Tree-based machine learning performed in-memory with memristive analog CAM.

Area of Science:

• Computer Science
• Materials Science
• Electrical Engineering

Background:

• Tree-based models like Decision Trees and Random Forests offer interpretability and perform well with limited data.
• Conventional von Neumann architectures struggle with optimizing these models for fast, large-scale inference due to memory access bottlenecks.
• Emerging memristor devices offer potential for novel computational paradigms.

Purpose of the Study:

• To propose and demonstrate the use of analog content addressable memory (CAM) as an in-memory computational primitive for accelerating tree-based machine learning inference.
• To develop an efficient mapping algorithm for programming Decision Tree paths into CAM rows.
• To evaluate the performance gains of this in-memory compute approach.

Main Methods:

• Proposed a novel analog CAM based on memristor devices for fast look-up table operations.
• Developed an efficient mapping algorithm to represent Decision Tree root-to-leaf paths as rows in the analog CAM.
• Leveraged in-memory compute capabilities of the CAM for model inference.

Main Results:

• Demonstrated that each root-to-leaf path of a Decision Tree can be programmed into a row of the analog CAM.
• Achieved few-cycle model inference, enabling significant acceleration.
• Showcased a throughput increase of approximately 1000× compared to conventional approaches.

Conclusions:

• Analog CAM serves as an effective in-memory computational primitive for accelerating tree-based model inference.
• The proposed in-memory compute concept dramatically enhances inference throughput for Decision Trees and Random Forests.
• This approach overcomes the limitations of von Neumann architectures for efficient tree-based model deployment at scale.