My research interests are at the intersection of machine learning, optimization, statistical learning, and computational neuroscience. I offer an optimization-based signal processing perspective to design and analyze deep learning architectures. Particularly, I am interested in representation learning and probabilistic generative models to develop deep interpretable, and efficient neural architectures. My research is related to a class of machine learning algorithms referred to as unrolled learning in the litearture.

My PhD research is divided into three tracks:

• Deep learning theory for model recovery and interpretability: I use a model-based optimization approach to improve the theoretical rigor of deep learning. This approach allows the designing of deep-learning-based provable algorithms. Moreover, it offers interpretability and mathematical reasons for representations of a new test example and extracts similar/dissimilar data from the training set. For deep sparse coding-based networks, I have shown that backpropagation not only accelerates learning but also guarantees a better model recovery. Check out this work for the interpretability of deep unrolled networks as well as their theoretical properties for model recovery.
• Advancement of inverse problems in engineering: Inverse problems are conventionally solved by slow and unscalable optimization techniques. Deep learning exhibits a superior performance in solving inverse problems at scale. However, inverse problems often suffer from data scarcity. Hence, a question arises concerning the imposition of an inductive bias on deep architectures to enhance their generalization in unsupervised or data-scarce inverse problems. My research addresses this question by designing structured deep networks that optimize a statistical model, demonstrating a superior performance in solving inverse problems with accelerated inference. In my recent work, I highlighted a superior generalization in data-limited regime in radar sensing. In my ICML publication, I have demonstrated the efficiency and state-of-the-art competitive performance of my approach for Poisson image denoising.
• Computational neuroscience: Deep learning can capture neural population dynamics in computational neuroscience. The black-box nature of deep learning, however, limits the unsupervised identification of factors driving neural activity. My research addresses this consequential drawback using interpretable learning; I associate the hidden network representations with a human-understandable model, linking them directly to stimuli and neural activity. My framework has deconvolved, for the first time, the single-trial activity of dopamine neurons into interpretable components in this abstract. Overall, this track aims to enable deep-learning applications to help answer scientific questions in computational neuroscience. I closely collaborate with Paul Masset from Murthy Lab.

During my two fantastic research internship experiences, I worked on speech enhancement. Here, I proposed channel-attention to improve multichannel speech enhancement. In another publication, a joint work with Kazuhito Koishida from Microsoft, I proposed a training framework for perceptual enhancement of stereo speech.