An energy-efficient in-memory computing architecture for survival data analysis based on resistive switching memories

Area of Science:

• Biomedical Engineering
• Computer Science
• Data Science

Background:

• Precision medicine requires extensive survival data analysis for treatment optimization.
• Machine learning (ML) and deep neural networks are standard tools for this analysis.
• Current Von Neumann architectures face performance and energy efficiency limitations.

Purpose of the Study:

• To propose an in-memory computing (IMC) architecture for accelerating ML in precision medicine.
• To leverage resistive switching memory (RRAM) crossbar arrays for efficient matrix-vector multiplication.
• To enhance the performance of the DeepSurv neural network for biomedical survival analysis.

Main Methods:

• Developed an IMC architecture using RRAM crossbar arrays.
• Implemented matrix-vector multiplication in a single computational step.
• Explored synaptic weight mapping strategies and programming algorithms to address RRAM non-idealities.

Main Results:

• Demonstrated significant performance improvements for DeepSurv acceleration.
• Identified a performance/energy trade-off influenced by weight mapping and programming algorithms.
• Showcased the suitability of the IMC architecture for this specific biomedical application.

Conclusions:

• The proposed RRAM-based IMC architecture offers a revolutionary approach to overcome computational bottlenecks in precision medicine.
• This architecture provides substantial performance and energy efficiency gains for DeepSurv models used in survival data analysis.
• The findings highlight the specialized advantages of IMC over commodity systems for demanding biomedical ML tasks.