Computer Vision News - February 2018

fast physics-inspired on-device pre- processing on noisy raw sensor data using state-of-the-art algorithms from the signal processing realm. This pre- processing component, he reveals, although currently under attack from those who espouse a strictly raw-data- driven end-to-end deep learning narrative, has been crucial in reducing the demand on the volume of training data required to train the forthcoming deep-learning component. The second component, termed as the deep-feature-extraction component, exploits a novel hybrid-CNN-LSTM- deep-learning architecture that has a few ‘convolution’ layers that feed into specialized layers with LSTM cells in the last few layers. This component ingests the pre-processed sensor data from the previous signal processing component and churns out highly discriminative deep features that are then passed on to the forthcoming typicality modeling component in the pipeline. The third component of their pipeline, the typicality modeling component, is in parts, inspired by the innovations such Universal Background Models and the i-vector approach that have attained much success in domains such as speech-based person identification and the general biometrics domain. The last component of their pipeline is the decision-fusion-hypothesis-testing component that entails both decision fusion across the several passive factors by using models from the menagerie of Graphical models as well as ideas from the Bayesian hypothesis testing body of literature to provide the final inference as well as the confidence scores associated with the inference(s). To sum up, they have three components: signal processing, deep learning, and hypothesis testing. The hypothesis testing ascertains the person’s identity: “ At the end of the day - Vinay comments - we are trying to answer the question that, given all of the data, can it tell that the person behind the phone is indeed the person they are claiming to be? ” In the process, they learned that implementing an end-to-end solution does not work. Different types of fancy LSTMs which are proposed didn’t work either. The third thing is that the old, shallow algorithms like the Gaussian Graphical Models and Kernel machines are not to be thrown away. In fact, they play a very important role, especially when it comes to providing explainable inference. One of the things that they have learned through experimentations and implementations, going through several cycles, is not to be stubborn about models. In some cases, the most elegant model that can be put into production in the shortest available amount of time happens to be a ‘shallow learning algorithm’. Vinay says: “ Embrace it. Don’t chase deep learning, deep learning, deep learning 20 Computer Vision News UnifyID Application 50-users gait clustering