Research/Blog
Meeting Minutes from Saturday 11th. May – AI Lab Sessions in Bangalore & Gurgaon
- May 13, 2019
- Posted by: CellStrat Editor
- Category: Uncategorized
CellStratAILab #disrupt4.0 #WeCreateAISuperstars
The CellStrat AI Lab met last Saturday in Gurgaon and some fabulous sessions were presented by Pratima, Vinay & Vishal
Vishal started the session talking about AI, its various use cases and how the world is adopting it, what salaries can be expected and how the big companies of the world are on hiring spree in this amazing domain. Some new areas of working came into picture (like AI activist or Drone Pilots etc.) while discussions happened over these in the session.
Pratima followed with presenting a session on Foundation topics in Mathematics for people who would like to get into Machine Learning. It was highly appreciated by all as Maths is one area feared by all but her explainations made the concepts very easy to understand.
Vinay explained the difficult deep learning topic of NLP or the Natural Language Processing giving many relatable examples in real life. He also demoed it through a coding example where when machine is fed some words, what could be possible output from the machine. He discussed the Word2Vec model in detail.
![](http://www.cellstrat.com/wp-content/uploads/2019/05/AILabGgn_110519-Collage-1024x484.png)
In Bangalore fabulous sessions were presented by Anupam, Merril and Sathvik.
First Anupam (team lead for NLP group) presented a fantastic introduction to BERT (Bidirectional Encoder Representations from Transformers) – a new NLP model published by Google. BERT has pre-trained language representations which have been trained with unsupervised bi-directional algorithms, and it has been trained on general-purpose text corpus of Wikipedia. NLP models such as Word2Vec or GloVe are “context-free” models which means the word “bank” means the same think in “bank deposit” and “river bank”. BERT, on the other hand, is context-sensitive and each word representation depends on other words in the sentence. BERT is also bi-directional which means that it learns the sentence context in forward as well as reverse mode. BERT provides enables common NLP tasks such as Question-Answering, sentence classification, next sentence prediction etc.
After this came a superb presentation on Masked RCNNs by Merril. Merril started with introducing image segmentation which can be of these types – threshold based, clustering-based, edge-based and region-based. Image Segmentation finds industrial uses in medical imaging tasks, satellite image analysis, biometrics etc. When it comes to instance segmentation, Masked RCNNs are probably the only algorithm available for this particular task. Here two or more objects of the same class in an image are marked with separate masks. E.g. two dogs in an image will be masked with different pixel level color codes. Instance Segmentation combines object detection (which classifies and creates bounding boxes around objects in an image) with semantic segmentation on each bounding box object, in order to do pixel level classification within each bounding box.
Masked RCNNs extend Faster RCNNs (which do object detection and localization) with a branch to predict the object mask. These use ROIAlign algorithm for Region of Interest detection while masking the localized objects.
Then came a powerful presentation by Sathvik on image captioning which is based on a paper titled “Show, Attend and Tell: Image Captioning with Visual Attention”. This model uses a complex set of deep neural networks involving CNN for feature detection in images and LSTM for caption prediction, with the caveat that the LSTM relies on Visual Attention vector developed with image features. This is a very sophisticated model with some intricate maths behind it and it helps develop high quality image captions.
Our AI Lab members and ML students are working on cutting-edge AI projects and their combined expertise continues to rise week after week.
![](http://www.cellstrat.com/wp-content/uploads/2019/05/AILabB_110519-Collage.png)
AILab Pictures from Bangalore Session