Research/Blog
Meeting Minutes from AI Lab session on Saturday 5th Oct in Bengaluru
- October 7, 2019
- Posted by: vsinghal
- Category: Computer Vision Deep Learning Machine Learning Natural Language Processing
#CellStratAILab #disrupt4.0 #WeCreateAISuperstars
Last Saturday saw some deep sessions on advanced AI solutions at the CellStrat AI Lab meetup.
![](http://www.cellstrat.com/wp-content/uploads/2019/10/Collage-1024x768.jpg)
Sentiment Analysis with Self-Attentive Embedding Model :-
First Bhanumathi presented an interesting session on Sentiment Analysis using a Self-Attentive Sentence Embedding model. This is based on ICLR 2017 paper by IBM Watson and University of Montreal researchers. Instead of a vector, the Sentence Embedding is represented by a matrix where each row focusses on a different part of the sentence. It uses a self-attention mechanism on hidden states of the bi-directional LSTM. This model gives high accuracy on author profiling, sentiment analysis and textual entailment.
![](http://www.cellstrat.com/wp-content/uploads/2019/10/Self-Attentive-Sentence-Embedding.png)
Image Credit : https://arxiv.org/pdf/1703.03130.pdf
Semantic Image Segmentation with Gated Shape CNNs :-
Then came a super presentation by Jani Basha on Semantic Image Segmentation with Gated Shape CNNs (Gated-SCNN). This involves a new two-stream CNN architecture for semantic image segmentation that explicitly wires shape information as a separate processing branch (the regular CNN branch processes usual image features such as shape, color, texture etc.) to improvise the current state of the art results i.e DeepLab v3. The shape stream uses a gated Convolutional Layer architecture and connects the intermediate layers of the two streams. In the end, a fusion module combines the two streams using the Atrous Spatial Pyramid Pooling (ASPP). A special loss function with dualtask regularizer focusses on predicting high-quality boundaries.
![](http://www.cellstrat.com/wp-content/uploads/2019/10/Gated-SCNN.png)
![](http://www.cellstrat.com/wp-content/uploads/2019/10/Gated-SCNN-2.png)
Image Credit : https://arxiv.org/pdf/1907.05740.pdf
Support Vector Machines :-
Finally, Vivek Singhal presented a session on Support Vector Machines (SVMs), which classify datasets based on finding the maximum margin hyperplane between two or more classes. The objective function involves minimizing weights while keeping the classes far apart from the margin. Linear SVMs can detect linear classification boundaries. Whereas non-linear SVMs with polynomial kernel or RBF kernel can detect non-linear classifier boundaries, by adding a polynomial feature or up-dimensionality change (e.g. 2D dataset can be classified when upsampled to 3D transformation) respectively.
Join our AI Lab to experience world-class AI research and development or to train yourself in advanced AL, ML and DL. Attend our AI Lab sessions in Bengaluru (Bellandur or Hebbal) this Saturday (12th Oct 2019) :-
Bellandur AI Lab (Saturday 12th Oct) :-
Register : https://www.meetup.com/Disrupt-4-0/events/263932144/
Topic : Diabetes prediction with ML, Style Transfer in Computer Vision, RL with DQN
Loc. : WeWork, Embassy Tech Village, ORR, BLR
Presenters : Dr Purnendu Sekhar Das, Gouthmanan Asokan, Niraj Kale
Hebbal AI Lab (Saturday 12th Oct) :-
Register : https://www.meetup.com/Disrupt-4-0/events/265383409/
Topic : Fake Video Detection, NLP
Loc. : WeWork, Hebbal, Bellary Road, BLR
Presenters : Pushparaj M., Abdul Azeez
See you this weekend for the AI Lab meetup ! Let’s put India, and Bengaluru, on the global AI map !
Questions ? Call me at +91-9742800566 !
Best Regards,
Vivek Singhal
Co-Founder & Chief Data Scientist, CellStrat
+91-9742800566