We morph a polyhedron till it thinks.
MORPHEDRON is a Vienna-based deep-tech company that builds a neural-network engine in which no weight is trained. Neurons are points in space, and every weight follows from the distance between them.
Modern AI scales the wrong way.
Modern AI gains capability from data and parameters. Data is finite, and scaling parameters is increasingly expensive. The resulting models are opaque, and those who deploy them face mounting explainability demands.
For decades the field has assumed that neural-network weights must be trained. We challenge this practice at the root and propose a new AI paradigm.
The world needs a new type of AI: linearly scalable, highly capable and inherently explainable. The state of the art lacks the fundamentals to achieve all three, and it will keep lacking them until its fundamentals change.
Neurons are points. Weights are distances.
MORPHEDRON trains no weights. It positions its cells in a D-dimensional Euclidean space, and the weight of every connection is computed from the distance between the connected cells. The trainable parameters of a MORPHEDRON model are coordinates.
A trained MORPHEDRON model is therefore a shape: a set of points you can render, inspect and hand to an auditor.
Where MORPHEDRON sits in the AI stack.
MORPHEDRON sits at the architecture and engine layer: beneath the models that are built with it, above the hardware that runs it. Complements, not competitors. Established architectures and general frameworks are toolkits you build with. MORPHEDRON is a distinct paradigm: its own architecture, whose fundamentals are published as DEBI-NN, plus a purpose-built engine.
- Application layerassistants, agents, products
- API layerhosting, serving
- Model layerlanguage, vision, tabular, spiking
- Architecture / Engine layerMORPHEDRON
- Infrastructure layercloud, accelerators
One principle. Compounding advantages.
Small by design
- Scales linearly. Parameters grow with neuron count, not with its square.
- Compact. Dramatically fewer trainable parameters at identical topology. Models are easy to store, share and deploy, because weights can be recalculated from distances.
- Built for small data. Suited to small, niche and imbalanced datasets, not only large ones. In continual learning, expanding a model's capability does not require expanding the training data at the same pace.
- Regularizes itself. Moving one cell changes every connection it has, so the trainer must find coordinates that satisfy all connected distances at once. In a multi-centric study on three cancer cohorts and five network layouts, MORPHEDRON models outperformed conventional networks of identical architecture while relying on little or no explicit regularization (Ecsedi et al., 2025).
- Prunes itself, reversibly. Cells that drift far away stop contributing during training, reversibly, and can be detected afterwards, without hand-set thresholds and without stored rollback states.
Human readable
- Explainable by construction. You inspect the model itself, not a surrogate fitted to it.
- Auditable. The model is a set of points you can render and inspect directly. Its complete specification is a set of coordinates plus one distance-to-weight function: a deterministic description of its decision logic.
- Calibrated without extra models. Predictive uncertainty and ensembles come from perturbing coordinates, with nothing extra to train.
- Visible while it learns. A model can be observed during training, after training and at inference time, and what you observe is its geometry.
Universal fit
- Any trainer. Optimize the metric that fits your model or data, including gradient-free training on objectives that backpropagation cannot use. Business rules can become the training objective.
- Any signal regime. One parameterization models conventional artificial and spiking networks alike, so builders of spiking systems need no surrogate gradients.
- Frame-free. Only relative position matters, so the model is invariant to how the space is oriented. No alignment or orientation has to be fixed by the builder.
- Backward compatible. A trained MORPHEDRON network can be deployed as conventional weight matrices on any existing inference stack.
- Two models in one. The geometric model can stay behind your walls and keep training, while the derived weights ship in your products without exposing the knowledge behind them.
Built for those who build.
MORPHEDRON is of interest to every AI model builder and deployer who wants to:
- cut the complexity of building models: less training data, fewer parameters, simpler training;
- protect know-how when deploying: keep the geometric model in-house, ship the weights;
- keep training in-house or in a self-controlled cloud: MORPHEDRON accesses no training data at all, and the engine is self-hosted.
Healthcare and medical devices
Where the research began: small, multi-center clinical cohorts, and AI embedded in regulated products.
Regulated decision systems
Credit, insurance, employment, critical infrastructure and the other high-risk uses listed in the EU AI Act, where a model must be presentable to an auditor.
Robotics and spiking systems
Weights and signal timing from geometry, sparsity as a property of the model, no surrogate gradients.
Any builder with small data
Niche, imbalanced or expensive datasets where large models cannot be trained at all.
Compliance, computed
A MORPHEDRON model's complete specification is a set of coordinates and one distance-to-weight function: a deterministic, auditable description of its decision logic. For the EU AI Act's high-risk obligations, applying from 2 December 2027 (Annex III) and 2 August 2028 (regulated products such as medical devices), the engine emits a provenance manifest per run, integrity hashes and a versioned audit log, on-premise and vendor-neutral.
Curious minds
Organizations operating in regulated domains, working with small or imbalanced data, or requiring a model that can be presented to an auditor are the intended near-term users. Evaluation access and pilot engagements are open.
Investors
A published scientific basis, a working engine, and an architecture whose compliance artifacts are emitted natively. Early discussions are welcome.
Published and peer-reviewed.
The fundamentals of MORPHEDRON and their advantages are published and demonstrated in peer reviewed manuscripts. A lot has happened since then.
- Papp L, Haberl D, Ecsedi B, Spielvogel CP, Krajnc D, Grahovac M, et al. DEBI-NN: distance-encoding biomorphic-informational neural networks for minimizing the number of trainable parameters. Neural Networks 167, 517–532 (2023). doi.org/10.1016/j.neunet.2023.08.026
- Ecsedi B, Boukhari A, Spielvogel CP, Haberl D, Ritter Z, Bundschuh RA, Lapa C, Hacker M, Hatt M, Papp L. Impact of regularization in optimizing distance-encoding biomorphic-informational neural networks for small nuclear medicine datasets. EANM Innovation 1, 100008 (2025). doi.org/10.1016/j.eanmi.2025.100008
- Boukhari A, Ecsedi B, Abdallah N, Haberl D, Spielvogel C, Papp L, et al. Three-dimensional visualization of DEBI-NNs for supporting interpretability of predictive models relying on neural networks. Proceedings of the ICCR (2024).
Founders.
MORPHEDRON FlexCo is a spin-off of the Medical University of Vienna, based in Vienna, Austria.
Laszlo Papp, PhD
Computer scientist and medical physicist (PhD, Award of Excellence), specialized in AI for more than a decade, and the inventor of the MORPHEDRON paradigm. He co-founded Dedicaid, acquired by Telix Pharmaceuticals in 2023. Before his academic career he spent ten years in the medical imaging industry as head of software R&D.
Martin Jankoski, MSc
A deep-tech operator (MSc in Innovation and Management, TU Munich) who spent five years as a venture capitalist at CNB Capital, leading deep-tech investments and holding board seats.
Frequently asked questions.
What is MORPHEDRON?
MORPHEDRON is a Vienna-based company that builds a neural-network engine in which no weight is trained. Its cells are positioned in space, and every weight is computed from the distance between connected cells. Its fundamentals are published under the name DEBI-NN.
Is MORPHEDRON a large language model or a world model?
No. MORPHEDRON is not an LLM, not a world model, and not a tuned version of any known model scheme. It supplies the building blocks from which such models can be built by others.
What is the role of the Medical University of Vienna?
MORPHEDRON is a spin-off of the Medical University of Vienna, where the research behind it began. MORPHEDRON is not a clinical or medical AI company: it builds a new neural-network engine with a wide range of use cases across industries. The paradigm originated in a high-risk healthcare environment because that is where the need for this kind of AI, small-data capable and auditable, became unavoidable.
Do you ship pre-trained models?
No. MORPHEDRON ships no pre-trained model and offers access to none. It supplies an engine with which builders train their own models.
Where does my data go?
Nowhere. The engine is self-hosted, on-premise or in your own cloud, and MORPHEDRON has no access to your training data.
How is a MORPHEDRON model explainable?
A trained model is a set of coordinates and one distance-to-weight function. It can be rendered and inspected directly during training, after training and at inference time. No post-hoc explainer approximates it, so there is no disagreement between explainers to reconcile.
What does this have to do with the EU AI Act?
High-risk obligations apply from 2 December 2027 for Annex III systems and from 2 August 2028 for AI embedded in regulated products such as medical devices. A deterministic, inspectable model specification, together with natively emitted provenance manifests, integrity hashes and audit logs, is what those obligations call for.
How mature is the engine?
It is a working engine whose core is validated on tabular cohorts. The product layer for deployment is on the roadmap.
Can a trained model run on my existing inference stack?
Yes. A trained MORPHEDRON network can be exported as conventional weight matrices and run on any existing inference stack.
How do we engage?
Evaluation access and pilot engagements are open to organizations in regulated domains, with small or imbalanced data, or needing a model that can be presented to an auditor. Investors are welcome for early discussions. Write to martin@morphedron.ai.