Unlocking New Frontiers: AI and the Sciences

0 30

On the artificial intelligence landscape, Stanford HI’s fall conference, “New Horizons in Generative AI: Science, Innovation, and Society,” highlighted the profound impact of AI on scientific exploration. Generative AI has attracted public attention for vision and language The conference was explored in depthilluminating the diverse spectrum of generative AI research from its application in the fields of science and innovation to its societal implications.

The first session of the day focused on how AI can provide new windows into the natural world to enhance human understanding. At the forefront of this intellectual odyssey were dynamic speakers pushing the boundaries of what AI can achieve in their respective fields. Aditya SheshadriAn assistant professor in Stanford’s Department of Earth System Sciences, he led the audience through his fascinating climate modeling using AI to address the uncertainties in weather forecasts. MIT PhD student Pratyusha Sharma Featured project CETI, research using machine learning to decipher the language of sperm whales. and computer scientist Alex Reeves It has opened up the language of proteins using excellent language models. Viewing protein sequences as linguistic codes, Reeves demonstrated how AI, particularly transformers, can decipher the complex structures and functions embedded in these sequences.

These speakers emphasized the transformative power of AI in shaping our understanding of science and the world around us. Look Full session hereOr read the highlights below.

Modeling the Earth’s climate

Sheshadri, an assistant professor of machine learning and atmospheric science at Stanford University, highlighted the uncertainty in climate forecasting and the limitations of current climate models.

In a recent study, she focused on the unique problem of atmospheric gravity waves, which play a significant role in Earth’s climate. Waves generated by processes such as hurricanes and air movement over mountains are challenging to model because of their large-scale nature—they can vary in size from one meter to 100 kilometers. Current climate models struggle to accurately represent these waves, leading to uncertainty in weather forecasts.

Sheshadri presented two approaches her research group has taken to improve atmospheric gravity wave modeling. The first involves replacing the traditional parameters with a neural network called WaveNet. This AI-based model has shown promising results in simulating atmospheric gravity waves.

The second approach captures a physics-based representation of atmospheric gravity waves but incorporates uncertainty using AI-based tools. Sheshadri explained how ensemble-based inversion and Gaussian process simulators can help quantify and measure uncertainties in the parameters governing atmospheric gravity wave processes.

Sheshadri concluded her talk by introducing DataWave, a collaborative gravity wave research project that integrates AI, climate models, and observations and data for a comprehensive climate research approach.

Deciphering the language of whales

Sperm whales are known for their complex social structure and communication using clicks. In the new research called Project CETI, Sharma showed how machine learning techniques and data collected in the Caribbean can help us understand and decode the communication of these marine mammals.

Collecting data on a sperm well is not an easy task. They live in the ocean and often in complete darkness. The team used tagging technology affixed to whales that had been released to record whale sounds and behaviors, though that data was limited, Sharma said.

Her team’s research shows that sperm-well communication is more complicated than previously thought. Instead of a fixed set of codas, they identified a composite code system with four independently variable characteristics (tempo, rhythm, ornamentation, and rubato), resulting in a more expressive form of communication. Sharma used visuals such as the exchange plot to show the variation and structure in whale speech.

Sharma touched on the use of predictive models to understand the structure of whale calls. Models of whale vocalizations are improved by increasing the amount of context of the inputs, revealing long-term dependencies in call structure. She also observed improvements in predictions by increasing the model’s expressiveness.

“This will probably help us understand the meaning of the whales’ sounds and maybe someday allow us to communicate with them,” Sharma said. “We hope that the algorithms and approaches we develop in the course of this project will empower us to better understand the other species we share the planet with.”

Solving protein structures

In this final talk of the session, Reeves, a computer scientist and entrepreneur, explored the application of language models in the field of biology, with a particular focus on proteins and sequences. Reeves, who holds two bachelor’s degrees in philosophy and biology from Yale and a Ph.D. in computer science from NYU, where he made important contributions to protein sequence modeling while at Meta.

Proteins, encoded by amino acid sequences, play critical roles in a variety of biological functions, from cancer treatment to plastic degradation and carbon fixation. The challenge lies in understanding the language of these sequences. Despite extensive databases of protein sequences, our understanding of their function and structure is limited.

Rivas proposed the idea of ​​looking at protein sequences as a language and used language models, similar to those used in natural language processing, to decode and extract information.

Rivas and his team trained the transformant models on a database of evolutionarily diverse proteins to see how well the models captured biological information. They found that certain points of interest in the model corresponded to the 3D structure of proteins, which could interpret protein folding.

Optimizing the models showed improved accuracy in protein structure prediction. The group introduced a model called “”.ESMFoldHe demonstrated state-of-the-art results in protein structure prediction, challenging the accuracy of existing methods such as alpha fold. This approach is faster and more efficient than traditional methods that predict protein structure by sequence, which involves searching through evolutionary processes. Databases.

One notable application of ESMFold is the rapid folding of a comprehensive database of metagenomic proteins, providing a comprehensive overview of their structures.

In addition, Rivas tapped the potential of these language models to be used for generation by designing new proteins by predicting the sequence of these language models. The experimental results suggest that the models can generalize well, and may even generate new proteins that have not been observed through natural evolution.

“New Horizons in Generative AI: Science, Innovation, and Society” held on October 24, 2023 at Stanford University. Learn more about See these speakers and the entire conference.

Source link

Leave A Reply

Your email address will not be published.

This website uses cookies to improve your experience. We'll assume you're ok with this, but you can opt-out if you wish. Accept Read More