Trustworthy tree-based machine learning by MoS₂ flash-based analog content-addressable memory with inherent soft...

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Advancements in artificial intelligence (AI) raise concerns about trustworthiness, particularly interpretability and robustness.
• Tree-based models like Random Forest offer interpretability but are computationally expensive to scale.
• Prior attempts to accelerate tree-based models using analog content-addressable memory (CAM) faced challenges due to device variations and susceptibility to adversarial attacks.

Purpose of the Study:

• To develop a hardware-software co-design for efficient inference with soft tree-based models.
• To address the limitations of traditional sharp decision boundaries in analog CAM implementations.
• To improve the robustness and accuracy of AI models against device variations and adversarial attacks.

Main Methods:

• A hardware-software co-design approach utilizing MoS2 flash-based analog CAM with inherent soft boundaries.
• Fabrication of analog CAM arrays for efficient inference.
• Experimental calibration and testing of the developed model on benchmark datasets.

Main Results:

• Achieved 96% accuracy on the Wisconsin Diagnostic Breast Cancer dataset using fabricated analog CAM arrays.
• Demonstrated high robustness to device variations, with only a 0.6% accuracy drop on MNIST under 10% device threshold variation.
• Outperformed traditional decision trees, which experienced a 45.3% accuracy drop under similar conditions.

Conclusions:

• The proposed MoS2 flash-based analog CAM offers an efficient and robust solution for AI inference.
• Soft boundaries in analog CAM provide inherent resilience to device variations and adversarial attacks.
• This hardware-software co-design approach significantly enhances the trustworthiness and scalability of tree-based AI models.