Seminars
Investigating DNN Accuracy Predictors for Network Architecture Search
Description
Advancements in neural network accuracy predictors have reduced architecture evaluation costs in NAS. However, a major trade-off is the lack of generalizability. Neural predictors typically rely on architecture-specific encodings within designated search spaces, limiting the scope of exploration for search algorithms. Specialized predictors are designed for individual search spaces, requiring additional training and labeling costs. A generalized neural predictor would overcome these constraints by accepting input from multiple search spaces and learning a robust architecture representation. This seminar investigates accuracy prediction methods for accurately estimating and ranking network performance across unseen search spaces, not limited to open-source NAS benchmarks.
Reference: Mills, Keith G., et al. "Gennape: Towards generalized neural architecture performance estimators." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 8. 2023.
Contact
Shambhavi.Balamuthu-Sampath@bmw.de
Supervisor:
Reparametrizable DNNs for efficient inference on edge
Description
Complicated Convolutional Neural Networks achieve higher accuracy but have drawbacks during deployment. Multi-branch designs are challenging to implement, slow down inference, and reduce memory utilization. Certain layers like the depthwise convolution or channel shuffle increase memory access cost and sometimes lack device support. Inference speed is influenced by multiple factors, not just floating-point operations (FLOPs). Nonetheless, multi-branch networks are beneficial for improved performance. To tackle this speed-accuracy trade-off, decoupling the training-time multi-branch architecture from the inference-time plain network architecture is possible. This seminar aims to investigate works that achieve fast inference through structural reparametrization of multi-branch networks.
Reference: Ding, Xiaohan, et al. "Repvgg: Making vgg-style convnets great again." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
Contact
Shambhavi.Balamuthu-Sampath@bmw.de
Supervisor:
Minimizing on-device latency measurements for HW-aware DNN optimization
Description
For resource-constrained edge hardware, the development and deployment of high-performance Deep Neural Networks (DNNs) across various applications requires the optimization of DNNs for both latency and accuracy. One way to achieve this is by Network Architecture Search (NAS), which aims to find efficient inference models for target hardware. However, it is an iterative process of evaluating numerous models that incurs significant computational costs. Additionally, accurately measuring on-device latency for each model within a large search space is both exhaustive and impractical. To address these challenges, this seminar focuses on the concept of adaptive sampling, which aims to minimize on-device latency measurements and accelerate hardware-aware DNN optimization. We thoroughly investigate the end-to-end DNN latency prediction methods that can be made sample-efficient. Furthermore, we aim to quantify the potential impact on the GPU hours saved within a typical NAS pipeline, comparing scenarios with and without the employing the latency estimator.
Contact
Shambhavi.Balamuthu-Sampath@bmw.de
Supervisor:
Accelerating End-to-End Autonomous Driving Models on Edge Hardware
Description
In recent years, foundation models have become popular as generic solvers for different tasks, such as text generation, image generation, and semantic segmentation.
In an effort to unify the challenges of autonomous driving, from perception, to occupancy, motion, and path planning, recent works have attempted to create foundation models for end-to-end autonomous driving.
In recent years, foundation models have become popular as generic solvers for different tasks, such as text generation, image generation, and semantic segmentation.
In an effort to unify the challenges of autonomous driving, from perception, to occupancy, motion, and path planning, recent works have attempted to create foundation models for end-to-end autonomous driving.
These models have performed exceptionally well and have proven to be a promising candate to solve the challenges of higher-levels of autonomous driving. However, the complexity of these models and their strictly sequential structure makes it difficult for them to meet real-time execution demands. In this seminar topic, the different approaches to end-to-end autonomous driving will be researched and compared. Then an analytical study will be performed to identify the hardware challenges and opportunities in accelerating them on edge.
Contact
Nael.Fasfous@bmw.de