Research/Blog
CELLSTRAT AI LAB – MINUTES FROM 2ND. MAR 2019 BLR & GGN MEETUPS
- March 4, 2019
- Posted by: CellStrat Editor
- Category: Artificial Intelligence Computer Vision Deep Learning Machine Learning Natural Language Processing
The AI Lab groups continues to get bigger by the week with increasing number of mid-tier professionals willing to learn, experience and practice deep learning skills. Interest rising both among technology developers and leadership level professionals in wide range of industry domains right from technology services companies to banking to manufacturing and automotive companies. Fantastic sessions were presented by some of the AILab members last Saturday followed by discussions on few concepts where members were facing difficulty.
In AI Lab session in Gurgaon, we discussed:
- Various use cases for data preparation in various industries.
- Need for data analysis for SMEs in manufacturing sector who can’t afford engaging expensive ML resources.
- Hyperparameters.
- Recurrent Neural Networks.
In AI Lab session in Bangalore, following presentations took place:
Shreyas presented Path Planning with RL, Anshumaan presented extensive discussion on Image Segmentation using FCNs, U-nets, CNNs etc. Half a dozen people joined over the Hangout including the Gurgaon AI Lab meetup group which was meeting parallelly as well. The drive for intense AI-led disruption continues non-stop. Deeper work and research is forthcoming.
Shreyas presented “Path Planning Algorithm” with Reinforcement Learning. He stated — Many years have passed since the first path planning algorithm was found. These algorithms have evolved a lot since then. Even the best of the path finding algorithms were incapable to give the ideal result which was needed to find the path. Especially in terrains where the algorithms couldn’t overcome the obstacle because that situation wasn’t predicted beforehand. The ideal path was found using trial and error methods after the unpredicted events led to the collision of the robot. This problem was resolved when the robots gained the ability to outperform itself in these situations by learning what went wrong on its own and making sure the exact same event doesn’t occur again. These Reinforcement Learning algorithms emerged in the past decade and were continuously improvised and were perfected to have a very high success rate.
![](http://www.cellstrat.com/wp-content/uploads/2019/03/020319_Tech-Slides_Collage-1024x508.png)
Anshumaan explained the “Semantic image segmentation: using FCNs, UNet and cutting-edge architectures like DeepLabv3+.
The deep lab architecture was explained in detail with concepts like Atrous Spatial Pyramid Pooling and encoder decoder networks and their respective significance in Semantic segmentation.
![](http://www.cellstrat.com/wp-content/uploads/2019/03/020319_Tech-Slides_DeepLab-Collage-1024x266.png)
![](http://www.cellstrat.com/wp-content/uploads/2019/03/AILab_020319-Collage.png)