Sakana AI achieved 96.7% accuracy on MNIST and 61.7% on CIFAR-10, using a biological learning method that bypasses backpropagation, the algorithm underpinning most modern AI (TechTimes). This approach trains convolutional networks without backpropagation's computational overhead or biological implausibility. While backpropagation has been deep learning's bedrock for decades, Sakana AI demonstrates competitive results with a fundamentally different, biologically plausible mechanism. This challenges established AI development. Consequently, AI research may pivot towards alternative, brain-inspired learning rules, potentially yielding more robust, efficient, and generalizable systems.
How Sakana AI's Brain-Like Learning Works
Sakana AI employs an Error Diffusion method to train Dale-Compliant Dual-Stream Networks (Marktechpost). Concurrently, its Continuous Thought Machine (CTM) leverages neuron dynamics synchronization for task solving (Sakana Ai). These mechanisms enable neural networks to learn via local rules, mimicking biological brains without global error signals. Sakana AI's benchmark performance without backpropagation suggests gradient descent is an architectural choice, not a fundamental requirement, opening avenues for energy-efficient and biologically plausible AI.
Architectural Innovations for Biological Plausibility
Sakana AI's dual-stream architecture maintains non-negative weights while enabling inhibitory effects (TechTimes). This design provides a nuanced simulation of biological neural activity. Furthermore, the Continuous Thought Machine (CTM) allows neurons to access their own behavioral history to compute outputs (Sakana Ai). This self-history mechanism supports complex, internal dynamic processing. Together, these architectural elements are crucial for mimicking biological neural processes, fostering complex interactions while ensuring network stability.
