Analyzing problems from a data viewpoint might reveal striking similarities between problems in different sectors. Despite coming from different demands and domains, many business problems frequently have a similar theme and solution.
Consider computer vision, for example. Since the pandemic ended, research has repeatedly shown that computer vision can measure and observe customer behavior in shopping aisles or identify empty shelves in both retail contexts. To enhance predictive maintenance procedures, the same technology can be implemented on the manufacturing floor in the meantime.
When these issues are examined from a data viewpoint, it becomes clear that the same data solutions—in this case, computer vision—can perform in support of various business objectives and settings, producing differing outcomes based on requirements, function, and training.
Alberto Rizzoli, co-founder and CEO of V7 Labs, and Matthew DeMello, senior editor of Emerj, recently discussed how new generative AI and NLP-based tools can be deployed in contexts across life sciences and retail to handle various problems that look much more similar from a data-based perspective.
The analysis of their talk that follows looks at two important takeaways for executives in a variety of industries:
Developing AI that can independently label data: utilizing self-learning artificial intelligence (AI) systems to autonomously label products, eliminating the need for laborious manual labeling.
Adapting AI systems to the particular requirements of the retail industry: modifying pre-made AI goods to match the unique information and looks of the merchandise sold by the shop.
Knowledge In Computer Vision, Deep Learning, And AI
In a nutshell: Alberto is the CEO and co-founder of V7. He is an Italian businessman. With a Google scholarship, he enrolled in Singularity University’s Graduate Studies Program at NASA Ames Research Park in 2015 after earning a Bachelor of Science in Management from Cass Business School. Italian President Sergio Mattarella gave Alberto an award in March 2016 for his contributions to the advancement of artificial intelligence for charitable purposes.
Building An AI-Reliant Data Labeling System
Alberto starts by pointing out how much the life sciences industry uses imaging to find new drugs and perform scientific studies. On the other hand, these fields include a large number of intricate images. He goes into more detail on the difficulty of working with a lot of cells on digital pathology slides. These slides may include thousands of cells, but only a small percentage of them may be malignant or have particular traits.
After that, Alberto talks about how big models are used in microscopy and radiography. In these areas, a great deal of effort has been made in identifying changes in cell segmentation over time. He uses 3D microscopy as an illustration. He proposes that 3D image cell segmentation can be done automatically via basis models. In this procedure, cells are identified using complicated microscope data.
After that, Robert talks about the idea of data labeling in AI systems. Labeling data is similar to appending bits of information to documents, photos, or any other type of media that the AI uses. The AI may carry out tasks embedded within the patterns it learns from the data as it “eats” more of this labeled material.
He observes that AI systems are now more self-supervised, which means that less intensive manual labeling is required because they can determine labels on their own to some level.
Alberto continues by describing how artificial intelligence (AI) can watch a YouTube video and pull labels from the subtitles. The AI begins to comprehend the connection between the labels and the content by attaching the labels to pictures. This kind of self-supervised learning is comparable to the way early models of image production deduced picture information from captions.
Alberto adds that when a lot of data needs to be gathered quickly—typically via the internet or a data repository—self-labeling or self-supervision becomes useful. When there isn’t enough time to manually label every piece of data, it helps, especially if the output isn’t very significant.
He gives the early development of massive language models as an example, which required scraping a sizable amount of text data from the internet. He does, however, draw attention to the fact that building these models requires more than just self-supervised learning; significant human intervention is also required to guarantee the models do not produce offensive or dangerous information.
To wrap up, Alberto talks about how difficult it is to keep labeling standards high enough to get the required accuracy.
The final objective is to guide the AI away from its initial “feral state,” which it got from raw internet data, and educate it to mimic activities that competent humans execute, even though self-supervision can serve as a basis. The goal is to hone the AI’s skills to the point where it can generate results that are on par with the caliber of work produced by human specialists.
Alberto stresses in his conclusion that, despite being costly, this investment in superior human input eventually improves the AI model’s precision and dependability.
Customizing AI Solutions To Meet The Special Retail Needs
He notes that because there are so many products on the market, each with a unique look and set of variations, the retail industry poses a unique set of obstacles. Retail requires more precise identification, in contrast to other computer vision applications where generic categories (such as vehicles or motorbikes) can be identified. Retail executives may need to identify a specific stock-keeping unit (SKU), such as a limited-edition Christmas version of a soft drink, to effectively manage their inventory. For example, simply recognizing any soft drink is not enough.
Alberto continues by pointing out that there are significant infrastructure difficulties associated with the quantity of data in retail. Therefore, to handle and manage this enormous amount of data effectively, having a strong data infrastructure—including appropriate tooling and storage methods—is essential.
The CEO of V7 further notes that given the extreme diversity of products, more than standard classification methods may be needed in retail settings. Rather than classifying data into established categories, he recommends using approaches like retrieval, in which the AI system retrieves particular items based on resemblance or relevance.
Alberto also suggests that companies think carefully about the AI use cases they wish to apply in the retail industry. While off-the-shelf solutions may exist, they frequently require customization and tuning to the unique data and product presentations of the business. Regardless of the application’s purpose—autonomous checkout, out-of-stock detection, or improving the online shopping experience by making product recommendations to users—this personalization is essential.
Robert makes a comparison between retail establishments and warehouses. He points out that the underlying ideas that guide the technology employed in these settings are rather similar: whether in a retail or warehouse scenario, both entail managing and tracking goods.
He observes that the technology underlying retail’s self-checkout systems is closely related to that of warehouse systems. For example, a lot of businesses are adopting robotic solutions that can pick products out of warehouses, and there’s a rising trend of directing human workers to the right place for certain items.
Furthermore, efforts are being made to improve warehouse safety by incorporating autonomous navigation capabilities into machinery such as forklifts. It is especially important in places where big things and high shelves might cause accidents with serious consequences.
He also highlights the possibility for automation in lesser jobs in addition to warehouses’ main purpose of storing and retrieving goods. Palletization, the process of placing goods on pallets for transportation, and routine object scanning as they enter and exit the warehouse are examples of these duties. By utilizing nearby robots or cameras, these procedures can be expedited, saving time and effort by requiring less physical labor.
Robert concludes by talking about the importance of infrastructure and its maintenance. Artificial intelligence (AI) could be used to monitor and assess infrastructure, identifying things like concrete cracks or worn-out areas. Robotic quadrupeds such as Boston Dynamics’ Spot robot inspect energy infrastructure and transmit data for artificial intelligence evaluation.
Like radiologists spotting medical problems, AI systems in a central hub can spot anomalies in infrastructure data. Large models are leveraged by AI to produce thorough Health Reports for infrastructure.