Meeting Minutes from Saturday 4th. May – AI Lab Session in Bangalore
- May 6, 2019
- Posted by: CellStrat Editor
- Category: Uncategorized
CellStratAILab #disrupt4.0 #WeCreateAISuperstars
The CellStrat AI Lab met last Saturday in BLR and some fabulous sessions were presented by Shreyas, Abdus and Anshumaan.
Shreyas started with a demo of the driverless cars technology whereby the autonomous system deploys image segmentation techniques for scene analysis and obstacle detection. Shreyas discussed world’s first Indian roads dataset published by Intel and IIIT Hyderabad. He mentioned use of ESPNet, a state of the art image segmentation technique that provides better accuracy on roads dataset compared to the previous techniques.
Then came an intuitive and rich presentation by Abdus who explained the Inception Module and it’s design in detail. Abdus discussed how an Inception Network can provide better accuracy by doing parallel convolutions with 1×1, 3×3 and 5×5 kernels in the same layer and then concatenating them in the next layer, in order to preserve spatial intelligence at different scales of image resolution. The Inception module also uses a bottleneck 1×1 filter before 3×3 and 5×5 kernels to reduce the total number of computations – this helps mitigate the performance cost of the Inception Module.
GoogLeNet (Inception V1) and the latest Inception-ResNet are examples of network architectures that use Inception Module for efficient image processing. Inception Network has been one of the best performing architectures for the ImageNet dataset.
Finally came a superb presentation by Anshumaan on MorphNet, a neural network design solution published by Google recently that helps develop faster and smaller neural networks. This solution offers techniques to automate the design of efficient neural networks. MorphNet designs a neural network by alternating between a cycle of shrinking and expanding networks. The shrinking happens with aid of a sparsifying regularizer and expansion happens via a width multiplier that gets applied to all layers. This has the effect of removing insignificant neurons and allocating more resources to more important neurons, while respecting the resource constraints (such as FLOPs or floating-point operations). When applied to standard networks such as ImageNet, Inception V2, AudioSet or ResNet, the MorphNet discovers novel structures in each case while improving the network performance by a good margin.
Overall it was a rich Saturday AI Lab meetup. The sophistication of our demos and presentations continues to increase dramatically week after week.