Machine Learning Startup Deci Releases Open Source Tool to Analyze Health of AI Training Datasets

Deep learning automation startup Deci AI Ltd. today announced the launch of a free, open-source artificial intelligence tool that can profile datasets for model training purposes.

In a statement, the company said DataGradients will enable data scientists to quickly generate insights about the datasets they plan to use to train new AI models to understand how well those models perform. said.

Deci is the creator of a machine learning development platform used to build, optimize and deploy AI models on cloud, edge or mobile devices. One of the challenges Deci aims to solve is the “AI efficiency gap”. This is a common problem for AI developers whose hardware cannot meet the model’s demands.

The company aims to solve this problem with an automated neural architecture building tool that helps optimize machine learning models for target hardware. Developers simply define the task they want their AI model to solve, provide a training dataset and specify the hardware, and Deci optimizes the model for the specific task and hardware.

DataGradients allow users to better understand how a model will perform even before building it. The startup says it’s particularly useful in computer vision, where model capabilities are directly related to the quality of the data used to train the model.

For AI developers, it is of utmost importance to be able to identify problems and weaknesses in their datasets in order to avoid training roadblocks and be able to perform their intended tasks satisfactorily. With a better understanding of the underlying dataset, developers can make smarter decisions about choosing the right model, the best loss function, and how to optimize, Deci said.

More specifically, DataGradients allow data scientists to analyze and establish the health of their datasets, identifying issues such as corrupted data, distributional changes between training and test datasets, and duplicate annotations. can be identified. Users are also provided with insights to help mitigate these issues and improve dataset quality to improve model performance.

Andy Thurai, Vice President and Principal Analyst at Constellation Research Inc., told SiliconANGLE that training deep learning models, including computer vision models, is highly dependent on the availability of high-quality training datasets to achieve desired accuracy. said it can be very difficult. “When you train a computer vision model on a substandard quality dataset, the results are often very unpredictable,” he explained.

Fortunately, Thurai said, data scientists have many data quality tools at their disposal to help determine if a dataset is fit for purpose. “There are many tools available on the market, but the open source nature of DataGradients may increase its profile among the developer community,” he added.

Deci co-founder and CEO Yonatan Geifman said DataGradients is all about streamlining the model development and training process with “very clear visibility” into the underlying datasets used. said. He said this is the third open-source tool released by the company, following the launch of the company’s PyTorch training library SuperGradients and object detection infrastructure model YOLO-NAS.

Image: svstudioart/Freepik