Machine Learning Startup Deci, Introduces Open-Source Tool for Assessing AI Training Dataset Health

Highlights:

• Using DataGradients, users can gain improved insights into the performance of their models even before their development.
• According to Andy Thurai, the Vice President and Principal Analyst at Constellation Research Inc., training deep learning models, such as computer vision models, can be incredibly challenging.

Deci AI Ltd., a startup focused on deep learning automation, recently unveiled a new AI tool to profile datasets designed for model training. This tool is open-source and free to use.

According to the company’s announcement, DataGradients empowers data scientists to rapidly generate insights on datasets they intend to use to train new AI models. By doing so, they can evaluate the model’s potential capabilities in advance.

Deci is a company that has developed a machine learning development platform. The platform helps build, optimize, and deploy AI models on various cloud, edge, and mobile devices. Deci aims to address a significant challenge, the “AI efficiency gap.” This is a common issue faced by AI developers, where the hardware they utilize could be more capable of fulfilling the demands of their models.

Deci has introduced the Automated Neural Architecture Construction (ANAC) tool to overcome this challenge. By leveraging ANAC, the company endeavors to optimize machine learning models specifically for the target hardware, providing a solution to this problem. All developers need to do is specify the task they want their AI model to tackle, provide the training dataset, and specify the hardware they plan to use. Deci handles the rest by optimizing the model to suit best the given task and hardware requirements.

Using DataGradients, users can gain improved insights into the performance of their models even before their development. The startup highlights its value, particularly in computer vision applications, as the effectiveness of models directly correlates with the quality of the training data employed.

AI developers place paramount importance on their ability to recognize any issues or weaknesses present in their datasets. This is essential to prevent obstacles during the training process and to guarantee that the AI model can sufficiently fulfill its intended tasks. Deci stated that AI developers with a solid comprehension of their dataset can make more intelligent choices regarding model selection, loss function, and optimization methods.

To be more precise, DataGradients allows data scientists to assess and ensure the integrity of datasets. This includes identifying corrupted data, detecting distributional shifts between training and test datasets, recognizing duplicate annotations, and other related issues. Additionally, users gain invaluable insights from DataGradients that aid them in mitigating these issues and enhancing dataset quality. This ensures that their models will deliver superior performance.

According to Andy Thurai, the Vice President and Principal Analyst at Constellation Research Inc., training deep learning models, such as computer vision models, can be incredibly challenging. This is primarily because attaining the desired level of accuracy in these models heavily relies on having high-quality training datasets. Thrai explained, “When you train computer vision models on datasets of subpar quality, the results are often very unpredictable.”

Fortunately, data scientists have access to various data quality tools that can aid them in evaluating the suitability of their datasets, as highlighted by Thurai. The Vice President at Constellation Research also added, “There are many tools available commercially, but the open-source nature of DataGradients may help it to gain more traction among the developer community.”

According to Yonatan Geifman, Deci’s Co-founder and Chief Executive, DataGradients focuses on simplifying the model development and training process by providing “crystal-clear visibility” into the datasets being utilized. Geifman mentioned that DataGradients is the third open-source tool released by the company, following the introduction of SuperGradients, their PyTorch training library, and YOLO-NAS, their object detection foundation model.