liveBook · Manning

AmoebaNet-B

As explained in Sutskever's List

AmoebaNet-B is a convolutional neural network architecture in the AmoebaNet family for visual recognition.

AmoebaNet-B is a convolutional neural network for visual recognition. Trained with GPipe across four accelerators in four stages, its 557 million parameters reached 84.4% top-1 accuracy on ImageNet-2012, then a single-model benchmark record. section 6.3.2

Its size also supported transfer learning: the ImageNet-trained model matched or exceeded results across several vision datasets. Convolutional layers had uneven computational costs, however, making the model harder to partition efficiently; on eight GPUs, the observed speedup was 3.5-fold. section 6.3.2

CIFAR-100 transfer

Fine-tuning the ImageNet-trained model reduced the CIFAR-100 error rate from 10.7% to 8.7%. section 6.3.2

How Sutskever's List covers it

How deep it goes
Reports model scale, pipeline setup, benchmark accuracy, and transfer results, but does not describe AmoebaNet-B’s internal architecture.
Examples
Uses ImageNet-2012 accuracy and fine-tuning on CIFAR-100 and other vision datasets as evidence.
What you can do after
Evaluate how pipeline parallelism scales a convolutional vision model and enables transfer results from ImageNet pretraining.
How it connects
Connects AmoebaNet-B to convolutional neural networks, GPipe, ImageNet pretraining, and transfer learning.
Cover of Sutskever's List
liveBookmanning.com
Read Sutskever's List on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Sutskever's List. The section links point to the parts of the book it draws on.