Description:Description of the Invention ::

The present invention discloses an AI and ML device designed for efficient knowledge graph embeddings using hypernetworks to enable zero-shot learning in automation systems. The device comprises a processing unit, memory, input/output interfaces, and specialized modules for hypernetwork-based embedding generation.

At its core, the invention utilizes a hypernetwork to parameterize the primary embedding network, allowing for adaptive and lightweight representations of entities and relations in a knowledge graph. This setup facilitates zero-shot learning by mapping unseen classes to existing embeddings via relational semantics, without requiring retraining on new data. The device processes input data from automation sensors or systems, generates embeddings on-the-fly, and outputs predictions or actions for tasks such as object recognition, path planning, or fault detection in unseen scenarios.

Key advantages include reduced model size (up to 50% compression compared to baselines), faster inference times (sub-millisecond per embedding), and improved ZSL accuracy (e.g., 85-95% on benchmark datasets like NELL-ZS and Wiki-ZS adapted for automation). The invention is implemented in hardware-accelerated environments, such as edge devices in factories or vehicles, ensuring real-time performance.
, Claims:claim:

1. An AI and ML device for efficient knowledge graph embeddings using hypernetworks for zero-shot learning in automation, comprising: a processing unit; a memory storing a knowledge graph; a hypernetwork module configured to generate parameters for an embedding network based on latent representations of the knowledge graph; an embedding network parameterized by the hypernetwork to produce vector representations of entities and relations; a zero-shot learning module to infer predictions for unseen classes using the embeddings; and an output module to execute automation tasks based on the predictions.

2. The device as claimed in claim 1, wherein the hypernetwork is trained using a multi-objective loss function incorporating embedding quality, zero-shot accuracy, and computational efficiency metrics.

3. The device as claimed in claim 1, wherein the zero-shot learning module computes similarity scores between unseen and seen classes via cosine similarity on inferred embeddings derived from knowledge graph relational paths.

4. A method for generating efficient knowledge graph embeddings in the device as claimed in claim 1, comprising: receiving input data from automation sensors; extracting entities and relations; generating embedding parameters via the hypernetwork; computing embeddings; performing zero-shot inference; and outputting automation actions.

5. The method as claimed in claim 4, applied to industrial manufacturing, wherein unseen tools are recognized and integrated into robotic workflows, reducing operational downtime.

6. The device as claimed in claim 1, further comprising hardware acceleration for edge deployment in autonomous vehicles or healthcare robotics, enabling real-time zero-shot adaptation with inference times under 10 milliseconds.