NeuroPixel represents a recurrent neural state on a spatial canvas. Local learned updates evolve the field; a shared dictionary reads the result. The graphics implementation executes perception, local updates, retina and decoding through floating-point image passes. It provides an inspectable artificial neural process, rather than a claim of a biological brain or a complete causal explanation.
Actual learned network trajectory of the selected VIS07 model. File docs/animations/actual_neural_trajectory.gif (339.6 KiB, SHA-256 3cc78dacc2d1…). Frames are actual GPU-rendered states of the selected model, labelled as such; no synthetic artwork. See the manifest.
Actual GPU-rendered states, selected VIS07 model, registered seed240 and fixed original TRAIN digit0. Playback is slowed for inspection. Colours display a projection of raw channels; magnitude is shown separately. The validated classifier uses eight updates and resets for each new input.
| Question | Complete measured result | Evidence and scope |
|---|---|---|
| Can rendering reproduce the neural computation? | 136 engineering checks passed on RTX3090, including pixel retina, state updates, masks, full logits and decisions. Learned deployment also retained all 320 tested decisions across five models. | GPU fidelity; FP32, fixed checks and explicit subset. |
| Does an external pixel-only task work? | 94.36% NeuroPixel versus 90.84% CNN; paired difference +3.52pp, exploratory t95 [1.93,5.10]pp. | VIS07: five paired fits, official UCI test1797images, DEV-only selection, near nominal parameter counts, differing functions/costs. |
| Can graphics execution be faster? | RGBA vector rendering measured 1.536× relative to the best CUDA comparator for C16/grid128. Other workloads do not uniformly win. | RENDER09: same GPU/outputs/FP32, CUDAeager andCUDAgraph, all seven workloads shown. |
| Does query-dependent spatial reading help binding? | GLOB05 attention minus active global uniform control +6.28pp, exploratory IC95 [4.11,8.45]pp. | GLOB05: twelve fixed bodies, equal registered budget/supervision; internal development, not external replication. |
| Are the negative results preserved? | Yes. OriginalH1 remains unsupported; DEV04 original precision target failed; long continuations are not stable. | DEV04, historical audit, limits. |
These are results of the specified recipes and device. They do not establish general architectural superiority, independent external replication or a Nobel-level discovery.
Capability figure from the pre-scientific README. Memory and rest-state behaviour are not established by the current evidence; see the manuscript for the bounded claims.
Scaling comparison. The original hypothesis H1 was not supported; see the results index.
The original scalar renderer uses RGBA32F textures and framebuffer passes. The vector renderer packs coefficients and performs four-channel fetches/DOT4. The advanced OpenGL 4.6 backend uses compute shaders, SSBO weights and shared pixel tiles. Its complete comparison passed 77 output gates and all 245 timing rows, preserving the original 143 rows without repeating them. Compute measures 2.084× relative to the best CUDA comparator for C16/grid128, and loses for C48/grid128. Recovery interrupted measurement sessions; results are descriptive technical measurements, not independent replications. All RENDER10 cases and limits.
The five advanced GPU backends on the same network state; provenance in advanced_GPU_backends_provenance.json. File docs/animations/advanced_GPU_backends.gif (71.3 KiB, SHA-256 efe22459cffc…). Frames are actual GPU-rendered states of the selected model, labelled as such; no synthetic artwork. See the manifest.
All graphics workloads and their comparators. File docs/animations/all_GPU_workloads.gif (161.4 KiB, SHA-256 87a3572584c1…). Frames are actual GPU-rendered states of the selected model, labelled as such; no synthetic artwork. See the manifest.
The first scalar implementation was 3.14–12.25× slower than the best measured CUDA comparator. It is retained as a within-run control. The vector variant improves that implementation; advantages are workload-dependent. CUDA Graph is included to reduce avoidable Python launch overhead. Technical timing blocks are not independent scientific replications.
Timing uses resident inputs, sixteen updates and the complete scanner. Setup/compilation, transfers, training and archival are distinct costs. Short NVML counter deltas cannot establish per-inference energy savings; zero readings are not zero consumption. Whole-system wall energy has not been measured.
RGB pixels feed a learned retina, neural state and shared decoder. No target label or class token is inserted as image input. Five new pairs received the same training examples, orders, labels and forty-epoch budget. Ten DEV-selected models were sealed before official test scoring. Code and original weights are preserved; inference can run through graphics, while this study's training used PyTorchCPU.
All five visual realizations of the VIS07 pairs. File docs/animations/external_vision_all_seeds.gif (126.6 KiB, SHA-256 50243fe28a7c…). Frames are actual GPU-rendered states of the selected model, labelled as such; no synthetic artwork. See the manifest.
The corpus is public UCI optical digits. The metadata documents separate writers for official TRAIN/TEST; exact input-image identities were disjoint. It is a bounded recognition benchmark, not natural-language reasoning, blind custody or independent laboratory confirmation. NCAclassification and WebGL execution have primary antecedents.
Finite-depth limits: state growth when the same input is extended beyond the eight trained updates. File docs/animations/depth_limits_and_state_growth.gif (133.7 KiB, SHA-256 09e0cca97300…). Frames are actual GPU-rendered states of the selected model, labelled as such; no synthetic artwork. See the manifest.
Eight updates are trained. Extending the same input to sixteen or thirty-two updates degrades accuracy and grows the recorded state magnitude. Ten extended-logit comparisons failed the original replay tolerance and remain archived, even though categorical decisions matched. A fixed same-host128-row diagnostic reproduced both code paths and archived values; it does not erase the earlier failed execution.
Dictionary scanner reading the canvas while it answers a question. File docs/img/thinking.gif (1315.2 KiB, SHA-256 339b5dcf6cd5…), generated by scripts/make_readme_figures.py.
The scanner projects state through a learned decoder. SCN06 demonstrates that identical present displays can hide state differences that affect a future answer. Seeing a colour is useful observability, not unique causal circuit identification or biological equivalence.
An interactive local viewer exposes all sixteen raw channels, selectable RGB projection, per-cell decoder probabilities and reset/depth controls. Archive mode passed fifty HTTP checks and browser controls. Live physical GPU integration passed 27 trajectory gates plus 52 HTTP checks, including all ten TRAIN resets and clean release. A separate CPU/Mesa execution passed the same integration checks; software and archive modes are labelled explicitly and are not presented as physical GPU execution.
The sustained energy protocol uses actual board counters over multi-second blocks. Its first execution passed all 77 output gates and preserved 103 of 105 blocks before the unchanged RAM floor stopped it. A recovery measured the two missing RGB8 block-2 keys; the completed cohort has 105 rows, 77 gates and zero repeated keys, with the per-case table in the recovered receipt folder. Full wall energy is unmeasured.
- Reproduction instructions and environments
- All-study evidence index
- Scientific manuscript: methods, results and open requirements
- Frozen plans, original checkpoints, datasets, raw predictions/states, failed attempts and hash manifests under
results/research/. - Scientific animation provenance: manifest. Animations derive from actual data; no synthetic brain artwork.
- Original README and prior versions are retained in the archive and Git history.
CPU studies used Python3.12.14/Torch2.6.0+cpu, two threads and a minimum8GiB availableRAM. GPU tests used RTX3090/NVIDIA581.29, Torch2.6.0+cu124 and ModernGL5.12/glcontext3. Training and heavy neural loading require the resource guard and shared GPU queue. Never rerun closed fits to improve an interval.
Code: MIT. Optical-digits data: Alpaydin & Kaynak(1998), UCI DOI10.24432/C50P49, CC BY4.0. Historical bAbI data: Weston and colleagues; the publisher dataset card declares CC BY3.0. The historical acquisition report retains its mirror and upstream-checksum limitations. Third-party sources and internal reproduction are explicitly distinguished. Independent replication, full energy accounting, broad stable memory and exceptional original utility remain open research requirements.