NVIDIA GTC, AI conference for innovators, technologists, and creatives-April 2021
Nvidia Clara
Nvidia is an American company specialized in the design of graphics processors, graphics cards and graphics chips for PCs and game consoles.
The computational performance of GPUs originally designed for 3D display can perform the mathematical calculations required to train Deep Learning models. This makes NVidia a key player in the field.
There are more than 1,200 start-ups in the healthcare field that use Nvidia Clara.
Nvidia used its annual NVIDIA GTC 2021 Artificial Intelligence conference to announce new investments in its unified platform for imaging, genomics, patient monitoring and pharmaceutical research: Nvidia Clara.
Nvidia, in recent years, has been investing heavily in the areas of healthcare most impacted by AI: medical imaging, genetic sequencing, conversational agents, pharmaceutical research and medical robotics.
Clara Medical Imaging
The Clara Imaging application framework provides developers and researchers with the tools they need to accelerate data annotation, design AI models, and deploy intelligent imaging workflows with pre-trained models.
Clara Parabricks
Clara Parabricks® combines GPU-accelerated sequencing software and data analysis in the field of genomics.
Clara Guardian
The Clara Guardian application framework provides advanced intelligent video analytics and conversational AI capabilities to healthcare facilities.
Clara Discovery
Clara Discovery brings together a set of AI frameworks, applications, and models to accelerate pharmaceutical research.
Clara AGX
Clara AGX™ is a software development kit that centralizes a low-power embedded computing solution. The kit helps hospitals develop and deploy AI on smart sensors such as endoscopes, ultrasound devices and microscopes.
Pre-trained models
Nvidia has also developed a library of 40 pre-trained models in the medical field: Cardiology, oncology, covid-19.
Nvidia offers some of the best pre-trained models in the field of medical image classification such as DenseNet121, ResNet, Unet.
- Classification-chest-xray: a densenet121 pre-trained model for detecting lung infections from chest X-rays.
- Clara-mri-seg-brain-tumors: a pre-trained model for volumetric (3D) segmentation of brain tumors (only CT from T1c images)
- Clara-train-covid19: a model for segmenting the lung region from 3D chest CT images.
In October 2020, they added to their library an NLP model trained on a large clinical and scientific corpus: the BioMegatron. This model can be used to extract key concepts and relationships from biomedical texts and build knowledge graphs that can drive research and discovery.
It has been developed from the NVIDIA Megatron model developed with the PyTorch language, based on a Transformer architecture.
BioMegatron was pre-trained on a cluster consisting of eight DGX-2s for approximately 400 hours, or roughly two weeks. The NVIDIA DGX-2 system embeds sixteen NVIDIA V100 Tensor Core GPUs, which allows the models to be trained with large datasets characteristic of the new NLP models (GPT-3, NeMo, BERT).
Transformers, private and public partnerships, federated learning.
Nvidia’s announcements at the GTC 2021 conference illustrate the major trends in Artificial Intelligence in the field of medicine.
- The use of NLP models based on Transformer architectures to discover new drugs in record time.
- The development of partnerships between researchers, companies and private laboratories.
- The development of “private-public” partnerships with university research centers to develop new models.
- The development of federated learning.
Discovering new drugs
Huang, Nvidia’s CEO, noted at the NVIDIA GTC 2021 keynote that the combination of AI and pharmaceutical research has created models that can define new drugs in less than two years, whereas this process used to take up to ten years.
Nvidia claims that its GPU platform can create a molecule with a single A100 GPU in about 0.3 seconds. With the new DGX SuperPOD, according to Huang, researchers could generate thousands of new molecule structures per second.
This biological revolution comes at a cost. The price of the least expensive DGX Superpod is $7 million. The bill can soar to $60 million for the most processor-intensive versions.
The Schroedinger Collaboration
The collaboration between Nvidia and Schroedinger will combine the DGX A100 platform with the chemical simulation software specialist’s machine learning techniques to accelerate drug discovery.
Collaboration with AstraZeneca
NVIDIA will collaborate with AstraZeneca on the creation of a Transformer-based model, the Megamolbart, to create chemical structures used in drug discovery.
The model will be open sourced and trained on the UK’s most powerful computer Cambridge-1. Available to researchers and developers in the NVIDIA NGC catalog, the Megamolbart will be deployable via the NVIDIA Clara Discovery platform for drug discovery.
Private-Public Partnership with the University of Florida
Nvidia will collaborate with the University of Florida to develop a model, the GatorTron, trained on records of more than 50 million interactions with 2 million patients.
The model is intended to identify patients for life-saving clinical trials and provide clinical decision support to physicians.
The development of federated learning
In October 2019, in collaboration with King’s College London, NVIDIA Research had launched the first federated learning system for medical image analysis.
Deep learning feeds on large amounts of patient data. When this data is spread over several sites, federated learning makes it possible to exploit it without circulating it, and therefore without exposing it to confidentiality risks.
It is by aggregating the models trained locally in their corner that we obtain a neural network that has learned with data that has never been gathered.
At the NVIDIA GTC 2021 conference, Nvidia announced the next generation of federated learning — Clara 4.0.
Following the announcement, David Ruau, vice president and head of Global Data Assets and Decision Science at Bayer, announced the adoption of Clara Federated Learning by his lab.
Appendix
Healthcare conference sessions NVIDIA GTC 2021
