March 11, 2019 feature
A new approach to overcome multi-model forgetting in deep neural networks
In recent years, researchers have developed deep neural networks that can perform a variety of tasks, including visual recognition and natural language processing (NLP) tasks. Although many of these models achieved remarkable results, they typically only perform well on one particular task due to what is referred to as "catastrophic forgetting."
Essentially, catastrophic forgetting means that when a model that was initially trained on task A is later trained on task B, its performance on task A will significantly decline. In a paper pre-published on arXiv, researchers at Swisscom and EPFL identified a new kind of forgetting and proposed a new approach that could help to overcome it via a statistically justified weight plasticity loss.
"When we first started working on our project, designing neural architectures automatically was computationally expensive and unfeasible for most companies," Yassine Benyahia and Kaicheng Yu, the study's primary investigators, told TechXplore via e-mail. "The original aim of our study was to identify new methods to reduce this expense. When the project started, a paper by Google claimed to have drastically reduced the time and resources required to build neural architectures using a new method called weight-sharing. This made autoML feasible for researchers without huge GPU clusters, encouraging us to study this topic more in depth."
During their research into neural network-based models, Benyahia, Yu and their colleagues noticed a problem with weight sharing. When they trained two models (e.g. A and B) sequentially, model A's performance declined, while model B's performance increased, or vice versa. They showed that this phenomenon, which they called "multi-model forgetting," can hinder the performance of several auto-mL approaches, including Google's efficient neural architecture search (ENAS).
"We realized that weight-sharing was causing models to impact each other negatively, which was causing the architecture search process to be closer to random," Benyahia and Yu explained. "We also had our reserves on architecture search, where only the final results are shed to light and where there is no good framework to evaluate the quality of the architecture search in a fair way. Our approach could help to fix this forgetting problem, as it is related to a core method that nearly all recent autoML papers rely on, and we consider such impact to be huge to the community."
In their study, the researchers modeled multi-model forgetting mathematically and derived a novel loss, called weight plasticity loss. This loss could reduce multi-model forgetting substantially by regularizing the learning of a model's shared parameters according to their importance for previous models.
"Basically, due to the over-parameterization of neural networks, our loss decreases parameters that are 'less important' to the final loss first, and keeps the more important ones unchanged," Benyahia and Yu said. "Model A's performance is thus unaffected, while model B's performance keeps increasing. On small datasets, our model can reduce forgetting up to 99 percent, and on autoML methods, up to 80 percent in the middle of training."
In a series of tests, the researchers demonstrated the effectiveness of their approach for decreasing multi-model forgetting, both in instances where two models are trained sequentially and for neural architecture search. Their findings suggest that adding weight plasticity in neural architecture search can significantly improve the performance of multiple models on both NLP and computer vision tasks.
The study carried out by Benyahia, Yu and their colleagues sheds light on the issue of catastrophic forgetting, particularly that which occurs when multiple models are trained sequentially. After modeling this problem mathematically, the researchers introduced a solution that could overcome it, or at least drastically reduce its impact.
"In multi-model forgetting, our guiding principle was to think in formulas and not just by simple intuition or heuristics," Benyahia and Yu said. "We strongly believe that this 'thinking in formulas' can lead researchers to great discoveries. That is why for further research, we aim to apply this approach to other fields of machine learning. In addition, we plan to adapt our loss to recent state-of-the-art autoML methods to demonstrate its effectiveness in solving the weight-sharing problem observed by us."
Efficient neural architecture search via parameter sharing. arXiv:1802.03268 [cs.LG]. arxiv.org/abs/1802.03268
© 2019 Science X Network