Associated Learning: Decomposing End-to-end Backpropagation based on Auto-encoders and Target Propagation
This addresses a scalability bottleneck in deep learning training for researchers and practitioners, though it is incremental as it builds on existing auto-encoder and target propagation ideas.
The paper tackles the inefficiency of backpropagation due to backward locking by introducing associated learning (AL), which modularizes networks into independent components to enable simultaneous weight updates, reducing training time complexity from O(nl) to O(n + l) while achieving comparable accuracy to BP.
Backpropagation (BP) is the cornerstone of today's deep learning algorithms, but it is inefficient partially because of backward locking, which means updating the weights of one layer locks the weight updates in the other layers. Consequently, it is challenging to apply parallel computing or a pipeline structure to update the weights in different layers simultaneously. In this paper, we introduce a novel learning structure called associated learning (AL), which modularizes the network into smaller components, each of which has a local objective. Because the objectives are mutually independent, AL can learn the parameters in different layers independently and simultaneously, so it is feasible to apply a pipeline structure to improve the training throughput. Specifically, this pipeline structure improves the complexity of the training time from O(nl), which is the time complexity when using BP and stochastic gradient descent (SGD) for training, to O(n + l), where n is the number of training instances and l is the number of hidden layers. Surprisingly, even though most of the parameters in AL do not directly interact with the target variable, training deep models by this method yields accuracies comparable to those from models trained using typical BP methods, in which all parameters are used to predict the target variable. Consequently, because of the scalability and the predictive power demonstrated in the experiments, AL deserves further study to determine the better hyperparameter settings, such as activation function selection, learning rate scheduling, and weight initialization, to accumulate experience, as we have done over the years with the typical BP method. Additionally, perhaps our design can also inspire new network designs for deep learning. Our implementation is available at https://github.com/SamYWK/Associated_Learning.