Skip to content

Archived Result Bundle for AEES Thesis and ICONIP 2026 Submission v0.1.1

Latest

Choose a tag to compare

@davidkfoss davidkfoss released this 22 May 00:47

Archived Result Bundle for AEES Thesis and ICONIP 2026 Submission

This release provides the archived experiment-result bundle used to regenerate the reported tables and figures for the AEES thesis and the ICONIP 2026 paper submission.

The reusable AEES implementation itself is maintained separately as the pulseopt package:

Provenance

This release provides the archived result bundle used for the AEES master’s thesis and the ICONIP 2026 conference paper submission.

This release corresponds to thesis experiment repository commit:

e2c8232

Reproducing the experiments should be done with pulseopt version 0.3.0.

Archive files:

  • archived_results_aees_thesis_iconip2026_v0.1.1.tar.gz
  • archived_results_aees_thesis_iconip2026_v0.1.1.sha256

Changes from v0.1.0

  • Added cifar_noise_ablation/ to the archived result bundle, containing preliminary CIFAR-100 gradient-noise ablation outputs.

Contents

The archive contains run-result JSON files and related compact result artifacts used by the reporting scripts in this repository. It is intended to support regeneration of aggregate tables and figures without rerunning the full training suite.

Expected extraction layout:

archived_results/
  checkpointing/
  cifar_clean/
  cifar_noisy/
  cifar_noise_ablation/
  clean_agnews/
  compute_overhead/
  instability_penalty_ablation/
  nlp_noise_ablation/
  noisy_agnews/
  sst2/
  trend_context_ablation/

Folder summary:

  • checkpointing/: peak-validation checkpoint rerun outputs used for checkpoint-diagnostics tables and figures.
  • cifar_noise_ablation/: preliminary CIFAR-100 gradient-noise ablation outputs.
  • cifar_clean/: clean CIFAR-100 experiment outputs.
  • cifar_noisy/: noisy CIFAR-100 runs, fixed-multiplier ablations, and related diagnostics.
  • clean_agnews/: clean AG News fine-tuning outputs.
  • compute_overhead/: timing outputs used for wall-clock overhead tables and figures.
  • instability_penalty_ablation/: supplementary ablation outputs for instability-penalty variants.
  • nlp_noise_ablation/: AG News sigma-axis falsification outputs.
  • noisy_agnews/: noisy AG News main and axis-ablation outputs.
  • sst2/: SST-2 fine-tuning outputs used in the thesis.
  • trend_context_ablation/: supplementary ablation outputs for trend-context variants.

The ICONIP 2026 paper submission uses a scoped subset of this archive, primarily:

archived_results/checkpointing/
archived_results/cifar_noisy/
archived_results/cifar_noise_ablation/
archived_results/noisy_agnews/
archived_results/nlp_noise_ablation/
archived_results/compute_overhead/

The thesis uses the broader archive, including clean CIFAR-100, clean AG News, SST-2, and supplementary ablation results.

What is excluded

This archive does not include model checkpoints, dataset caches, HuggingFace caches, raw temporary logs, or large transient training artifacts.

The archive is not intended to provide bitwise replay of every training trajectory. Instead, it provides the archived result files treated as the authoritative numerical source for the reported aggregate tables and figures.

How to use

Download the archive and extract it at the repository root so that the extracted directory is named:

archived_results/

The repository separates three artifact roles:

archived_results/        canonical archived result inputs
generated_artifacts/     curated tables, CSV summaries, and figures used in the thesis / paper submission
reproduced_artifacts/    local regenerated outputs for comparison

Regenerated outputs should be written to reproduced_artifacts/ and compared against the curated files in generated_artifacts/ before replacing any committed artifact.

Regenerating tables and figures

Both table-generation scripts (scripts/tables/) and figure-generation scripts (scripts/plots/) are configured with default paths. In normal use they run directly with no arguments: they read from archived_results/ and write to reproduced_artifacts/tables/... and reproduced_artifacts/figures/... respectively.

uv run python -m scripts.plots.cifar.plot_cifar_noisy_curves
uv run python scripts/tables/make_nlp_noise_ablation_table.py

To regenerate everything at once, use the bundled runners (from the repo root):

make plots      # or: ./scripts/plots/run_all.sh
make tables     # or: ./scripts/tables/run_all.sh
make artifacts  # both

A few figure scripts produce one figure per variant and accept an optional selector — cifar.plot_cifar_noisy_curves takes --noise-setting, and controllers.plot_arm_selection / controllers.plot_reward_per_arm take --setting. Omit the selector to emit all variants. Pass --runs-root/--out-dir to any script to override its default paths; check a script's --help for details.

When a figure script cannot find its inputs, it may exit non-zero and write a <figure-name>.MISSING.md file in the output directory explaining what was missing.

The TikZ method schematic is generated separately by:

bash scripts/plots/method/build_aees_overview.sh

Reproducibility notes

The archived files are the source used by the table and figure generation scripts. Re-running the original training jobs with the same seeds is expected to reproduce the same qualitative behavior and similar aggregate results, but exact bitwise or epoch-level reproduction across machines is not guaranteed.

Possible sources of variation include GPU architecture, CUDA/cuDNN kernels, PyTorch and torchvision versions, HuggingFace/transformers versions, DataLoader behavior, plotting backends, fonts, and other runtime details.

PDF and PNG figures may differ slightly across systems. For numerical checks, compare generated summary CSV or text files where available.

Checksum

A SHA-256 checksum file is provided alongside the archive. After downloading both files, verify with:

shasum -a 256 -c archived_results_aees_thesis_iconip2026_v0.1.1.sha256

License

The repository code is released under the MIT license. The archived result artifacts are provided for reproducibility of the thesis and ICONIP 2026 paper submission.