Technews
Intel and Mila Strengthen their Open Innovation Commitment to Responsible AI
Author: Kannan Keeranam, Director, Cloud & AI Strategy and Execution at Intel
Intel announced it has strengthened its strategic research and co-innovation collaboration with Mila, an artificial intelligence research institute based in Montreal. As part of a renewed three-year commitment, Intel and Mila researchers will focus on Responsible AI and developing advanced AI techniques for problems such as climate change, new materials discovery, and digital biology. Solving these complex challenges requires deep AI research and a commitment to open innovation to advance state-of-the-art (SOTA) AI.
Automated AI-Driven Discovery of Novel Materials
As scientists and organizations work to tackle climate change, one of the most relevant research topics for this is the AI-driven discovery of novel materials that can drastically reduce costs and technologies’ carbon footprint. While advances in chemical simulation techniques, such as density functional theory, have created various methods capable of simulating important properties of complex material systems, these techniques have been limited in the complexity of materials systems they can model given the unfavorable scaling of computational cost as the number of atoms increases. AI-techniques, especially graph neural networks (GNNs), have made headway ascribable to their ability to approximate chemical simulations with significantly lower computational cost, particularly as the system size increases. This holds tremendous promise in using AI-enabled simulated techniques to replicate materials systems of greater complexity and applicability to modern technological and societal challenges, such as climate change.
As part of this engagement, Intel and Mila will collaborate on developing scientific and technological innovations to improve the performance of GNNs on atomistic simulations, like the Open Catalyst dataset. These efforts can democratize researchers’ ability to engage with atomistic material data by enhancing the technology pipeline related to it. The research teams will work to create learning-based frameworks to improve searches within materials design applications. These frameworks can draw upon ideas from reinforcement learning, search algorithms, generative models, as well as other machine learning algorithms, including generative flow networks pioneered by Mila researchers. The primary focus of this research track will be on algorithm development for materials design challenges, as well as ecosystem development by creating toolkits for common challenges researchers encounter.
Additionally, Intel and Mila teams will apply Natural Language Processing-based (NLP) techniques to journal articles, papers, websites and/or patents. This will enable researchers to harvest and apply technical knowledge contained in text so they can discover novel material systems, understand the various physical and chemical phenomena involved in materials synthesis to speed up discovery routes, and develop methods to create structured data from unstructured data. The knowledge discovered from this line of research will provide valuable synergies for the aforementioned research efforts to significantly advance AI-driven materials design.
Causal Machine Learning for Climate Science
Standard physics-based climate models can help predict the effects of climate change, but they are complex and computationally expensive. They often take months to run — even on specialized supercomputing hardware — which reduces the frequency of simulation runs and the ability to develop granular, localized predictions.
Intel believes that global climate change is a serious environmental, economic and social challenge that warrants an equally serious response by governments and the private sector. It has actively taken steps to reduce its own environmental footprint for years and recently announced plans to achieve net-zero greenhouse gas emissions in its global operations by 2040. With today’s announcement, Intel and Mila are joining forces to further address this challenge via the development and application of new and advanced AI techniques. With this collaboration, Intel and Mila will build a new type of climate model emulator based on causal machine learning (ML) to identify which variables are predictive out of extremely high-dimensional inputs to traditional climate models. The teams will begin with probabilistic predictions of regional rainfall and temperature, learning from ensembles of climate model simulations in the CMIP6 data archive. The next step will be to develop more sophisticated algorithms for working with regional climate drivers such as land cover changes.
Ultimately, the project will enable significant advancements in climate science and directly inform policy by enabling local and regional predictions of the effects of climate change. The project will also advance cutting-edge causal ML due to the vast number of relevant variables and complex interrelationships between them, with a relatively large number of causal connections and numerous confounders.
Digital Biology: Accelerating the study of molecular drivers of diseases and drug discovery
Biology is an exciting frontier in natural sciences. Now more than ever before, it is coming within the purview of computing with availability of high-resolution data, advent of breakthroughs in AI, and growth in compute density driven by Moore’s law. The time is ripe to finally usher in the era of precision medicine, where we learn from data of all to bring benefit to one and everyone.
To realize this vision, Intel and Mila will develop AI techniques to:
- Understand molecular drivers behind diseases. Predicting complex phenotypes including diseases based on genotype of single-nucleotide polymorphisms (SNPs), formally known as polygenic risk score (PRS) prediction, has been a long-standing challenge in digital biology. As most phenotypes are affected by many SNPs across the genome, the main computational challenge is jointly learning the causal effects of all the SNPs in the genome on the phenotypes, using large-scale population data. With millions of SNPs detected so far (e.g., UK Biobank dataset), an exact solution is computationally intractable.
- Identify the most promising drug molecules. It takes more than 10 years and $2.5 billion to develop a new drug and identify the most promising target drug molecules. This presents a big opportunity for ML-based methods. With the amount of data already generated in the biomedical domain and represented as graphs (e.g., scientific literature, known drug molecules, drug-protein bindings, etc.) there’s an opportunity to leverage open-source ML frameworks for drug discovery. For example, TorchDrug, https://torchdrug.ai/Mila’s ML framework for drug discovery based on Graph ML, Deep Generative Models, and Reinforcement Learning, can be used to significantly cut down on the cost and time to market for drug development.
These and many such problems of broad scientific interest are computationally intractable if we attempt to find exact solutions but can be framed as learning problems: discovery of gene regulatory networks as causal discovery and drug discovery as active learning. Using this approach, Intel and Mila will jointly take on these challenges to build novel high performance for 1) causal AI methods and 2) AI algorithms to identify promising drug molecules and usher in the long-promised era of precision medicine.
Key Contributors:
Intel Labs: Gal Novik, Bharat Kaul, Santiago Miret, Sanchit Misra, Somdeb Majumdar
Mila: Professor Yoshua Bengio and research team
Notices & Disclaimers
Your costs and results may vary.
Intel does not control or audit third-party data. You should consult other sources to evaluate accuracy.
Statements in this document that refer to future plans or expectations are forward-looking statements. These statements are based on current expectations and involve many risks and uncertainties that could cause actual results to differ materially from those expressed or implied in such statements. For more information on the factors that could cause actual results to differ materially, see our most recent earnings release and SEC filings at www.intc.com.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.
