Transcengram
Transcengram is developing an explainable, white-box deep neural network architecture called CRATE, mathematically derived from optimization learning objectives and published at NeurIPS 2023. The company builds a parallel, distributed, multi-layer transformation autoencoder system toward general intelligence, with vision-centric multimodal alignment enabling geometric and semantic 3D and 4D generation and prediction. Its embodied brain training approach enables autonomous concept discovery and task-oriented closed-loop interactive co-evolution learning, with most training data automatically generated rather than manually collected.
- Artificial Intelligence
Funding
Founders
Founder details are not available yet.
Product
Problem
Current AI systems rely on opaque, black-box deep learning models that lack mathematical interpretability and struggle to learn generalizable intelligence from real-world physical environments. Existing approaches often require extensive manual data collection and annotation, limiting their ability to scale to autonomous physical tasks.
Solution
Transcengram is developing an explainable, white-box deep neural network architecture called CRATE, which is mathematically derived from optimization learning objectives. The company builds a parallel, distributed, multi-layer transformation autoencoder system as a neural architecture toward general intelligence, with vision-centric multimodal alignment enabling geometric and semantic 3D and 4D generation and prediction. Its embodied brain training approach enables autonomous concept discovery and task-oriented closed-loop interactive co-evolution learning, with most training data automatically generated rather than manually collected.
Target Audience
Primary customers are AI research institutions and companies developing embodied intelligence systems, autonomous robots, and multimodal AI applications that require interpretable, generalizable learning from physical environments.
Features
- White-box CRATE deep neural network architecture that is mathematically fully interpretable, derived from optimization learning objectives (published at NeurIPS 2023)
- Closed-loop autoencoding, self-learning network system forming a neural architecture toward general intelligence
- Binocular vision memory learning system achieving correct geometric reconstruction and semantic recognition of the real 4D world for vision-centric multimodal alignment
- Memory-based motor control model enabling autonomous concept discovery for embodied brain training and task-oriented closed-loop interactive co-evolution learning
- Lightweight data collection approach where most data is automatically generated, eliminating manual teleoperation and reducing annotation dependency
- Model adaptable across various environments with high generality across levels, scenes, and tasks