configuration, applies selective protection, and represents a configurable compromise among accuracy, energy, execution time, and reliability.

The selected boundary K* and early-exit configuration E* completely determine the computational graph executed during inference.The first K* backbone stages run on the SNN accelerator, while the remaining stages run on the ANN accelerator.
The learned controller decides whether to terminate at an intermediate ANN exit or continue to the final detection head, while the SNN backbone always executes completely.Dynamic execution therefore affects only the ANN component.
Selective bit-level protection is applied to the ANN component after configuration selection without changing network architecture or requiring retraining.The protection preserves the original parameters while improving robustness against transient hardware faults.
The deployed configuration optimizes a compromise among detection accuracy, energy consumption, execution time, and reliability, with objective weights adaptable to deployment priorities.The methodology is independent of the underlying hardware platform.
The evaluation covers the dynamic architecture, detector reliability, cross-boundary reliability, and final reliability-aware boundary and early-exit selection.

A. Experimental Setup

Experiments evaluate dynamic early exit and reliability-aware design on Gen1 event-camera detection using fixed temporal preprocessing, validation-selected operating points, and held-out testing.

The Gen1 experiments use 10 temporal bins at 240×304 and report detection accuracy as mAP@0.5.The dataset contains cars and pedestrians and was recorded by a moving vehicle.
Operating points are selected on 20 329 validation images and re-evaluated on 30 605 held-out test images, with K = 4 deployed and K ∈ {0, 2, 4, 8} assessed for cross-boundary reliability.Both floating-point and integer implementations are evaluated.
At τ = 0.90, the early exit achieves 0.6695 mAP@0.5 and 16.22 mJ, reducing energy by 7.4% with a 1.4% relative mAP drop and exiting on 19.5% of validation samples.The static K = 4 detector achieves 0.6789 mAP@0.5 at 17.53 mJ, while always exiting reaches 10.85 mJ and 0.5919 mAP@0.5.
The dynamic exit saves energy more than latency because the SNN backbone accounts for about 96% of GPU inference latency, limiting latency reduction to about 3.1%.Only the final ANN detection head is skipped, structurally capping dynamic energy savings at 38.1%.
The learned multi-dimensional controller outperforms the confidence threshold at most non-trivial exit fractions, with a peak gain of +0.0326±0.0062 mAP@0.5 at 80% exit.The paired-t gain interval excludes zero at 10 of 12 non-trivial exit fractions; at 8% exit, the difference is −0.0020.
The controller improvement is mainly attributable to multi-dimensional detection-state features: a non-reinforcement linear gate recovers approximately 86% of the best learned-controller gain, while scalar-confidence correlation with true exit loss is −0.108.

2) Exit Placement:

Exit placement and reliability analysis identify a single stride-16 ANN exit and a dominant floating-point exponent vulnerability, motivating targeted protection in the hybrid detector.

Exit Placement: The stride-16 exit is selected because it reaches 0.6719 mAP@0.5 at a 9.7% exit rate, exceeding stride-32 and stride-8 under the deployment-relevant near-lossless criterion.Always-exit accuracy is also highest at stride-16, with 0.5781 mAP@0.5 versus 0.4510 at stride-32 and 0.4080 at stride-8.
Number of ANN Exits: The temporal exit never fires because the ASAB bridge requires all T = 10 temporal bins before producing the fused ANN representation.Across the evaluated grid, the temporal exit rate is 0.000 and the average exit timestep is 10 out of 10.
Number of ANN Exits: The dual-exit cascade collapses to the single stride-16 exit because the preceding stride-8 exit fires on 0% of frames at the accuracy-matched operating point.The dual-exit configuration reaches 0.6722 mAP@0.5 at about a 17% overall exit rate and shows no catastrophic failures in the realistic fault region.
Number of ANN Exits: Adding a second exit raises router overhead from 0.25 mJ to about 1.25 mJ, a 5× increase without accuracy or reliability benefit.The final architecture therefore uses one ANN early exit at stride-16.
Bit-level Vulnerability Analysis: Bit 30, the most significant floating-point exponent bit, causes significant-or-worse accuracy degradation in 58.8% of injections, while the other 31 bits cause only one significant degradation over 3155 injections.The dominant vulnerability is localized to the ANN floating-point representation.
Bit-level Vulnerability Analysis: Flipping exponent bit 30 rescales affected weights by approximately 2^128 or 2^-128, whereas the int8 SNN backbone degrades gracefully without catastrophic failures over the evaluated fault range.Protecting the known vulnerable bit requires approximately 3% storage overhead.

2) Layer-wise Criticality:

Bit-30 faults are concentrated near the SNN–ANN boundary, but selective protection removes catastrophic failures across realistic fault rates. Increasing the SNN share improves robustness under aggressive stress while preserving a boundary-dependent accuracy trade-off.

Layer-wise Criticality: The ASAB bridge has catastrophic fraction 1.0 under bit-30 faults, followed by the first ANN block at 0.76, while later blocks become less vulnerable.The detection head records catastrophic fraction zero despite containing the most floating-point parameters.
Multi-bit Reliability: Protecting bit 30 removes catastrophic failures across realistic BER values, with mean mAP@0.5 declining from 0.692 clean to 0.686 at 10^-5 and 0.677 at 10^-4.At higher stress, mean mAP@0.5 falls to 0.572 at 10^-2 and 0.194 at 5×10^-2.
Multi-bit Reliability: The int8 SNN degrades earlier but more gradually, whereas the protected fp32 ANN retains higher moderate-BER accuracy before dropping sharply at extreme rates.Neither substrate produces catastrophic trials across the realistic BER range of 5.59 × 10^-9 to 2.04 × 10^-5.
Validation: The protected reliability trend remains unchanged across independent subsets, 100 versus 1000 Monte Carlo trials, and a larger 2048-image evaluation subset.Increasing the Monte Carlo count tightens the Wilson confidence interval without changing reliability behavior.
Cross-boundary Reliability: Across protected boundaries, all configurations degrade gracefully without catastrophic failures in the realistic technology range, while greater SNN computation improves accuracy retention under aggressive faults.The fully SNN K = 8 configuration retains the highest stress-range accuracy, whereas K = 4 has the highest clean detection accuracy.

D. Reliability-aware Boundary Selection

BLADE jointly selects the SNN–ANN boundary and ANN early-exit configuration across reliability, accuracy, execution time, and energy. The preferred deployment changes with the target objective balance, with selective bit protection preserving realistic-range reliability.

D. Reliability-aware Boundary Selection: BLADE jointly optimizes the SNN–ANN boundary and ANN early-exit configuration using reliability, detection accuracy, execution time, and energy consumption.Reliability is evaluated through retained detection accuracy over the complete BER sweep.
D. Reliability-aware Boundary Selection: BLADE selects different feasible deployments according to objective priorities: K = 8 with stride-16 exit gives reliability retention 0.965 and 10.20 mJ, while K = 4 gives mAP@0.5 0.691 and 15.82 mJ.The energy figure is reported on the compute-only basis used in Table VIII.
D. Reliability-aware Boundary Selection: No single boundary optimizes all objectives simultaneously, so deployment depends on the target balance among reliability, accuracy, execution time, and energy consumption.The K = 4 configuration remains free of catastrophic failures throughout the realistic technology range after bit-level protection.
D. Reliability-aware Boundary Selection: Bit-30 corruption causes significant-or-worse degradation in 58.8% of injections, while approximately 3% storage overhead for protection eliminates catastrophic failures at realistic fault rates.The fully SNN configuration achieves the highest reliability retention under aggressive faults.
Scope: The methodology requires candidate boundaries, ANN early exits, and per-candidate accuracy, energy, execution-time, and reliability metrics, supporting application across hybrid SNN–ANN architectures.Future work targets additional event-based tasks, activation and permanent faults, and validation on emerging neuromorphic hardware.