Ryzen AI NPU team

1. Exploring how scale of an image can be used to determine optimal complexity of a Deep Neural Network.

1. Designed an innovative approach using Inter-Class Similarity Metrics to optimize the size of Neural Network architecture. This approach compared similarities between different image class groups, ensuring enhanced efficiency in models like YOLOv5, EfficientNet, and MobileNet.2. Implemented the integration of optimization techniques with an automated graph network that allows seamless parameter grouping and facilitate Structural pruning to remove parameters physically. Proved instrumental in enhancing model efficiency by systematically eliminating redundant network components without compromising accuracy.

1. Engineered Video Analytics system on a Disaggregated Architecture of a cluster of heterogeneous NVIDIA Jetson devices.2. Implemented fast & memory efficient pipelines to optimize the performance of Deep Learning models on ML acceleration hardware by addressing performance bottlenecks.3. Orchestrated deployment on Kubernetes and streamlined machine learning workflows with Kubeflow integration.4. Utilizing sparsity, quantization-aware training (QAT) & post-training quantization (PTQ), operator fusion, pruning, ONNXRuntime & TensorRT optimizations to minimize latency & enhance throughput.5. Developed custom C++ operators (LayerNorm, GeLU) for TFLite, enabling real-time model inference on NPU.6. Deployed License Plate Recognition CI/CD pipeline using PyTorch & Docker with 1.93x increased processing rate.7. Developed a Split Computing algorithm (model sharding & model merging) to ensure optimized memory and computing resource allocation for deployment (CPU & GPU), efficiency and scalability for Deep Neural Networks.8. Published a thesis contributing to advancements in Machine Learning Infrastructure and Optimizations.

1. Developed a Qt (C++) application for recognizing Gujarati Script using EfficientNet B3 model for optical and hand-written text with an accuracy of 99.7%.2. Recognized text was further processed by using a Bi-LSTM model to rectify grammatical errors. Increased the recognition accuracy of conjunct consonants by 12% in comparison with traditional OCR models.3. The results of this project is accepted for publication in Procedia Computer Science, 2022.