Abhisek Kundu

About Abhisek Kundu

I am a machine learning expert with demonstrated history of successfully working in the…

Experience

1. Ai Ml Scientist

   Intel Labs

   Jun 2016 — Present · Bengaluru, IN

   General: Efficient Deep Learning via Low-precision and/or Sparse Representation; Projects:1.[Lead] AutoSparse: Automated Sparse Training for Deep Learning: Achieved state-of-the-art accuracy with best average sparsity for dynamic sparse training via learnable thresholds for Resnet50 (80% sparse) and MoblieNetV1 (75%) on ImageNet1K data. Fully auto-tuned sparsity hyper-parameter significantly reduces the training cost by eliminating the need for many trial runs.2.[Lead] Prediction of Faults in Chip Design: Achieved high accuracy results predicting design faults in chip layout using deep learning, significantly expediting the chip design verification process 3. Influenced incorporating BFloat16 instruction support into Intel Xeon Server Chips for acceleration of AI workloads.4.[Lead] K-TanH: Achieved state-of-the-art GNMT training accuracy using a novel, hardware-efficient approximation to TanH activations. K-TanH delivers >5X speed up over Intel SVML TanH on CPUs. Other activations can be approximated using same technique5. ML-based understanding of Physics of Red Giant Stars: Successfully predicted parameters governing the dynamics of red giant stars using deep learning (collaboration TIFR India), achieving speed up over SOTA methods for the same task6.[Lead] Ternary Residual Networks(TRN): Achieved state-of-the-art accuracy for Deep Learning inference with no retraining using sub-8-bit representation: INT8 activations-Ternary weights. Delivered 2X performance projection for FPGA (PoC on ND5 Chip) over NVIDIA\'s T40 projection using TRN techniques (effectively INT7) for Compiled Instance Neural Networks with almost no multiplication (accuracy drop only 0.22% from baseline). 7.(Ongoing) Sparse, low-precision DL Training/Inference: To achieve 10x speed up exploiting sparsity and low-precision in activations, weights, and gradientsAchievements:1. Intel Gordy Award (Intel\'s highest technical award)2. Intel Divisional Recognition Award

Skills

• Randomized Algorithms
• Machine Learning
• Teaching
• Optimization
• Compressive Sensing
• Numerical Linear Algebra