Hi, This is Imtiaz

Deep generative networks with continuous piecewise affine (CPA) activations (e.g. relu, leaky-relu) are Affine Spline Operators, that induce a CPA partitioning of space, mapping the latent space domain to the image of the generator via affine operations (left). We build on this theory to develop a change of variables analogy for the density transformation incurred by a CPA generator and provide provable methods for sampling the modes, anti-modes or uniformly from a generator manifold. Our method requires no retraining or labels, and we show that it works for many pretrained GAN/VAE, e.g. using our method we can reduce bias in StyleGAN2 by 41% for the same truncation psi of 0.5.

We study the efficacy of Neural Implicit Representation for limited view CT image reconstruction. We explore Neural Implicit Representation learning framework with 1) differentiable rendering based (e.g. NeRF) frameworks where a sinogram to volume mapping is implicitly learned by rendering the ground truth sinogram using a differentiable CT renderer 2) neural representation upsampling (e.g. COIL, SIREN), coordinate learning neural networks on the input sinogram, is used to interpolate/upsample the input sinogram for inverse reconstruction with an inverse solver.

Coherent light sources create chaotic sensor readings called speckle, while passing through scattering media. We build from scratch, a differentiable RTE renderer for incoherent light (left) and retrofit the RTE equation to account for wave effects of light (middle). Using our method, we are able to reproduce speckle effects for coherent rendering, such as the memory effect- local correlations for smaller shifts of the light source (right). The renderer is built using the JAX autodiff library. As part of the research, we also explore the Mitsuba library for incoherent speckle rendering.

I have developed a visualization toolkit for Texas Children’s Hospital, that allows physicians to monitor ECG and CVP signal waveform variations at higher precision. The toolkit also provides a TSNE based ECG annotation/EDA framework which makes assessing the change of waveforms due to physiological events, more interpretable and accessible.

Biomedical data is at the mercy of the acquisition device. A model trained on data from one device tends to perform below par on data from a different one. In this project, I used an adversarial domain adaptation technique where a conditional adversarial discriminator forces the model to derive domain agnostic features.

Interpretability of machine learning algorithms, especially, neural networks, has been of utmost interest to researchers in recent years. My research explores the interpretability of 1DCNNs from a Finite Impulse Response/Signal Processing perspective.

Embedded Application of Computer Vision Algorithms have wide potential in assistive technologies. Two key challenges remain vital for the success and outreach of such technologies: real-time algorithms and the price tag. My research incorporated the use of cheap webcams with a real-time embedded application. It also utilized Head Related Transfer Functions to modulate stereo sounds for blind assistance.

Music Information Retrieval is the field of research that enables your Shazam, that powers your Spotify recommendations. I was introduced to it through the IEEE Signal Processing Cup 2017, and have fallen in love with it ever since. My interests lie in cover song identification and music beat tracking.