Research/Blog
CellStrat AI Lab – Minutes From 30th. Mar, 2019 Bangalore Meetup
- April 1, 2019
- Posted by: CellStrat Editor
- Category: Artificial Intelligence Computer Vision Deep Learning Machine Learning
CellStratAILab #disrupt4.0 #WeCreateAISuperstars
AI Lab is the new name for global disruption. We have an increasing group of Lab researchers and learners presenting their research and presentations on advanced AI algorithms.
We had a fabulous Lab meetup on Saturday 30th. Mar., in Bengaluru with members travelling from far off locations as well as other remote locations attending via Hangout.
Among the sessions presented, Ramapriya had a superb session on building Recommender Systems using Model Based Approach, in particular Deep AutoEncoder Networks. This is a research paper by Nvidia titled “Training Deep AutoEncoders for Collaborative Filtering” (https://arxiv.org/abs/1708.01715).
Recommenders can be built with group-based models such as similarity matches based on item-item or item-user groups. Or they can be built with model-based methods. The latter include AutoEncoders which this paper proposes. Here both Encoder and Decoder are feed forward neural networks. Taking example of Movie Ratings from Netflix dataset, idea is to :-
1) predict the ratings for non-rated movies (Sparse Matrix) in the first pass (this produces dense matrix) using RMSE loss minimization.
2) and then re-feed predictions from first pass above for a second run to improve on predictions.
Pushparaj presented interesting ideas in Machine Vision, which involve detecting defects of various nature on the manufacturing assembly line. He discussed traditional Machine Vision techniques which help us detect defects in manufactured objects or material such as glass, metal, sheet, clothing etc.
Sathvik presented a fabulous paper on Double Attention Networks (https://arxiv.org/abs/1810.11579). These are called A2 Nets which allow double attention mechanism in CNNs in order to capture long range spatial correlations in images. The way they do this is that one multiplies the feature maps in a prior CNN layer with those from a later CNN layer and replace the previous layer with this new transformed set of feature maps. This has the effect of capturing the correlation from higher order features in lower level layers thus ensuring better accuracy across long-range pixel correlations.
This paper on Double Attention Networks represents state-of-the-art research as it was presented at NIPS 2018 just a few months back.