Summary of Compressing Vae-based Out-of-distribution Detectors For Embedded Deployment, by Aditya Bansal et al.
Compressing VAE-Based Out-of-Distribution Detectors for Embedded Deployment
by Aditya Bansal, Michael Yuhas, Arvind Easwaran
First submitted to arxiv on: 2 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes a novel approach to designing out-of-distribution (OOD) detectors for embedded cyber-physical systems. OOD detectors are critical components in these systems, as they can prevent potentially unsafe actions by identifying samples outside a machine learning model’s training distribution. However, implementing OOD detectors using deep neural networks is challenging due to memory and power constraints on embedded systems. The authors leverage variational autoencoder (VAE) based OOD detectors, which operate in latent space, and apply quantization, pruning, and knowledge distillation techniques to compress the VAE model while maintaining its performance. These techniques are combined to develop lean OOD detectors that can perform real-time inference on embedded CPUs and GPUs. The authors propose a design methodology for this compression process and demonstrate its effectiveness with two existing OOD detectors on a Jetson Nano platform, achieving significant reductions in memory and execution time while maintaining AUROC within 5% of the baseline. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper is about making machine learning models safer for use in robots and self-driving cars. It’s trying to solve a problem where these models can make mistakes if they’re shown something new or unexpected. The authors are using special kinds of neural networks called variational autoencoders, which help identify when something is outside the model’s usual behavior. They then applied some clever tricks to shrink the size of this model while keeping it accurate, so that it can run quickly and efficiently on tiny computers. This is important because these models need to make decisions in real-time, like stopping a car from crashing or preventing a robot from causing harm. The authors showed that their approach works well on small computers and maintains its accuracy, making it a big step forward for safety-critical applications. |
Keywords
» Artificial intelligence » Inference » Knowledge distillation » Latent space » Machine learning » Pruning » Quantization » Variational autoencoder
