Brain-Inspired AI Designs Promise 50 Per Cent Efficiency Boost

Researchers are increasingly recognising the potential of merging neuroscience and artificial intelligence to create more efficient and powerful computing systems. Shanmuga Venkatachalam, Prabhu Vellaisamy, and Harideep Nair, from Carnegie Mellon University, alongside Wei-Che Huang, Youngseok Na, and Yuyang Kang et al., present a novel approach with NeuroAI Temporal Neural Networks (NeuTNNs). This work details a new microarchitecture and design framework that draws directly from biological principles, specifically, neuron models with active dendrites, to enhance both capability and hardware efficiency. By introducing NeuTNNGen, a tool suite translating PyTorch models into application-specific NeuTNN layouts, the team demonstrates significant performance gains and a reduction in synaptic costs of up to 50%, paving the way for a new generation of energy-efficient NeuroAI systems.

Biologically inspired NeuTNNs and automated design with NeuTNNGen offer promising avenues for next-generation neural networks

Scientists are pioneering a new era in artificial intelligence by directly integrating insights from neuroscience, a field they term NeuroAI. Central to this breakthrough is NeuTNNGen, a PyTorch-to-layout tool suite enabling the creation of application-specific NeuTNNs, streamlining the design process and unlocking new levels of performance.
Compared to existing temporal neural network designs, NeuTNNs demonstrate superior performance and efficiency across a range of applications. Researchers validated NeuTNNGen’s capabilities through three key examples: UCR time series benchmarks, MNIST design exploration, and the creation of Place Cells for neocortical reference frames.

These implementations showcase the versatility of the new architecture in handling diverse data types and computational tasks. Furthermore, the study explores synaptic pruning techniques, achieving a 30-50% reduction in synapse counts and associated hardware costs without compromising model precision across various sensory modalities.
This research represents a significant step towards building brain-like computing systems with dramatically improved energy efficiency. NeuTNNGen facilitates the design of specialized, energy-efficient NeuTNNs, paving the way for the next generation of NeuroAI systems. By embracing principles from cortical columns and reference frames, the team has created a microarchitecture that leverages active dendrites for more powerful and nuanced computation. The study meticulously constructs NeuTNNs through a six-layer hierarchical abstraction, beginning with synapses and culminating in cortical macrocolumns, significantly expanding upon the four layers present in previous TNN designs.

Each NeuTNN neuron incorporates active dendrites containing both distal and proximal segments, functionally equivalent to TNN columns, allowing for contextual clustering and classification. NeuTNNGen, a PyTorch-to-layout tool suite, facilitates the design of application-specific NeuTNNs, translating software models into hardware implementations.

This framework supports rapid design space exploration, identifying optimal models for diverse applications, and then leverages commercial EDA tools to assess post-synthesis and post-place-and-route Power-Performance-Area (PPA) metrics. Post-layout PPA results were reported for 45nm and predictive 7nm CMOS technologies, evaluating time-series clustering, MNIST, and Place Cells designs utilising the TNN7 library for optimised implementations.

The experimental setup involved three key applications: UCR time series benchmarks, MNIST design exploration, and Place Cells modelling for neocortical reference frames. Synaptic pruning was also explored, achieving a 30-50% reduction in synapse counts and associated hardware costs while preserving model precision across various sensory modalities.

NeuTNN minicolumns, serving as foundational building blocks, were stacked to form layers and cascaded to create multilayer networks, potentially replacing multilayer TNNs with shallower, more energy-efficient architectures. NeuTNN segments operate similarly to TNN point neurons, but with varied response characteristics, integrating signals to influence somatic potential and enabling hierarchical processing.

Expanded NeuTNN architecture and efficient hardware implementation via NeuTNNGen enable low-latency inference

NeuTNNs, a new category of temporal neural networks, achieve a six-layer abstraction hierarchy extending beyond previous TNN designs which featured only four layers. This expanded architecture incorporates active dendrites and a hierarchy of distal and proximal segments, mirroring biological neocortical structures.

NeuTNN neurons demonstrate significantly richer functionality compared to classic point neurons, comprising active dendrites each containing multiple segment types. Each NeuTNN segment operates equivalently to a TNN point neuron, while a NeuTNN active dendrite is computationally comparable to a TNN column.

The research introduces NeuTNNGen, a PyTorch-to-layout tool suite designed for application-specific NeuTNNs, supporting fully configurable multilayered networks and automated compatibility checks. Synaptic pruning, explored within NeuTNNGen, reduces synapse counts and associated hardware costs by 30-50% while maintaining model precision across diverse sensory modalities.

Implementation of reference frames (RFs) in CMOS with spiking NeuTNNs has been demonstrated for the first time, reporting corresponding hardware results. Post-layout power-performance-area (PPA) results were obtained for 45nm and predictive 7nm CMOS technologies across several benchmarks. Time-series clustering, expanding on previous work, alongside MNIST and Place Cells in RF, were used to evaluate performance.

Leveraging the TNN7 library, optimized NeuTNN implementations were realised, demonstrating the framework’s capabilities. These networks incorporate findings from neuroscience, specifically a neuron model featuring active dendrites and a hierarchical structure of distal and proximal segments, to enhance both capability and hardware efficiency.

The team developed NeuTNNGen, a PyTorch-to-layout tool suite, enabling the design of application-specific NeuTNNs for diverse tasks. Demonstrations using UCR time series benchmarks, MNIST design exploration, and Place Cells for neocortical reference frames showcase NeuTNNGen’s effectiveness. Furthermore, the application of synaptic pruning reduced synapse counts by 30-50% without significantly impacting model precision across different sensory modalities.

This suggests a pathway towards creating energy-efficient neuromorphic processing units. The authors acknowledge that the current work focuses on specific applications and that further research is needed to explore the full potential of NeuTNNs across a wider range of problems. Future work will likely involve expanding the tool suite and investigating more complex network architectures, potentially leading to more biologically realistic and efficient AI systems.