Brain-inspired Predictive Coding Improves the Performance of Machine Challenging Tasks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Backpropagation is the standard for training artificial neural networks but faces criticism for biological implausibility.
• Backpropagation exhibits limitations in specific machine-challenging tasks (MCTs), indicating room for performance improvement.
• The human brain's learning mechanisms offer a potential alternative for enhancing machine intelligence.

Purpose of the Study:

• To investigate if a brain-mimicking learning mechanism can improve machine learning performance on MCTs.
• To evaluate the effectiveness of predictive coding, a biologically plausible algorithm, as an alternative to backpropagation.
• To address limitations in incremental learning, long-tailed recognition, and few-shot recognition using predictive coding.

Main Methods:

• Employed predictive coding, a biologically plausible learning algorithm, for training artificial neural networks.
• Conducted experiments on representative MCTs: incremental learning, long-tailed recognition, and few-shot recognition.
• Compared the performance of predictive coding-trained networks against backpropagation-trained networks.

Main Results:

• Predictive coding robustly outperformed backpropagation across all tested MCTs.
• Predictive coding-based incremental learning effectively alleviated catastrophic forgetting.
• Predictive coding mitigated classification bias in long-tailed recognition and enabled accurate few-shot recognition.

Conclusions:

• A learning mechanism mimicking the human brain (predictive coding) enhances machine learning performance on challenging tasks.
• Predictive coding offers a more biologically plausible and effective alternative to backpropagation for specific AI applications.
• The study highlights the potential of predictive coding for advancing general machine learning capabilities.