The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Probabilistic Numeric CNNs with Roberto Bondesan - #482
May 10, 2021
Roberto Bondesan, an AI researcher at Qualcomm, shares his groundbreaking work on probabilistic numeric CNNs, which leverage Gaussian processes for enhanced error correction. He delves into innovative adaptive neural compression techniques that optimize data transmission efficiency. The conversation also touches on the exciting intersection of quantum computing and AI, where Bondesan discusses the future potential of combinatorial optimization in revolutionizing logistics and design. His insights bridge physics and advanced AI applications, highlighting a promising frontier in technology.
AI Snips
Chapters
Transcript
Episode notes
Roberto's Path to AI
- Roberto Bondesan's background is in physics, where he applied deep learning to problems like characterizing phases of matter.
- This led him to AI research at Qualcomm, inspired by the spherical CNNs paper by colleagues Taco Cohen and Max Welling.
Probabilistic Numeric CNNs
- Probabilistic numeric CNNs address deep learning applications for signals sampled non-uniformly, like irregular time series.
- They consider continuous signal formulations, quantifying uncertainty from discretization, unlike traditional discrete approaches.
PDE Layer Innovation
- Probabilistic numeric CNNs use a novel convolutional layer defined by linear partial differential equations (PDEs).
- This PDE layer allows the network to operate on continuous functions rather than discrete vectors, enabling analytical computations.