This is a joint effort with the Google Brain team to show the potential of Swift for Tensorflow (S4TF) as a next generation system for deep learning and differentiable computing. This work brings together methods from DCNN to solve the task of pixel-level classification ("semantic image segmentation"). Atrous spatial pyramid module to encode multi-scale contextual information and encoder-decoder structure to capture sharper object boundaries are used to recover the spatial information. The whole pipeline is first implemented using TensorFlow 2.x API. Currently, implementing the model using S4TF.
Unpaired image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between two image domains. Neural Style Transfer is one way to perform image-to-image translation, which synthesizes a novel image by combining the content of one image with the style of another image. Unlike recent work on “neural style transfer”, we used CycleGAN method which learns to mimic the style of an entire collection of artworks, rather than transferring the style of a single selecterd piece of art.
This project studied the Integrated Gradients (IG) attribution method, an Explainable AI technique introduced in the paper Axiomatic Attribution for Deep Networks. The IG method is successfully implemented on an image classification task using the Inception V1 network trained on ImageNet dataset. The project also investigated the effect of using different baselines to determine the sensitivity of this method to the input baseline hyperparameter.
In this project, two prominent aspects of applied deep learning are explored, i.e. facial detection and algorithmic bias. Deploying fair, unbiased AI systems is critical to their long-term acceptance. Consider the task of facial detection: given an image, is it an image of a face? This seemingly simple, but extremely important, task is subject to significant amounts of algorithmic bias among select demographics. Here, we investigated one recently published approach to addressing algorithmic bias.
This project implements MonoSLAM on the wheeled robot using the 1-Point RANSAC for EKF filtering. To make the implementation quite, the is search space is narrowed down for each feature from the whole image to a small ellipse where the feature is predicted to lie. For the closed-loop SLAM operation, a custom deep CNN is built which combines semantic segmentation, VAE, and triplet embedding network. The trained network is used to construct a global feature space to describe both the visual appearance and semantic layout of an image.
This project demonstrates how to train a Variational Autoencoder (VAE) on the MNIST dataset using TensorFlow APIs . A VAE is a probabilistic take on the autoencoder, a model which takes high dimensional input data compress it into a smaller representation. Unlike a traditional autoencoder, which maps the input onto a latent vector, a VAE maps the input data into the parameters of a probability distribution, such as the mean and variance of a Gaussian. This approach produces a continuous, structured latent space, which is useful for image generation.
This project explores building a Recurrent Neural Network (RNN) for music generation using TensorFlow APIs . A model is trained to learn the patterns in a dataset of thousands of Irish folk songs, represented in the ABC notation. The RNN model is based off the LSTM architecture, where a state vector is used to maintain information about the temporal relationships between consecutive characters. For inference, output distributions are iteratively sampled and are encoded to generate song in the ABC format.
In this project, the Actor-Critic method is implemented to train an agent on the Open AI Gym CartPole-V0 environment to control cartpole. In Cartpole, a pole is attached by an un-actuated joint to a cart, which moves along a frictionless track. The pole starts upright, and the goal is to prevent it from falling over by applying a force of -1 or +1 to the cart. A reward of +1 is given for every time step the pole remains upright. An episode ends when (1) the pole is more than 15 degrees from vertical or (2) the cart moves more than 2.4 units from the center.
Over years the design of structures advanced with the availability of historical data and improvement of probabilistic methods. In these structural engineering design approaches, the initial problem constraints are governed by past influences. As climate change and other man-made events rapidly alter our environment, this work seeks to explore a new paradigm in which past influences are removed in order to drive innovation. The removal of the predetermined loading conditions will yield a structure that is not specifically designed for given hazards, but adaptable for any environment.
The force distribution due to axle loads, with respect to the moving referential is modelled using the Finite Element Analysis software called PLAXIS 3D. The matrix of dynamic multipliers for each static point load is genertaed and applied along the railway track with each time step. The results can be used for gauging the effect of changing various geological and physical parameters on the distribution of force due to axle load. Once calibrated to geotechnical and physical data corresponding to sites above actual underground train tracks, the model can be used to predict the effect of vibrations on existing infrastructure.
The multi-hazard effects on moment resisting (MR) industrial unbraced and braced steel frame buildings are assessed independently under non-stationary earthquake and wind induced forces. Joint probabilities of failure are determined in terms of probability density functions (PDFs) and cumulative distribution functions (CDFs) by conducting fragility analysis against each natural hazard independently.