![]() In recent research, three methods have emerged as especially important (and interesting) strategies for model compression. Model compression aims to alleviate the costs of large model sizes (like the ones mentioned above) by representing the model in a more efficient format with minimal impact on its performance. Even in the absence of such applications, the costs imposed by a large model size are well-worth navigating around.Īs such, the problem of model compression is important. Indeed, many applications of deep learning are applied to devices in which resources are constrained, like an offline Google Translate app. Computational resources are limited, and a network that requires too much of it will either train too slowly or cannot be stored at all. ![]() ![]() Models with large architectures and complex operations may take a long time to inference over many inputs. Models with vast oceans of parameters to tune take a long time to train, which can be impractical. ![]() With the advent of convolutional neural networks and transformers to handle complex image recognition and natural language processing tasks, deep learning models have skyrocketed in size.Īlthough the increase in size is usually associated with an increase in predictive power, this supersizing comes with undesirable costs. Creative techniques to make complex models smaller ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |