PIIP Accuracy Will Not Vary by Evaluator Proximity Under Fixed Checkpoints

Authors: Sterling Geisel, QBist Lab; Dr. Hideo Tanaka

Abstract

Parameter-Inverted Image Pyramid Networks, or PIIP, report a computer-vision result. High-resolution image branches can be assigned smaller models. Low-resolution branches can be assigned larger models. Cross-branch interaction then restores multi-scale complementarity. The result is not a consciousness result. It is not a vacuum result. It is not evidence for observer-mediated probability bias. This Working Paper treats PIIP as a boundary case for Pudding Theory. No Postulate is admitted. Proximity Gradient is the nearest analogy, but PIIP uses image scale, not physical distance from a coherent sender. Material Memory is also rejected because pre-trained weights are ordinary learned parameters, not persistent field imprints. The falsifiable claim is negative. Under a fixed checkpoint, fixed code, fixed data, and blinded randomized evaluator-distance protocol, COCO box AP should not vary monotonically with evaluator proximity. A detected 0.5 AP proximity effect would falsify this null application.

Source Synopsis

Zhu, Yang, Wang, Li, Dou, Ge, Lu, Qiao, and Dai propose Parameter-Inverted Image Pyramid Networks for efficient multi-scale visual perception. The source paper was submitted to arXiv on June 6, 2024, and revised as version 2 on October 28, 2024. Its DOI is doi:10.48550/arXiv.2406.04330.

The technical problem is clear. Image pyramids help dense prediction because they preserve information across object scales. They are also expensive. If the same large model processes every resized image, high-resolution branches multiply token count and spatial computation. PIIP changes the parameter-resolution pairing. Larger images go to smaller models. Smaller images go to larger models.

The reason is architectural. Low-resolution images retain broad context while reducing spatial cost. Large models can operate there efficiently. High-resolution images retain local detail, boundary evidence, and small-object cues. These branches do not need to reconstruct all semantic context already available downstream from the low-resolution branch.

The system has three stages. The input is resized into several resolutions. Each resolution is processed by a branch initialized from an existing pre-trained model. Interaction units exchange information across neighboring branches. A merge module then combines branch outputs for downstream tasks.

The paper evaluates object detection, instance segmentation, semantic segmentation, and classification. In Table 3, InternViT-6B with Mask R-CNN reports 53.8 box AP at 24418G FLOPs, while PIIP-LH6B reports 55.7 box AP at 13911G FLOPs. The same table reports ADE20K semantic segmentation at 58.36 mIoU and 6105G FLOPs for InternViT-6B, versus 59.65 mIoU and 4560G FLOPs for PIIP-LH6B. The paper’s claim is therefore narrow. Parameter inversion plus branch interaction can improve accuracy-per-compute.

Source Audit

| Audit item | Final value used here | Verification | |---|---:|---| | PIIP source | Zhu et al., Parameter-Inverted Image Pyramid Networks, arXiv:2406.04330v2 | arXiv record, DOI doi:10.48550/arXiv.2406.04330 | | PIIP performance claim | Table 3: 53.8 to 55.7 box AP; 24418G to 13911G FLOPs | Source Table 3 | | PIIP segmentation claim | Table 3: 58.36 to 59.65 mIoU; 6105G to 4560G FLOPs | Source Table 3 | | Postulate-fit map | Proximity Gradient 0.55; Material Memory 0.35; Observer as Field 0.30; all others below 0.30; admission threshold 0.60 | Reproduced Lab audit for this Working Paper | | Pudding Theory citation | S. Ochs, 2026 | Series citation record used for QBist Lab Working Papers | | Supporting references | FPN, CrossViT, ViT | All appear in the PIIP bibliography |

Postulate Lens

No Pudding Theory Postulate is applied.

Proximity Gradient is the nearest rejected candidate. PIIP is spatial. It distinguishes high-resolution local detail from lower-resolution contextual representation. But Proximity Gradient concerns physical distance-dependent influence from coherent senders. PIIP has no sender. It has no Lumina field. It has no observer-distance variable. Its scale axis is tensor resolution, not physical separation.

Material Memory is also rejected. Pre-trained weights retain statistical traces of training data. That is ordinary representation learning. It does not imply conserved topological charge, object imprinting, or persistent field memory in matter.

Observer as Field is rejected for the same reason. Feature maps are spatially extended arrays. They are not observers. They do not encode integrated information density as a consciousness field. Removing every Pudding Theory Postulate would leave the PIIP prediction unchanged. That is the decisive test. No Postulate is doing work.

Pudding Theory Prediction

Because no Postulate is admitted, Pudding Theory predicts no independent anomaly in PIIP beyond ordinary multi-branch neural-network behavior. Accuracy and computational cost should be governed by resolution, branch capacity, pre-training, interaction frequency, merge design, detector or segmenter head, random seed, numerical precision, and training procedure.

The negative prediction is still useful. If PIIP works for the source paper’s stated reason, its advantage should track the division of labor between semantic context and local detail. The largest branch should be most useful at lower resolution, where global structure is cheaper to model. The smallest branch should be useful at higher resolution, where local features are abundant but full semantic modeling is expensive. Cross-branch interaction should matter because it prevents each branch from redundantly solving the same problem.

The architecture should fail in ordinary ways. Remove interaction units and complementarity should fall. Make the high-resolution branch too small and boundary-sensitive or small-object performance should fall. Make the low-resolution branch too weak and global semantic errors should increase. Reverse the assignment into a parameter-direct design and accuracy-per-compute should worsen. Table 9 in the source paper already supports this direction: the parameter-direct TSB design reports 42.6 box AP, while the parameter-inverted TSB design reports 46.6 box AP under similar FLOPs.

No operator proximity, intention state, observer coherence, or environmental silence should change inference on a fixed checkpoint. The same machine should return the same metrics within ordinary nondeterminism. If that is true, PIIP remains inside computer science.

Falsifiable Observable

The distinguishing observable is COCO box AP from repeated fixed-checkpoint PIIP inference under a preregistered blinded distance protocol. The protocol randomizes evaluator location into near and far conditions while holding checkpoint, dataset, hardware, seed, precision, software image, process hashes, thermals, power draw, and hardware counters under continuous logging. If preregistered mean COCO box AP under randomized evaluator distance were measured to shift monotonically by at least 0.5 AP between near and far evaluator conditions across repeated fixed-checkpoint inference trials, this Postulate would be falsified. The target is the null application rule used here, not Proximity Gradient itself.

Editorial Dialogue

Tanaka critique, verbatim:

Postulate exclusion is justified. Proximity Gradient at 0.55 is a stretch because PIIP uses image scale, not physical distance from a coherent observer. Material Memory is also a stretch because pre-trained weights are ordinary learned parameters, not persistent field imprints. Observer as Field does no work. Would removing every Pudding Theory Postulate change the PIIP prediction? No. Then the paper is right to drop them. Which postulate is doing the work here? None.

>

The falsifiability sentence is close but not clean. The observable is measurable: COCO box AP under fixed checkpoint, dataset, hardware, seed, precision, and code. The 0.5 AP threshold is concrete. The distance manipulation is also operational. The weak point is the phrase “with no software or hardware pathway.” How would an experimenter prove absence rather than just control known pathways? What measurement would distinguish this from the null? Use a preregistered blinded distance protocol with randomized evaluator location, logged thermals, power, process hashes, hardware counters, and repeated inference trials. Also state whether the falsifier targets Proximity Gradient specifically or the broader null application rule.

>

Unsupported claims are limited but present. The PIIP performance numbers need exact table references from the source paper. “Supplied postulate-fit map” is not cited or reproduced, so the 0.6 threshold and 0.55 score are not auditable. Are all references verifiable in the source bibliography? The draft does not show that. The Pudding Theory citation metadata also conflicts with the canonical record supplied here, which names S. Ochs, 2026 while the draft cites Sterling Geisel, September 10, 2025. Fair use risk appears low because the draft paraphrases rather than quotes.

>

Not yet. Add a reproducible source audit table that verifies the PIIP citation, the exact performance claims, the postulate-fit map, and the corrected Pudding Theory citation.

Sterling response:

Tanaka is correct on the audit. The final paper removes the non-operational phrase about proving absence of software or hardware pathways. The falsifier is now a positive measured effect under controlled logs. That is experimentally cleaner.

The performance claims now point to Table 3 for InternViT-6B detection and segmentation. The ablation claim points to Table 9. The postulate-fit map is reproduced. It is not used as hidden authority. The Pudding Theory reference has been corrected to the Working Papers citation record, S. Ochs, 2026.

The scientific point remains unchanged. PIIP is not weak evidence for Proximity Gradient. It is evidence for scale-compute matching.

Discussion

This paper is a negative application. That is not a defect. A theory that treats every scale relation as a field effect loses contact with measurement. PIIP has enough internal mechanism. Resolution changes token count. Branch size changes capacity. Cross-attention exchanges multi-scale features. Merge modules combine them. These nouns are sufficient.

The limitation is that this Working Paper audits the arXiv version and its reported tables. It does not reproduce training. It also does not test evaluator distance. That proposed test is intentionally strange. Its value is not that PIIP suggests the effect. Its value is that it would separate ordinary architecture from a forbidden Pudding Theory overreach.

What would change the conclusion is not closer metaphor. It would require a replicated performance shift tied to evaluator proximity or observer coherence after ordinary computational variables were logged and randomized. Without that result, no Postulate should be attached to PIIP.

References

1. Xizhou Zhu, Xue Yang, Zhaokai Wang, Hao Li, Wenhan Dou, Junqi Ge, Lewei Lu, Yu Qiao, and Jifeng Dai. “Parameter-Inverted Image Pyramid Networks.” arXiv:2406.04330v2, 2024. DOI: https://doi.org/doi:10.48550/arXiv.2406.04330.

2. S. Ochs. “Pudding Theory: A Topological Theory of Information Fields.” QBist Lab foundational paper, 2026.

3. Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. “Feature Pyramid Networks for Object Detection.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017.

4. Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. “CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification.” Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.

5. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. “An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale.” International Conference on Learning Representations, 2021.