[Paper Review] Bottom-up and top-down approaches for the design of neuromorphic processing systems: Tradeoffs and synergies between natural and artificial intelligence

Charlotte Frenkel, David Bol+1 more|arXiv (Cornell University)|Jun 2, 2021

Advanced Memory and Neural Computing5 citations

TL;DR

This paper compares bottom-up (biologically inspired) and top-down (application-driven) design approaches for neuromorphic systems, analyzing tradeoffs in circuit design, learning paradigms, and system efficiency. It identifies synergies between natural and artificial intelligence, proposing a framework for neuromorphic intelligence with on-chip learning and event-based processing for low-power edge computing.

ABSTRACT

While Moore's law has driven exponential computing power expectations, its nearing end calls for new avenues for improving the overall system performance. One of these avenues is the exploration of alternative brain-inspired computing architectures that aim at achieving the flexibility and computational efficiency of biological neural processing systems. Within this context, neuromorphic engineering represents a paradigm shift in computing based on the implementation of spiking neural network architectures in which processing and memory are tightly co-located. In this paper, we provide a comprehensive overview of the field, highlighting the different levels of granularity at which this paradigm shift is realized and comparing design approaches that focus on replicating natural intelligence (bottom-up) versus those that aim at solving practical artificial intelligence applications (top-down). First, we present the analog, mixed-signal and digital circuit design styles, identifying the boundary between processing and memory through time multiplexing, in-memory computation, and novel devices. Then, we highlight the key tradeoffs for each of the bottom-up and top-down design approaches, survey their silicon implementations, and carry out detailed comparative analyses to extract design guidelines. Finally, we identify necessary synergies and missing elements required to achieve a competitive advantage for neuromorphic systems over conventional machine-learning accelerators in edge computing applications, and outline the key ingredients for a framework toward neuromorphic intelligence.

Motivation & Objective

• To analyze the tradeoffs and synergies between bottom-up (biologically inspired) and top-down (application-driven) design approaches in neuromorphic systems.
• To identify key challenges in achieving efficient, versatile, and adaptive neuromorphic systems for edge computing.
• To survey silicon implementations across analog, mixed-signal, and digital neuromorphic circuits.
• To establish design guidelines for integrating synaptic plasticity and on-chip learning in neuromorphic hardware.
• To propose a framework for neuromorphic intelligence that bridges biological and artificial intelligence

Proposed method

• Conducts a comparative analysis of neuromorphic system design across three circuit styles: analog, mixed-signal, and digital, focusing on processing-memory co-location.
• Evaluates bottom-up approaches based on biological neural architectures and top-down approaches targeting practical AI applications.
• Examines time-multiplexing, in-memory computation, and novel devices as mechanisms to blur the boundary between processing and memory.
• Reviews silicon implementations of spiking neural networks with on-chip online learning and synaptic plasticity.
• Analyzes key performance metrics such as energy efficiency, inference speed, and adaptability across different neuromorphic platforms.
• Proposes a framework for neuromorphic intelligence integrating event-based processing, low-power operation, and real-time adaptation

Experimental results

Research questions

• RQ1 What are the key tradeoffs between bottom-up and top-down design methodologies in neuromorphic systems?
• RQ2 How do different circuit design styles (analog, mixed-signal, digital) impact the co-location of processing and memory in neuromorphic chips?
• RQ3 What role does on-chip learning play in enabling adaptability and privacy in edge neuromorphic applications?
• RQ4 How can neuromorphic systems close the efficiency and versatility gap compared to conventional AI accelerators?
• RQ5 What synergies between natural and artificial intelligence are necessary to achieve competitive neuromorphic intelligence?

Key findings

• Neuromorphic systems based on spiking neural networks achieve significantly higher energy efficiency than conventional deep learning accelerators, with the human brain operating at ~20 W versus AlphaGo Zero’s ~1 kW.
• Bottom-up approaches enable tighter integration of plasticity and event-based processing, supporting real-time adaptation and low-latency inference.
• Top-down approaches demonstrate strong performance in narrow AI tasks such as image recognition and keyword spotting, but require extensive retraining for new tasks.
• End-to-end spike-based processing of speech and biosignals (e.g., EEG, ECG) is feasible with platforms like HICANN-X and ReckOn, avoiding computationally expensive off-chip preprocessing.
• On-chip learning is essential for autonomous adaptation in wearable and robotic systems, reducing reliance on centralized computing and preserving privacy.
• The NeuroBench initiative is emerging as a critical benchmark suite to enable fair, standardized comparisons across heterogeneous neuromorphic hardware and software platforms.