We propose a training scheme to train neural network-based source separation algorithms from scratch when parallel clean data is unavailable. In particular, we demonstrate that an unsupervised spatial clustering algorithm is sufficient to guide the training of a deep clustering system. We argue that previous work on deep clustering requires strong supervision and elaborate on why this is a limitation. We demonstrate that (a) the single-channel deep clustering system trained according to the proposed scheme alone is able to achieve a similar performance as the multi-channel teacher in terms of word error rates and (b) initializing the spatial clustering approach with the deep clustering result yields a relative word error rate reduction of 26 % over the unsupervised teacher.
<div class="section abstract"><div class="htmlview paragraph">The development of perception functions for tomorrow’s automated vehicles is driven by enormous amounts of data: often exceeding a gigabyte per second and reaching into the terabytes per hour. Data is typically gathered by a fleet of dozens of mule vehicles which multiply the data generated into the hundreds of petabytes per year. Traditional methods for fueling data-driven development would record every bit of every second of a data logging drive on solid-state drives located on a PC in the vehicle. Recorded data must then be exported from these drives using an upload station which pushes to the data lake after arriving back at the garage.</div><div class="htmlview paragraph">This paper considers different techniques for curating logged data. These curation methods are performed to maximize the usefulness of the data throughout its lifecycle and minimize the amount of data necessary for perception development and validation The reduction of logged data has the effect of not only curtailing storage costs, but also minimizing latency for data availability and maximizing possible campaign drive time. Advanced techniques considered include: (a) the real-time evaluation of sensor and bus signals, (b) the application of artificial intelligence (e.g. for similarity-based image discovery and dynamic scene content detection), and (c) using function prototyping to inform the data curation process. In addition, various possible integration points for the curation techniques are considered in the data ingestion pipeline - from the cloud in the datacenter to the in-vehicle logging system on the edge. The trade-offs for each proposed techniques are considered across the various components of the data ingestion pipeline to explore the feasibility of - and to arrive at some conclusions for - designing more intelligent data collection campaigns.</div></div>
A number of studies have investigated the training of neural networks with synthetic data for applications in the real world. The aim of this study is to quantify how much real world data can be saved when using a mixed dataset of synthetic and real world data. By modeling the relationship between the number of training examples and detection performance by a simple power law, we find that the need for real world data can be reduced by up to 70% without sacrificing detection performance. The training of object detection networks is especially enhanced by enriching the mixed dataset with classes underrepresented in the real world dataset. The results indicate that mixed datasets with real world data ratios between 5% and 20% reduce the need for real world data the most without reducing the detection performance.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
customersupport@researchsolutions.com
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.