ECCV 2026 NS-PFS OMT Image + Tabular NeuroAI
Interactive Research Demo

iSyncTab

Neural Synchrony Arena

Explore cross-modal feature sequencing through an interactive multimodal AI experience. Challenge the HAM10000-trained iSyncTab model, explore your own multimodal datasets, and see how neural synchrony connects heterogeneous representations.

🧑 Human VS 🤖 iSyncTab
⚡ Enter the Arena Find it on ECCV Page

Choose Your Experience

The Arena provides pretrained HAM10000 inference, external dermoscopy, compact image-tabular experiments, and small audio-video token learning.

Live
🎮
HAM10000 Challenge Human versus iSyncTab skin-lesion classification game using the released HAM10000 checkpoint.
Live
🔬
External Dermoscopy Upload a dermoscopic image outside HAM10000. The existing Arena will automatically request the metadata needed by the HAM-trained checkpoint.
Live
🧩
Small Image-Tabular Lab Upload a compact ZIP dataset and train a lightweight generalized iSyncTab experiment on ZeroGPU.
Live
🎬
Small Audio-Video Lab Upload precomputed audio-video token features and run a compact iSyncTab_AV training experiment.

🎮 HAM10000 Challenge

Upload a HAM10000-style dermoscopic image, make your own prediction, and then challenge iSyncTab.

HAM or external dermoscopy: upload an original HAM10000 image to load its metadata automatically, or upload any other dermoscopic image and the Arena will open the manual metadata panel for out-of-dataset inference.
Round 01 Can you beat a multimodal neural synchrony model?
Connecting...
01 · Dermoscopic Image
Upload an image. If its original filename matches a HAM10000 image ID, the Space will retrieve its metadata from Kaggle and keep the true label hidden until reveal.
02 · What is your prediction?
NS-PFS
Neural Synchrony
0 Human wins
0 iSyncTab wins
0 Ties
0 Scored rounds
Round result
You —
iSyncTab —
True label —
Model probabilities
Multimodal inference
NS-PFS order preview
⚕️ Research demo only. iSyncTab-HAM10000 is a research artifact and is not a medical device. Predictions are not intended for diagnosis, treatment, patient management, or clinical decision-making. HAM10000 metadata is retrieved from its Kaggle distribution at runtime; the original dataset images are not bundled with this Space.

🧩 Small Image-Tabular Lab

Train a compact generalized iSyncTab experiment on your own paired image-tabular dataset. This public lab is intentionally capped for ZeroGPU and demonstration use.

Dataset

Upload one ZIP containing metadata.csv and the paired images. All CSV columns other than the image and label columns are used as tabular features. Low-cardinality categorical values are encoded automatically.

Expected ZIP format 
dataset.zip
├── metadata.csv
└── images/
    ├── sample_001.jpg
    ├── sample_002.jpg
    └── ...

metadata.csv:
image,label,age,feature_2,category,...
images/sample_001.jpg,class_a,42,...
Experiment Result

The quick lab freezes the ImageNet-pretrained ResNet-50 stem and trains the iSyncTab projection, tabular tokenization, NS-PFS/OMT path, and classifier. Use the full PyPI/GitHub workflow for serious experiments and hyperparameter tuning.

— Test Accuracy
— Paired Samples
— Encoded Tabular Features
NS-PFS order preview
Upload a compact paired dataset to run the generalized image-tabular workflow.

🎬 Small Audio-Video Lab

Run iSyncTab_AV on small paired precomputed audio and video token representations. This follows the public audio-video API directly.

Token Dataset

Upload one NPZ with arrays audio, video, and y. iSyncTab_AV expects token-level representations from your chosen upstream audio and video encoders.

Expected NPZ format 
audio: (N, audio_len, audio_dim)
video: (N, video_len, video_dim)
y:     (N,)

Example:
np.savez(
    "av_tokens.npz",
    audio=X_audio,
    video=X_video,
    y=labels,
)
Experiment Result

The lab trains a compact NS-PFS + OMT audio-video classifier from the supplied token representations. For raw WAV/MP4 datasets, extract token features with your preferred upstream encoders first.

— Test Accuracy
— Paired Samples
— Audio / Video Tokens
NS-PFS order preview
Upload precomputed token representations to run the generalized audio-video workflow.