Meeting Minutes from AI Lab session on Saturday 26th Oct in Bengaluru
#CellStratAILab #disrupt4.0 #WeCreateAISuperstars
Last Saturday we had another fantastic set of AI presentations at our AI Lab meetup.
Face Recognition System with MTCNN :-
First, Jani Basha presented a fabulous model on face recognition system, which involved detecting and recognizing faces from images and videos. One can also use this model to identify humans from live webcams. The model is based on MTCNN face detector which is inspired by David Sandberg’s FaceNet model and Arxiv Paper 1604.02878 by Zhang et al. (2016). This model takes care of poses, illuminations and occlusions as well. It involves a cascaded three-stage multi-task deep convolutional network that predict face and landmark correlation in coarse-to-fine manner.
Fake Video Detection :-
Next came an amazing session by Pushparaj M. on Fake Video Detection. This model was based on Arxiv paper 1809.00888 which focuses on detecting videos created with two recent techniques being used for forging faces in videos, viz Deepfake and Face2Face. Normal image forensics are not able to detect such forgeries. This paper proposes two networks which focus on mesoscopic properties of images. These networks did well in detecting forged videos created with Deepfake and Face2face.
Deepfake is a technique which replaces the face with somebody else’s face. It achieves this by training two autoencoders (AEs) on two people’s faces and then mixing up the encoder weights from both AEs. Then the test image is encoded with Encoder of Network A and decoded with Decoder of Network B. This technique is applied to successive image frames of a video to create facial forgery.
Face2Face does a photorealistic transfer of image facial expression from source to target person. The final image synthesis involves overlaying the target face with morphed facial blendshape to fit the source facial expression.
The solution uses mesoscopic analysis (intermediate between microscopic and macroscopic). A Meso-4 CNN network with 4 CNN layers (with batch normalization, pooling and dropout) followed by a dense layer can classify the forged videos. An alternative MesoInception-4 network, which added Inception modules (Szegedy et al) to this CNN, also does well for this detection. The tests demonstrate a very successful detection rate with more than 98% for Deepfake and 95% for Face2Face.
Monte Carlo Simulation :-
Finally, Atmabit Pattanaik presented a superb session on Monte Carlo simulation, a data distribution technique used for predictive analytics. Atmabit presented several interesting use cases of Monte Carlo, including the probability of winning a bet for a particular number or color while playing roulette casino, predicting random walk outcome, and finally stock price prediction.
Wish to join our amazing AI Lab or learn advanced AI via our extensive ML/DL courses or AI Internship / Research program ? If yes, visit our AI Lab meetups in BLR or Gurugram this Saturday (2nd Nov) :-
Bellandur BLR AI Lab meetup :-
Register : https://www.meetup.com/Disrupt-4-0/events/vcqljryzpbdb/
Topic : NLP with BERT, Cancer Prediction with ML
Loc. : WeWork, ETV, ORR, BLR
Presenters : Amit Kumar, Gouthaman Asokan, Dr. Purnendu Sekhar Das
Infantry Road BLR AI Lab meetup :-
Register : https://www.meetup.com/Disrupt-4-0/events/265386248/
Topic : Deep Compression for Model Deployment, Statistics for ML
Loc. : WeWork Prestige Central, Infantry Rd, BLR
Presenters : Madhav Kopalle, Darshan G.
Gurugram AI Lab meetup :-
Register : https://www.meetup.com/Disrupt-4-0/events/265652913/
Topic : Parameter Estimation with SciPy, Maths in ML
Loc. : The Office Pass, DLF CyberCity, Gurugram
Presenters : Jaidev Deshpande, Anshu Sobti, Akarsh Verma
See you this weekend for the AI Lab meetup ! Let’s disrupt the world with AI, together !
Questions ? Call me at +91-9742800566 !
Co-Founder & Chief Data Scientist, CellStrat