Gary Marcus: Deep Learning Alone Won't Get Us to AGI

Summary

Gary Marcus argues in Wired that deep learning, despite its impressive achievements, has fundamental limitations that prevent it from achieving human-level artificial general intelligence. He contends that deep learning systems lack robust reasoning, generalization, and common sense, and that reaching AGI will require integrating additional approaches beyond neural networks. Marcus calls for hybrid systems that combine deep learning with symbolic reasoning and other AI paradigms.

Key Points

• •Deep learning excels at pattern recognition but struggles with systematic generalization, causal reasoning, and transferring knowledge to new domains.
• •Current deep learning systems lack common sense understanding and robust reasoning capabilities necessary for human-like intelligence.
• •Marcus advocates for hybrid AI architectures combining deep learning with symbolic AI and structured knowledge representations.
• •Overreliance on scaling deep learning may lead to diminishing returns without addressing fundamental architectural limitations.
• •Achieving AGI likely requires interdisciplinary insights from cognitive science, linguistics, and neuroscience beyond purely data-driven approaches.