New Neural Network Models Enhance Predictions in Medicine, Finance, and Genetics
Vector Institute Faculty Member David Duvenaud and his collaborators have published a groundbreaking paper titled "Scalable Gradients for Stochastic Differential Equations" in the journal Artificial Intelligence and Statistics. This research introduces a novel method using backpropagation to fit stochastic continuous-time models, effectively combining deep neural networks with stochastic differential equations (SDEs). Historically, SDEs were limited by their inability to scale to large parameter sets, restricting their application in complex data domains. By generalizing previous work on Neural Ordinary Differential Equations, the team developed an algorithm that allows for scalable gradient-based optimization of SDEs. This advancement enables the creation of more sophisticated prediction models capable of handling uncertainty and unseen interactions in continuous time. The potential applications are vast, offering improved accuracy in forecasting stock market prices, modeling medical data, and tracking human genetic evolution over time. This breakthrough addresses previous scalability limitations, allowing these models to integrate seamlessly with modern deep learning architectures containing millions of parameters, thereby enhancing predictive capabilities across physics, finance, and healthcare sectors.
Wire timeline
New Neural Network Models Enhance Predictions in Medicine, Finance, and Genetics
Vector Institute Faculty Member David Duvenaud and his collaborators have published a groundbreaking paper titled "Scalable Gradients for Stochastic Differential Equations" in the journal Artificial Intelligence and Statistics. This research introduces a novel method using backpropagation to fit stochastic continuous-time models, effectively combining deep neural networks with stochastic differential equations (SDEs). Historically, SDEs were limited by their inability to scale to large parameter sets, restricting their application in complex data domains. By generalizing previous work on Neural Ordinary Differential Equations, the team developed an algorithm that allows for scalable gradient-based optimization of SDEs. This advancement enables the creation of more sophisticated prediction models capable of handling uncertainty and unseen interactions in continuous time. The potential applications are vast, offering improved accuracy in forecasting stock market prices, modeling medical data, and tracking human genetic evolution over time. This breakthrough addresses previous scalability limitations, allowing these models to integrate seamlessly with modern deep learning architectures containing millions of parameters, thereby enhancing predictive capabilities across physics, finance, and healthcare sectors.
Vector Institute for Artificial Intelligence