GABE: Groupwise Affine Basis Encoding

Key Points

• The aim is to introduce GABE, a method for efficiently encoding neural network weights.
• Decomposes neural network weights into three components: shared mean weight, low-rank basis, and per-layer coefficients.
• Utilizes a groupwise approach for similar layers in the neural network.
• Identifies three interpretable components for weight representation.
• Demonstrates that this method serves as an addressable memory system for neural networks.