Train-by-Reconnect: Decoupling Locations of Weights from their Values
This work addresses a foundational problem in deep learning by decoupling weight locations from values, potentially offering new training paradigms for neural networks.
The paper tackles the problem of understanding what distinguishes trained deep neural networks from untrained ones, proposing that weight locations hold most of the information. It introduces Lookahead Permutation (LaPerm), a method that reconnects weights to achieve performance similar to or better than Adam in various initialization scenarios, including sparse and single-value cases.
What makes untrained deep neural networks (DNNs) different from the trained performant ones? By zooming into the weights in well-trained DNNs, we found it is the location of weights that hold most of the information encoded by the training. Motivated by this observation, we hypothesize that weights in stochastic gradient-based method trained DNNs can be separated into two dimensions: the locations of weights and their exact values. To assess our hypothesis, we propose a novel method named Lookahead Permutation (LaPerm) to train DNNs by reconnecting the weights. We empirically demonstrate the versatility of LaPerm while producing extensive evidence to support our hypothesis: when the initial weights are random and dense, our method demonstrates speed and performance similar to or better than that of regular optimizers, e.g., Adam; when the initial weights are random and sparse (many zeros), our method changes the way neurons connect and reach accuracy comparable to that of a well-trained fully initialized network; when the initial weights share a single value, our method finds weight agnostic neural network with far better-than-chance accuracy.