EfficientNet and ImageNet

-


                                    @image taken from original research paper. 


Hello, i am back with another article related to CNN's

The article deals with a new method of improving the efficiency of CNN's by using "Compound scaling method", in old models scaling was done, random without a specific patter, but using CSM, the model achieved better accuracy.

But what is compound scaling, well if we wanna use 2 to power n computational resource, then we can increase the network depth, width and image size by alpha to power n, betta to power N and image size by gamma to power N, where alpha, beta, and gamma are constant coefficient determined by small grid search on original small model.

Result: Compound Scaling Method are 2.5 % better than solely scaling Method

parallel to that i would like to tell u about ImageNet Competition, which was help from 2011 to 2017, and ConvNet, GoogleNet, SENet, were a few winners of the competition, the task in those competition was to localize objects, the higest accuracy so far achieved was 87 percent.

Research Paper Link: https://arxiv.org/abs/1905.11946

Comments

Post a Comment

Popular posts from this blog

WORLD MODELS!

-
World models are GAN Models of popular reinforcement learning environments. How are they used?? well the are first trained in unsupervised manner to learn representations of our training environment and then fed into a agent. The agent is actually trained on a fixed policy now. The agent is also able to learn in an hallucinogenic environment and then transfers that learning to real world. To have  a demonstration visit https://worldmodels.github.io reference: https://arxiv.org/abs/1803.10122 read this research article by David Ha.
Read more