Projects

Deep Neural Networks with Trainable Activations and Controlled Lipschitz Constant

Organization: Indian Institute of Technology Madras

Team Size: 4

Tools used: Python, Visual Studio Code, LaTeX.

Responsibilities:

  • We studied a variational framework to learn the activations of a deep neural network while controlling its global Lipschitz regularity.

  • Neural networks with second-order bounded-variation activations were considered and a global bound for their Lipschitz constants was provided.

  • The solution to the variational problem could be expressed in the form of a deep-spline network with continuous piecewise linear activation functions.

Frequency Synchronization for OFDM-Based Massive MIMO Systems

Organization: Birla Institute of Technology and Science, Pilani

Team Size: 3

Tools used: MATLAB, LaTeX.

Responsibilities:

  • We studied frequency synchronization for the OFDM-Based Massive MIMO system.

  • The study dealt with channel estimation, carrier frequency offset (CFO) estimation, and analysis of Mean square error as a function of signal-to-noise ratio (SNR), in the presence of as well as the absence of CFO.

  • Additionally, deep learning algorithms were implemented for channel estimation.

Malaria Parasite Detection using Convolutional Neural Networks

Organization: Birla Institute of Technology and Science, Pilani

Team Size: 2

Tools used: Python, Jupyter Notebook, LaTeX.

Responsibilities:

  • Investigated the performance of pre-trained (ResNet-50) and the customized convolutional neural network (CNN) for the classification of a cell image into parasitized and uninfected classes.

  • Obtained the accuracy of 97.2% and 95.4% for the ResNet-50 and the custom network, respectively.

  • As part of the novel idea, an architecture based on support vector classification with radial basis function kernel was also implemented.

Design of Variable Latency Speculative Han-Carlson Adder

Organization: Birla Institute of Technology and Science, Pilani

Team Size: 3

Tools used: Cadence Virtuoso.

Responsibilities:

  • Implemented variable latency speculative Han-Carlson adder using the CMOS TSMC 180 nm library.

  • Results revealed that the proposed adder outperforms the conventional Ripple carry adder and non-speculative architectures in terms of speed.

Non-Local Means Image Denoising with a Soft Threshold

Organization: Birla Institute of Technology and Science, Pilani

Team Size: 3

Tools used: MATLAB, LaTeX.

Responsibilities:

  • Studied an improved method of Non-Local Means (NLM) image denoising, which outperformed the classic patch-wise NLM.

  • Incorporated soft threshold into the NLM algorithm (NLM-ST) and used MCLA (Monte-Carlo Linear Aggregation) to improve the denoising performance, keeping SURE (Stein’s Unbiased Risk Estimate) minimum.

  • As a part of the novel idea, when the noise intensity was high, Donoho’s wavelet denoising algorithm was used to remove noise initially and the denoised image was further improved by the NLM-ST method.

Implementation of Polar Codes for Cascaded Binary Symmetric Channel

Organization: Birla Institute of Technology and Science, Pilani

Team Size: 3

Tools used: MATLAB, LaTeX.

Responsibilities:

  • Implemented Polar Codes for a Binary Symmetric Channel as well as Cascaded Binary Symmetric Channel.

  • Studied and compared their performance in terms of Bit Error Rate (BER) and Frame Error Rate (FER).

Audio chord recognition using deep learning and other mathematical tools

Organization: Jadavpur University, Kolkata

Team Size: 4

Tools used: Python.

Responsibilities:

  • Chords were automatically identified from an audio sample.

  • Theoretical concepts of signal processing were utilized and simulations were performed on Python.