Research/Blog
Meeting Minutes from AI Lab session on Saturday 27th July in Bengaluru
- August 2, 2019
- Posted by: vsinghal
- Category: Computer Vision Deep Learning Natural Language Processing
#CellStratAILab #disrupt4.0 #WeCreateAISuperstars
CellStrat AI Lab had intense Lab meetups last Saturday in Bengaluru at our Bellandur and Hebbal locations.
CellStrat AI Researcher Abdul Azeez spoke on Object Detection using a variety of techniques, starting with a traditional method called Viola/Jones Face Detector. This is a popular method for real time object detection (e.g. used in cameras). Training is slow for this model but inference is very fast. This model has been trained on 5000 faces and 9400 non-face images. It uses AdaBoost classifier with an incremental “Integral Image” concept. After this, Abdul discussed more modern techniques for object detection including CNN classification and localization models, Pascal, ImageNet and COCO datasets, IOU and mAP techniques, TensorFlow-based DL approaches including RCNN, Faster RCNN, YOLO models etc. Overall it was a very extensive and comprehensive discussion on a variety of Object Detection protocols.
At our Bellandur AI Lab, NLP researcher Deepti Gupta presented a fantastic session on Speech To Text using a LibriTTS corpus which involves .wav format audio files sampled at 24KHz rate. The signal is decomposed with Fast Fourier Transform (FFT) into frequency domain, this produces multiple sinusoidal waves with different amplitude and frequencies. These are combined into one sinusoidal sine wave. Deep Learning models like Gaussian Mixture variational autoencoder (GMVAE) and WaveRNN models are used to process the LibriTTS corpus for Text to Speech or Speech to Text models.
Then came a superb presentation on Inception V3 Networks by AI Researcher Abdus Samad. He started with basics of inception network and GoogLeNet. Then he described how the authors of inception V2 upgraded the network from V1. Then he explained upgrades used in Inception V3. In V3, authors used RMSProp, Batch Normalization, grid size reduction technique, and Label smoothing. Inception V3 is a CNN model trained on more than a million images on the ImageNet database. The network is 48 layers deep and can classify the images into 1000 object categories, such as animals, mouse, keyboard etc.
Want to learn advanced AI and ML with help of India’s No 1 AI Research Lab ? Then join us for our Saturday AI Lab tomorrow in Bengaluru :-
Bengaluru AI Lab :-
Register : https://www.meetup.com/Disrupt-4-0/events/jvfhvqyzlbfb/
Topic : 3D CNNs, Object Detection with Faster RCNNs, Ranked Retrieval for Text Analytics
Date : Saturday 3rd Aug 2019, 10:30AM – 5:00PM
Loc. : WeWork, Embassy Tech Village, ORR, BLR
See you TOMORROW (3rd Aug) for the AI Lab meetup! Let’s disrupt the world with AI, together !
Questions ? Call me at +91-9742800566 !
Best Regards,
Vivek Singhal
Co-Founder & Chief Data Scientist, CellStrat
+91-9742800566