“AItalk” is a series of webinars hosted by E2E Networks in the domain of computer vision and deep learning AI systems. In one of the webinars, the CEO and COO of the Tessellate Imaging, Mr. Adesh Shrivastava (CEO & Co-Founder at Tessellate Imaging) and Mr. Akash Deep Singh (COO & Co-Founder at Tessellate Imaging) respectively, were invited for a talk on ‘Challenges of Deep Learning in Computer vision’ on Tuesday, February 18, 2020.
In this webinar, computer vision application in the industry, applications of deep learning in computer vision, tessellate history, research, developments, bottlenecks of applying deep learning to computer vision, existing solutions and frameworks, and what’s MonkAI were discussed. Adesh and Akash discussed MonkAI’s features and future, and the future of computer vision application topics.
Computer Vision and Deep Learning
Adesh introduced the history of computer vision, deep learning and real-life applications of AI (Artificial Intelligence). The visual perception of living beings was evolving for millions of years and resulted in today’s modern-day complex visual perception. The technology has been developed and evolving for object detection via image processing. It has become a very crucial parameter for advanced and modern technological solutions in every sector. This has increased the speed of developments in image processing from compression, an enhancement to automating object detection in videos.
Image processing was used primarily for noise removal, media compression, medical imaging, manufacturing, but now it is used in advanced technologies such as AI for various purposes. AI has gained many real-life applications in science fiction movies, shopping complexes, automatic and driverless cars, robots, and applications in deep learning for medical sciences.
Challenges of Deep Learning in Computer vision
Computer vision industry is growing with a rate of 32% CAGR and estimated to be a 49 billion dollar industry by 2023 with advances in algorithms, data and hardware. Whereas, still, about 40% of organisations have not yet defined AI in their organisational strategies and goals. Computer vision and deep learning have challenges in every face that include development, production, and executive level. As the AI has a black box nature, the technology, people’s understanding and processes face development, maintenance and deployment obstacles.
AI systems have concerns regarding limited transparency, biases, and attributions. Other challenges include troubleshooting, limited resources and lack of standardised testing approaches for AI systems due to limited understanding about AI applications. Privacy and safety of deployed AI systems and data ownership are becoming prime concerns nowadays and leading to huge debates recently. AI development, deployment, and maintenance are different from standardised software life cycles which results in development and production level challenges. Precise estimation of efforts is required in the development of a particular application, idea validation, dependency management of AI systems are unclear while developing any system.
Monitoring and logging of existing AI systems and experiment management are also part of the black-box nature of AI systems and thus difficult to manage. Glue code and supporting systems challenges are hard to solve. Still, Tessellate Imaging is refining the toolkit with their learning and experiences in working with different projects by developing toolkits and processes. This has led to the development of MonkAI, an open-source toolkit for the development of computer vision.
MonkAI Toolkit for Computer Vision
Akash demonstrated and gave an idea of the MonkAI platform. Computer vision and deep learning have three essential components of research: (large labelled image or video) datasets, (training and inference) algorithms, and (efficient) deployment. He then demonstrated some work done by Tessellate in computer vision for image classification. A simple, invariant and unified open-source MonkAI toolkit and library can be used to create custom models and switch research between frameworks.
MonkAI has a standard syntax and unified wrapper. Currently, it supports Pytorch, Keras, and MXNet. It helps transfer learning-based custom image classification, custom neural network building and debugging toolkits. It has SOTA deep neural network-based object detection workflows. Monk Studio is a GUI (Graphical User Interface) based approach to deep learning to build computer vision applications. Tessellate team is actively working for the development of future features such as one-click deploy to the cloud, GPU optimisation, image segmentation, multiple imaging modalities, paper to code, GANs (Generative Adversarial Networks) and so forth.
After the demonstration of MonkAI image classification and references to useful lecture links for image processing and computer vision, the webinar ended with a Q&A (Questions & Answers) session. Adesh and Aksh clarified attendees' doubts regarding computer vision, deep learning and MonkAI platform.