Towards efficient and reliable artificial intelligence through neuromorphic principles

Area of Science:

• Artificial Intelligence
• Neuromorphic Engineering
• Sustainable Computing

Background:

• Current AI relies on large neural networks trained on GPUs, leading to high costs and energy use.
• This hardware-centric approach risks favoring algorithms suited to current hardware over inherently superior ones.
• Existing AI models often lack reliability, failing to quantify uncertainty and producing confident incorrect outputs.

Purpose of the Study:

• To propose a shift from current AI paradigms towards more efficient and reliable systems.
• To outline key neuromorphic engineering principles for designing future AI.
• To explore how brain-inspired computing can address limitations in current AI.

Main Methods:

• Discussing six core neuromorphic principles: stateful recurrent models, extreme dynamic sparsity, backpropagation-free learning, probabilistic decision-making, in-memory computing, and hardware-software co-design.
• Surveying relevant prior research in each principle area.
• Identifying future research directions.

Main Results:

• Identification of six key neuromorphic principles applicable to AI algorithms, architectures, and hardware.
• Potential for these principles to guide the development of more efficient, reliable, and sustainable AI systems.
• Highlighting the synergy between neuromorphic engineering and AI advancement.

Conclusions:

• Achieving efficient and reliable AI necessitates embracing neuromorphic principles.
• Brain-inspired computing offers a path to overcome limitations of current AI scaling.
• Future AI development should integrate algorithmic, architectural, and hardware co-design informed by neuroscience.