Design and Implementation of a High Speed CMAC Neural Network Using Programmable CMOS Logic Cell Arrays

NeurIPS 1990

TL;DR

This paper presents a high-speed, hardware implementation of the Cerebellar Model Articulation Controller (CMAC) neural network using programmable CMOS logic. They developed two VME bus boards capable of implementing multiple independent CMAC networks with a million adjustable weights, achieving response times under 1 millisecond, making them suitable for real-time robotic control and signal processing.

What Problem Does This Paper Address?

Traditional CMAC neural networks, while effective for tasks like robotic control and pattern recognition due to their ability to learn complex mappings and generalize, often suffer from slow response times when implemented in software. This slowness limits their application in real-time systems where rapid decision-making is crucial. The authors aimed to overcome this by designing and implementing a compact, high-speed CMAC system using dedicated hardware, specifically programmable CMOS logic cell arrays. This would allow for significantly faster processing than software-based implementations, opening up new possibilities for real-time applications.

Methods

The core of the paper is the hardware design of a CMAC associative memory system. They developed two custom devices implemented using XILINX programmable logic cell arrays:

  • Input Associative Mapping Device: This device takes input vectors and converts them into traditional RAM addresses. It employs a recursive bit mapping scheme and pipelined logic for efficient hashing of virtual receptive field addresses to physical RAM addresses. This process generates an 18-bit RAM address for each input vector.
  • CMAC Response Accumulation Device: This device acts as an accumulator. As addresses are generated by the mapping device, this unit sums the corresponding weight values from a data RAM. During training, it also handles weight adjustments.

The final VME bus system integrates these components. It uses one CMAC associative mapping device (handling overlapping receptive fields sequentially) and one CMAC response accumulation device (providing eight parallel output channels). A 1-million weight memory is implemented using static RAM SIMMs, and a TMS320E15 microcontroller manages VME bus communications. The system is highly configurable, allowing for 1 to 8 independent CMAC networks, with selectable input/output sizes and a configurable number of overlapping receptive fields.

Datasets

Not discussed in the paper. The paper focuses on the hardware implementation and its performance characteristics rather than training on specific datasets.

Key Insights and Ideas

  • Hardware Acceleration for CMAC: The primary insight is that dedicated hardware implementation can dramatically improve CMAC performance, making it viable for real-time applications where software implementations are too slow.
  • Configurability for Versatility: The design emphasizes flexibility, allowing users to configure the number of independent CMAC networks, input/output dimensions, and receptive field overlap. This makes the hardware adaptable to a wide range of problems.
  • Pipelined and Recursive Mapping: The use of pipelined logic and a recursive mapping scheme for address generation is a key technical innovation that contributes to the system's speed.
  • Parallel Output Channels: Providing multiple parallel output channels in the accumulator device further enhances the system's throughput for applications requiring multiple outputs.

Key Results

The paper demonstrates that the implemented CMAC system achieves very fast response times:

  • Response Times: For typical CMAC networks with 1 million adjustable weights, 32 integer inputs, and 8 integer outputs, response times (both for recall and training) are consistently well below 1 millisecond.
  • Scalability: The performance remains high even with a varying number of overlapping receptive fields (generalization parameter) between 8 and 256.
  • System Capacity: Each board can implement up to 8 independent CMAC networks with a total of one million adjustable weights. Each network can be configured with 1 to 512 integer inputs and 1 to 8 integer outputs.

Limitations and Future Work

  • VME Bus Dependency: The current implementation is specific to the VME bus. The authors mention that a commercial version for the PC-AT bus is under development, suggesting a desire for broader platform compatibility.
  • Commercialization: The VME bus design is noted as "currently under development," implying it's not yet a fully commercialized product.
  • Future Work: While not explicitly detailed, the success of this high-speed implementation naturally points towards its application in more complex real-time control and pattern recognition tasks, as well as further optimization and integration into larger robotic systems.

Summary updated

BibTeX
@inproceedings{1990design,
  title = {{Design and Implementation of a High Speed CMAC Neural Network Using Programmable CMOS Logic Cell Arrays}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure