Research
TinyML to reduce Model Size
Lightweight machine learning models to run on Edge AI devices.
With limited resources (Processing speed, Memory, Low power range)
Model reduction and On-chip training
Approximate Computing
Exploit the tradeoffs between Accuracy vs Power Consumption and Hardware Resources
Error Tolerant computing with In-exact processing elements like Multipliers and Adders
Approximate MAC
To design an approximate Logarithmic MAC
To explore techniques like Radix Booth Encoding and Use of Barrel Shifters
To design accuracy configurable adder
Approximate CORDIC Activation Function
To design an approximate Sigmoid and Tanh
To explore techniques like Loop perforation and approximation
To implement and test the design on TSMC 45nm Technology
Cross-Layer Approximation GEMM Unit
To design accuracy-configurable hardware to cater to the needs of different CNN layers
To exploit Data precision formats like BF16, TF32, etc.
To verify and test the GEMM design for TinyYolo