license: apache-2.0
library_name: pytorch
pipeline_tag: image-classification
description: "Evaluation results achieved using a Level 2 CubicNN (Megacube) trained for 180 epochs."
tags:
- architecture-only
- vision
- classification
- experimental
datasets:
- cifar10
- cifar100
metrics:
- accuracy
model-index:
- name: CubicNN
results:
- task:
type: image-classification
name: Image Classification
dataset:
name: CIFAR-10
type: cifar10
metrics:
- type: accuracy
name: Top-1 Accuracy (Level 2 Megacube)
value: 87.14
- type: accuracy
name: Top-5 Accuracy (Level 2 Megacube)
value: 99.50
- task:
type: image-classification
name: Image Classification
dataset:
name: CIFAR-100
type: cifar100
metrics:
- type: accuracy
name: Top-1 Accuracy (Level 2 Megacube, 180 epochs)
value: 56.42
- type: accuracy
name: Top-5 Accuracy (Level 2 Megacube, 180 epochs)
value: 83.23
license: apache-2.0 library_name: pytorch pipeline_tag: image-classification tags:
architecture-only
vision
classification
experimental datasets:
cifar10
cifar100 metrics:
accuracy model-index:
name: CubicNN results:
task: type: image-classification name: Image Classification dataset: name: CIFAR-10 type: cifar10 metrics:
type: accuracy name: Top-1 Accuracy value: 87.14
type: accuracy name: Top-5 Accuracy value: 99.50
task: type: image-classification name: Image Classification dataset: name: CIFAR-100 type: cifar100 metrics:
type: accuracy name: Top-1 Accuracy value: 56.42
type: accuracy name: Top-5 Accuracy value: 83.23
Model Card for CubicNN
CubicNN is a percent-routed 3×3×3 hierarchical neural network. Each vector is routed into 27 “cells,” refined through neighbor mixing, and expanded hierarchically (kilocube → megacube → …). The design is experimental, aiming to explore recursive cube routing as a structural principle for neural networks.
Model Details
Model Description
Developed by: Chris (independent project)
Model type: Hierarchical cube-based classifier
Framework: PyTorch
License: Apache 2.0
Finetuned from model: Trained from scratch
Model Sources
Repository: Hugging Face Repo
Paper: Not published
Demo: Coming soon
Uses
Direct Use
Image classification on MNIST, CIFAR-10, and CIFAR-100.
Research into cube-based recursive architectures.
Downstream Use
Finetuning on small to medium image datasets.
Experimental multimodal setups using UniversalCubeClassifier.
Out-of-Scope Use
Safety-critical applications.
Production-scale deployment.
Bias, Risks, and Limitations
Only tested on academic datasets (MNIST, CIFAR).
No fairness or robustness auditing.
Scaling beyond 3 levels is computationally prohibitive.
Recommendations
Use for research and educational purposes. Avoid applying in real-world decision-making systems.
How to Get Started with the Model
from huggingface_hub import hf_hub_download
import torch
from model import CubicConfig, CubeClassifier
ckpt = hf_hub_download(repo_id="/CubicNN", filename="runs/best.cube")
cfg = CubicConfig(dim=64, levels=3, n_out=100, in_ch=3)
model = CubeClassifier(cfg)
state = torch.load(ckpt, map_location="cpu")
model.load_state_dict(state["model"], strict=False)
model.eval()
Training Details
Training Data
MNIST
CIFAR-10
CIFAR-100
Training Procedure
Optimizer: AdamW
Epochs: 100 (default)
Batch Size: 128 (train), 256 (eval)
Precision: mixed precision (torch.amp)
Evaluation
Architecture Hierarchy
CubicNN follows a recursive 3×3×3 cube hierarchy:
Level 0: Node × 27 = Cube
Level 1 (Kilocube): Cube × 27 — Toy scale
Level 2 (Megacube): Kilocube × 27 — Low-end consumer scale
Level 3 (Gigacube): Megacube × 27 — High-end consumer scale
Level 4 (Teracube): Gigacube × 27 — Commercial scale
Level 5 (Petacube): Teracube × 27 — Beyond commercial scale
The reported evaluation metrics were achieved using Kilocube (Level 1) and Megacube (Level 2) models.
Evaluation
