Source: Deep Learning on Medium
NVIDIA recently hosted our first AI Workshop on Research, Trends, and Practices. The workshop brought together data scientists who bring models to production, and researchers who investigate new methods to have a forum to exchange learnings, ideas and best practices.
With AI capabilities becoming an imperative in organisations large and small, public and private, we aimed to uncover what’s the latest across a broad range of topics including Deep Learning, Computer Vision, Generative Models and Attacks, Reinforcement Learning, Natural Language Processing, ML Systems Engineering, and Platforms from experts across academic institutions as well as from the leading technology labs in the industry.
The workshop had speakers that included Prof. Bill Dally, NVIDIA Research; Prof. Balaraman Ravindran, IIT Madras; Prof. Venkatesh Babu, IISc Bangalore; Prof. Pushpak Bhattacharya, IIT Bombay; Alex Sergeev, Uber Engineering; and Abhishek Saxena, Walmart Labs. Additionally Liwei Zhao, Shaohui Jiao and Tianho Xu from NVIDIA China spoke about Scaling Deep Learning inference in Chinese AI companies in the space of E-commerce, retail, and micro-feed entertainment.
Onto the highlights from the sessions:
- Prof Bill Dally shared insights on how hardware accelerated computing combined with mixed precision arithmetic can improve the efficiency of training and inferencing workloads while Alex Sergeev from Uber Engineering spoke on Horovod and simplicity of scaling jobs across multiple GPU nodes
- The first session brought out how to decompose RL problems using Successor RL and also described Risk Averse Imitation Learning by Prof Ravi, IIT Madras. The other talk by Prof.Richa, IIIT Delhi shed light on Adversarial Perturbations for preserving privacy and data-fine tuning
The next session on computer vision covered quite a few aspects:
- Prof. Vineeth from IIT Hyderabad helped discover methods towards Explainable Deep Learning while Prof. Venkatesh Babu, IISc Bangalore spoke on adversarial attacks on models using Network for Adversary Generation with class impressions as ‘proxy’ data.
- Prof. Mayank highlighted performing bio-metric fusion using DL while Dr. Siddhartha Chaudhuri, discussed how adjacency and symmetry are effectively characterized by their hierarchical organization of parts and by using novel neural network architecture for encoding and synthesis of their structures
The NLU sessions covered:
- Prof. Mausam, IIT Delhi discussed on using deep learning for various language processing tasks while Prof. Partha, IISc discussed about named entities in an image plus knowledge graphs and how it is applied to Knowledge-aware Visual Question Answering
- Prof. Pushpak, IIT Bombay brought out techniques on how to tackle low-resolution language modelling that are very relevant to Indic setting.
- The MLSys sessions covered experiences building platforms from key companies. Ram Madhavan, Linkedin and Abhishek Saxena, Walmart Labs where tools like Tony were explored that are frameworks to natively run deep learning frameworks on Apache Hadoop
On day two of the workshop, NVIDIAs Architects from China spoke about about how Chinese giants are adapting AI into their workloads, and leading research globally in many areas.
- Liwei Zhao spoke about how Chinese Companies are executing AI first. They are no more only e-commerce, or bite-sized content, or ride-haling companies, but innovating their business models with AI driven customized and real time customer experiences
- Dr. Shaohui Jiao delved into recommender systems and using software tooling to maintain models in production. She also spoke to the various types of recommender machines being used like DeepFM, Deep Interest Networks, etc
- Tianho Xu spoke about deploying inferencing models at scale — leveraging Hardware and software stack to deliver maximum throughput in minimum latency.
Summary and next steps
The inaugural workshop established a common platform between the researchers, scientists and ML engineers who are solving problems at the scale of billions for India across social and commercial areas. It was encouraging to see the need that this workshop fulfilled bringing together people united by their needs to solve AI challenges like no other. The researcher and practitioner communities are excited to work together and leverage each other strengths and expertise. NVIDIA will to continue to nurture this platform to spread these learnings for the benefit of delivering the good AI holds for India.
We hope the brief shared here serves as good food for thought. Stay tuned for more from us at NVIDIA on how AI research and practices are shaping up in the country. Check out our Developer blog for the latest on NVIDIA tech.
NVIDIA presentations are available here. Please reach out to us for access to other presentations.