Developing a deep learning algorithm, or any artificial intelligence (AI) algorithm for that matter is nothing short of a long process. From conception to production and implementation, the process goes through many iterations and is owned by multiple collaborators. Considering the mere process of obtaining and organizing your input data could take a magnitude of time, actually developing AI for your organization is extraordinarily complex. Allegro.ai attempts to organize and simplify the process for you, albeit if you’re a serious player.
Allegro.ai was founded by Nir Bar-Lev (the product leader veteran), Moses Guttman (the deep learning and computer vision expert) and Gil Westrich (the polymath software engineer). Guttman was the first student of Lior Wolf, Israel’s prized computer vision pioneer who championed deep learning in Israel, and also holds the record of the youngest PhD professor ever in the Tel Aviv University faculty.
Bar-Lev has pedigree in product and technology leadership experience, as he led Google’s first expansion into Israel on the product side, along with Yossi Matias, the engineering lead for the R&D office in Israel. In his early days at Google, he worked on building Google’s voice recognition system in 2005. At the time, they used statistical systems because machine learning hadn’t truly proved effective in the field. According to Bar-Lev, over the last few years, Google re-wrote the entire voice recognition technology stack by using deep learning.
Bar-Lev is a graduate of MAMRAM, a coveted elite computer unit of the Israeli army, while Guttman and Westrich both served in unit 8100, the lesser-known top secret unit like 8200, although the nature of their work lied in hardware inventions. During Bar-Lev’s time at Google, he led Google’s Search Advertising in Europe, Middle East and Africa, among other product leadership roles.
Essentially, allegro.ai’s platform covers three main bottlenecks in the computer vision algorithms product lifecycle: managing data, labeling data and developer operations (DevOps) for experimentation and deployments. It’s also the ideal conducive environment (cloud or on premise) to facilitate teamwork, which often is the case for building complex computer vision algorithms. They also enable image and video annotations with their proprietary auto-annotation process, a fast-tracking tool that can change the estimated time to completion of ground-truth datasets by an order of magnitude. These tools combined form the allegro.ai platform – the ultimate scaling tool for development and production of computer vision algorithms and products.
In the end of 2015, Bar-Lev left Google and joined his two co-founders to start allegro.ai. He was looking to solve significant problems, be apart of the AI revolution, and see changes unfold in front of his eyes. A big problem that Bar-Lev pondered was the process of building software at the enterprise-grade level. “Once you do solve problem, you need to scale it while minimizing costs and maintaining quality” explained Bar-Lev, who drew the comparison to a car production line that enabled the scaling of manufacturing cars while maintaining quality and achieving cost effectiveness. Back in his early days at Google, Bar-Lev witnesses first hand how Google’s huge investment in infrastructure for operating at scale was on of its untold secrets for its massive success. “The things we can accomplish today with software aren’t just because of new and improved chipsets, to a considerable extent, it’s because of a multibillion dollar industry of software infrastructure and tools that enable companies to scale their software development and deployment.” This effect was evident at the time from the Whatsapp acquisition, a startup that was composed of 50 engineers, serving hundreds of millions of users.
Deep learning is a fundamentally different paradigm for software than traditional software/hardware development and management. As a result, there’s a huge shortage on both talent, tools and infrastructure to develop high-quality products, at scale and with commercially viable time-to-market and return on investment.
These challenges are common to all deep learning applications, not just computer vision. In computer vision specifically, the problems get more exacerbated because (1) data is huge and it needs to be managed in a unique way for deep learning. As an intuition to this – data is really the raw material to create the software in the first place under the deep learning paradigm; and (2) in many computer vision cases, the ultimate deep learning system needs to run on or near some edge device (camera or similar device) where compute power and memory is limited. Conversely, deep learning is a highly compute and memory-intensive process. So this becomes another challenge to deal with.
Compounded with the rise of connected and cognizant devices that are critical to the AI-laden industries, like autonomous cars and drones, or security cameras, which are all predicated on processing intensive deep learning, allegro.ai is the perfect solution. Allegro.ai was designed to provide a company with the tools to build computer vision algorithms, but also the necessary tools to manage and personalize them after production. “Data is the raw material in deep learning. The feedback mechanism is critical to the scientific experimentation paradigm of algorithm development and production” says Bar-Lev.
“If you want to run statistics, amateurs will use Excel, but professionals will use SAS or MATLAB. If you want to dabble in Artificial Intelligence algorithms, you might try leveraging the tools being rolled out by some of the large cloud providers. But if artificial intelligence is fundamental to your product, you’re going to use our platform” explained Bar-Lev. Allegro.ai’s development platform is used by automotive manufacturers, healthcare conglomerates, defense contractors and other industry giants.
Allegro.ai has built a software stack for deep learning, where data is the raw material, and inherently connected the feedback loop to the life-cycle of the platform. They raised over $11 million in total funding to date, recently announced a new strategic investment from The Hyundai Motor Company and are currently hiring new employees to join their stellar team. They’re also supporting the Israeli AI startup community with their newly formed “Startup Program” dedicated to streamlining startups’ AI development. It’s impossible to truly build deep learning computer vision algorithms without a multi-faceted platform. But as the industry grows, and the speed of development accelerates in the technology world, allegro.ai is scaling up to scale the development of computer vision products.