Research/Blog
Meeting Minutes from Saturday 27th. Apr AI Lab Session in Bangalore
- April 29, 2019
- Posted by: CellStrat Editor
- Category: Uncategorized
#CellStratAILab #disrupt4.0 #WeCreateAISuperstars
We had another fabulous AI Lab meetup full of intense demos and discussions.
The day started with Shubha presenting a superb presentation on time-series modeling for stock market predictions. She referred to traditional statistical models for time series such as ARIMA, ARMA and SARIMA as well as modern Deep Learning RNNs / LSTMs which are currently used instead of traditional methods. Shubha followed this up with a deep-dive on the algorithmic details of time-series modeling. She discussed how we can create a model which takes the previous time step (or time steps) output and learns to predict the next time-step value. Shubha explained how previous time step values are fed to future time step calculation in form of additional features in order to account for sequential dependency. She then showed a time-series code demo that included predicting Stock Prices for 150 stocks with 46 Anonymized features at 5 min intervals.
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-12.png)
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-13.png)
Shubha also referred to two more advanced papers, one where one can predict stock market based on NLP on news feeds, and the second one a Uber use case to predict driver demand in a certain area based on features like extreme events (such as weather anomalies). The latter uses an LSTM Autoencoder to detect latent features (Feature Extractor) and feeds the latent values learnt to a regular LSTM Decoder (Forecaster) for final driver demand prediction.
After this, Ramapriya discussed the Netflix recommendation engine problem and how he was able to train a model on this dataset. He reviewed the Data Science Project Methodology, which is a best practices scheme for running Data Science projects. Ramapriya discussed the results of Netflix collaborative filtering model to predict user ratings based on historical 3 months / 6 months / 1 year / all user rating data.
Then came Harshit’s excellent presentation on how to train OpenAI Gym Racing Car model using Deep Q Learning. He explained basics of RL including policy optimization, Bellman Equation, Value Iteration algorithm and Q Learning as well as Epsilon Greedy policy to exploit Exploration vs Exploitation concepts in RL. Q Learning involves updating the Q Table based on an iterative process of state-action trials. After explaining the basics, Harshit walked through a detailed code demo on how to build an RL model for Car Racing in OpenAI Gym using Q learning with a Greedy policy setting.
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-10.png)
![](http://www.cellstrat.com/wp-content/uploads/2019/04/image-11.png)
The sophistication of AI Lab experts continues to rise week after week. We are surely on way to becoming one of the world’s most powerful AI research labs.
![](http://www.cellstrat.com/wp-content/uploads/2019/04/AILabB_270419_Collage.png)