Introduction : I am an Assistant Professor in the Department of Electrical Engineering, IIT Delhi. As a graduate student at University of California Berkeley, my doctoral research has been in the field of nanomagnetism and spintronics. There I have mostly studied the quantum mechanical phenomena of spin transfer at the interfaces of heavy metal- ferromagnetic metal systems both through simulations and experiments and tried to implement low power logic and memory schemes using such metallic spintronic systems. At IIT Delhi, as a faculty member, I am trying to explore new kind of brain inspired computation schemes that have memory and computing intertwined- hardware machine learning or neuromorphic computing to be specific, using such metallic spintronic devices, which act as low power non volatile elements in these systems. I call my research group - Natural and Artificial Intelligence Through Spintronics (NAITS). Natural intelligence refers to the more biology inspired compuation models like Spiking Neural networks and Artificial Intelligence refers to the less biologicall inspired computation models like Non-spiking Fully Connected Neural Networks used abundantly in the machine learning community.
Research : Neural Network algorithms are currently being widely used by the machine learning and data sciences community to solve various classification, recognition and prediction tasks. These algorithms, if implemented on hardware instead of software, can provide further advantages owing to the parallel architecture and the principle of memory embedded computing inherent in these algorithms. Spintronic devices owing to their non-volatility can make excellent memory elements for hardware implementation of Neural Networks. In that context, in my research group at IIT Delhi, we are simulating hardware neural networks at system, circuit and device level, using spintronic devices (domain wall or skyrmion) as synaptic elements in these networks. We simulate different generations of neural networks: second generation (Non spiking Fully Connected Feedforward type with backpropagation algorithm based learning) and Third generation (Spiking Network with Spike Time Dependent plasticity based learning, inspired by biological data from the brain itself). Schematics of both are shown below:
(a) Ferromagnetic domain wall based device, benchmarked against experiments, that work as synaptic element in hardware nerual network, proposed by Sengupta et al. ( IEEE Transactions on Biomedical Circuits and Systems 10 (6), 2016) (b) Skyrmion based device, proposed by us, that acts as synaptic element (c) Proposed implementation of hardware Fully Connected non-spiking Neural Network with backpropagation algorithm. Reference: Utkarsh Saxena, Divya Kaushik, Mudit Bansal, Upasana Sahu and Debanjan Bhowmik. Low energy implementation of feed-forward neural network with back-propagation algorithm using a spin orbit torque driven skyrmionic device. IEEE Transactions on Magnetics 54 (11) (2018)
(a) Transistor circuit that drives current into skyrmionic / domain wall synaptic device to modulate its conductance, and hence weight as exponential function of time difference between pre and post neuron pulses (Spike Time Dependent Plasticity or STDP) (b) Spiking Neural Network based on STDP synapses which can train on different datasets through supervised learning Reference: Upasana Sahu, Kushaagra Goyal, Utkarsh Saxena, Tanmay Chavan, Udayan Ganguly and Debanjan Bhowmik. Skyrmionic implementation of Spike Time Dependent Plasticity (STDP) enabled Spiking Neural Network (SNN) under supervised learning scheme. IEEE International Conference on Emerging Electronics (ICEE), Bangalore, 2018. A more detailed overview of our group's research can be found here. It is an invited talk I gave at the 84th Annual Meeting of Indian Academy of Scienes held at Varanasi. Please visit my Publications page here for more details on our research. Personal : I am also an active blogger and I write several fiction/ non-fiction posts based on various other interests I have. Please check them out over here. I do have the dream of uniting my technical research with the ideas I conveyed in my blog posts some day.