Feature Disentangling and Combination Implemented by Spin-Orbit Torque Magnetic Tunnel Junctions

Feature disentangling plays a vital role in diverse domains such as image processing, drug development, and pattern recognition. However, this process is computationally intensive. Recently, spintronic devices, such as magnetic tunnel junctions (MTJs), have been increasingly employed to accelerate AI in hardware. Herein, an efficient algorithm that both disentangles and generates combined features from data is proposed. The effectiveness of spin-orbit torque magnetic tunnel junctions (SOT-MTJs) in disentangling and combining features of emoji images is experimentally demonstrated. The algorithm is further validated on real-world facial datasets, including CelebFaces Attributes (CelebA) and Protein Families Database (Pfam). The results confirm that SOT-MTJs can be effectively used as true random number generators. When combined with the algorithm, these devices enable both feature disentangling and integration, thereby expanding their potential application landscape.