CBCL-PR-Jetson

A TensorRT FP16 / INT8 reproduction of CBCL-PR (Ayub & Wagner, IEEE TCDS 2023) on Jetson Orin Nano, plus an ablation study quantifying how feature-backbone quantization interacts with class-incremental learning, and a short research note on a hypothesized latency channel for trust collapse in human-robot interaction.

This repository is a faithful reimplementation of the original aliayub7/CBCL-PR codebase, ported to modern PyTorch and extended with TensorRT engine builds and on-device benchmarks.

Status

Phase 1 complete: upstream audit and project scaffolding (see docs/AUDIT.md).

Phases 2 through 7 in progress.

Hardware

Phase	Device
FP32 baseline + ablation training	NVIDIA DGX Spark (GB10 Blackwell, 128 GB unified)
ONNX export	DGX Spark
TRT FP16 / INT8 engine compile	Jetson Orin Nano Super 8GB, JetPack 6.2, TensorRT 10.3.0
Latency benchmarks	Jetson Orin Nano Super 8GB

Layout

upstream/        Untouched clone of github.com/aliayub7/CBCL-PR for reference
src/             Modern reimplementation
scripts/         Feature extraction, ONNX export, engine build, benchmark drivers
configs/         YAML configs per dataset / per quantization mode
calibration/     INT8 calibration data and scripts
benchmarks/      Latency harness output
results/         CSVs and plots checked into git
tests/           Numerical-equivalence and smoke tests
docs/            AUDIT.md, REPRODUCIBILITY.md, RESEARCH_NOTE.md

Reproducibility

Target: a graduate student with one Jetson Orin Nano can reproduce all benchmark numbers in under one hour. See docs/REPRODUCIBILITY.md (added in Phase 5).

Headline result (CIFAR-100, ResNet-34 backbone, 5 seeds per precision)

All numbers measured on a Jetson Orin Nano Super 8GB at MAXN_SUPER, JetPack 6.2, TensorRT 10.3.0. Power is full-board VDD_IN sampled at 200 ms cadence during 15 s of sustained inference. Energy per image converts power and throughput into the actual deployment-relevant metric.

Precision	Final-increment acc	p95 B=1	p95 B=64	Throughput B=64	Mean board power B=64	Energy / image B=64
FP32 (Jetson TRT)	0.6700 +/- 0.0014	2.71 ms	137.9 ms	469 img/s	22.0 W	46.9 mJ
FP16 (Jetson TRT)	0.6704 +/- 0.0013	1.34 ms	55.3 ms	1182 img/s	21.9 W	18.5 mJ
INT8 (Jetson TRT)	0.6719 +/- 0.0007	1.23 ms	22.8 ms	2856 img/s	20.0 W	7.0 mJ

Idle baseline: ~6.9 W. FP32 vs INT8 at B=64: 6.1x throughput, 6.7x energy efficiency, identical accuracy. All three precisions also agree within statistical noise on the full per-increment forgetting trajectory across all 50 class-incremental updates, not just at the final increment. INT8 quantization on a prototype-based CL backbone is therefore a free deployment win on the Jetson Orin Nano edge platform: no measurable accuracy or forgetting-dynamics degradation, six-fold gains in throughput and energy.

At B=1 (single-image HRI, where latency dominates): INT8 = 1.23 ms vs FP32 = 2.71 ms (2.2x faster) at 13.4 W vs 20.3 W (6.9 W headroom for other workloads on a power-constrained robot).

Charts:

• results/cifar100_resnet34_pareto_b64.png - latency vs accuracy at batch 64
• results/cifar100_resnet34_pareto_b1.png - latency vs accuracy at batch 1
• results/cifar100_resnet34_forgetting_curves.png - per-increment accuracy across all 50 increments
• results/cifar100_resnet34_energy.png - energy per inference and throughput vs batch size, per precision

See docs/RESEARCH_NOTE.md for the full position paper and the proposed Wizard-of-Oz follow-up tying this to perceived-competence latency thresholds in human-robot interaction.

License

PolyForm Noncommercial 1.0.0. The upstream CBCL-PR code retains its original MIT license, see upstream/LICENSE.

Citing the original work

@ARTICLE{ayub_tcds_2023,
  author={Ayub, Ali and Wagner, Alan R.},
  journal={IEEE Transactions on Cognitive and Developmental Systems},
  title={CBCL-PR: A Cognitively Inspired Model for Class-Incremental Learning in Robotics},
  year={2023},
  doi={10.1109/TCDS.2023.3299755}
}