Research/Blog
Meeting Minutes from Saturday 20th. Apr AI Lab Session in Bangalore
- April 22, 2019
- Posted by: CellStrat Editor
- Category: Artificial Intelligence Computer Vision Deep Learning Machine Learning Reinforcement Learning
#CellStratAILab #disrupt4.0 #WeCreateAISuperstars
CellStrat AI Lab had an intense meetup last Saturday in Bangalore.
The day started with an intense presentation by Shreyas (team lead – RL, Driverless Cars), who presented ESPNet (Efficient Spatial Pyramid for Image Segmentation https://arxiv.org/abs/1803.06815) – a fast and efficient CNN particularly useful for semantic segmentation of high resolution images under resource constraints. ESPNet uses a new module ESP, which is 22 times faster and 180 times smaller than state of the art image segmentation algorithms. A demo on Cityscapes driverless dataset shows improved performance on many different criteria.
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-9.png)
Next came a superb presentation on Masked RCNNs (used for instance segmentation applications) prepared by Merril and presented expertly by Anshumaan (Team Lead – Vision group). Anshumaan discussed the differences between ROI pooling and ROIAlign; the latter is used for instance segmentation. Masked RCNNs are the only known algorithm for instance segmentation use case.
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-6.png)
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-7.png)
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-8.png)
Finally, Pushparaj presented an amazing session on how a signal tranformation from time-step mode to frequency domain can make the performance of CNNs much faster. A convolution in time space becomes single element-wise multiplication in frequency space. This data tranformation is achieved via Fast Fourier Transform (FFT) and ends up in data format involving frequencies, from which one can remove noise frequencies easily leading to efficient data space. Then an inverse FFT recovers the data back to time-space format. Overall, this makes CNNs significantly faster and efficient.
![](http://www.cellstrat.com/wp-content/uploads/2019/04/AILabB_200419_Collage-1.png)