Nehete Hemkant
Prime Minister Research Fellow (PMRF)
Microelectronics and VLSI Group,
Indian Institute of Technology Roorkee, India
Research Interests
FPGA based Hardware Accelerators
Developing Learning Algorithms for Deep Learning.
TinyML for Mobile and Edge Computing
In-memory computing Architectures
New Updates
March 2025: Work on Spatial-frequency domain non-local attention based restoration network got accepted at IJCNN'25.
March 2025: One paper got selected for CVPR'25 Workshop: New Trends in Image Restoration and Enhancement (NTIRE).
December 2024: Work on Approximation aware training and its CiM architecture got accepted at IEEE Open Journal on Nanotechnology.
November 2024: Work on ML models for estimating magnetic parameters got accepted at IEEE Open Journal on Nanotechnology.
April 2024: One paper got selected for CVPR'24 Workshop: New Trends in Image Restoration and Enhancement (NTIRE).
March 2024: Filled Patent on Feature Enhancement with Channel Fusion in Convolution.
February 2024: Paper accepted at IEEE Transactions on Electron Devices on Predicting formation on AFM Skyrmion with ML.
June 2023: Two papers accepted at IEEE NMDC 2023 Conference.
May 2023: Two papers accepted at IEEE NANO 2023 Conference.
May 2023: Two papers accepted at SPIE Optics+Photonics 2023 Conference.
March 2023: Received Prime Minister Research Fellowship.
March 2023: Paper accepted at IEEE Open Journal on Nanotechnology on SOT based MAAP unit for merged pooling and activation function.
January 2023: Paper accepted at IEEE Transactions on Electron Devices on Hybrid Multilevel STT/DSHE Memory Architecture for Training CNNs.
October 2022: Two papers accepted at SPIE Photonics West 2023 Conference.
Selected Publication
Fourier Prior-Based Two-Stage Architecture for Image Restoration
~CVPR'24 Workshop: New Trends in Image Restoration and Enhancement (NTIRE)
This work presents a novel two-stage architecture intended to improve images that have been deteriorated by rain, haze, blur, and other environmental factors. We propose the Fourier prior to improve the generalization ability of image restoration models. This is based on an important finding: degradation can be effectively mitigated by replacing the Fourier amplitude of degraded images with that of clean images. Indicating that phase preserves background structures while amplitude contains information about degradation. As a result, a two-stage model is presented, comprising the Phase Refinement Unit and the Amplitude Refinement Unit, which independently restore phase and amplitude information, respectively of degraded images.
Approximation-Aware Training for Efficient Neural Network Inference on MRAM Based CiM Architecture
~IEEE Open Journal of Nanotechnology
This work proposes approximation-aware-training, in which group of weights are approximated using a differential approximation function, resulting in a new weight matrix composed of approximation function's coefficients (AFC). The network is trained using backpropagation to minimize the loss function with respect to AFC matrix with linear and quadratic approximation functions preserving accuracy at high compression rates. This work extends to implement an compute-in-memory architecture for inference operations of approximate neural networks. This architecture maps AFC to crossbar arrays directly. This reduces the number of crossbars, lowering area and energy consumption. Integrating magnetic random-access memory-based devices further enhances performance by reducing latency and energy consumption.