Episode Summary
Executive Summary: Max Welling discussed Qualcomm and university research spanning generative modeling, symmetry-aware graph neural networks, low-precision inference, quantum-inspired neural methods, neural augmentation, and federated learning. The central theme was hybrid AI: combining strong classical/engineering models with learned neural components to improve data efficiency, robustness, and transfer across domains, especially in wireless, medical imaging, chip design, and privacy-preserving mobile learning.
Main Topics: Hybrid AI and Neural Augmentation (Priority: 5/5): Welling framed neural augmentation as combining hand-designed scientific/engineering models with neural networks that correct or nudge them, rather than replacing them. He argued this yields better data efficiency, robustness, and domain transfer. Quantum Mechanics and Quantum Machine Learning (Priority: 5/5): He described quantum as both an intellectually interesting mathematical framework for neural nets and a likely future computing platform. He expects hybrid classical-quantum systems within about a decade, while also exploring quantum-inspired classical models. Generative Modeling, Symmetries, and Graph Neural Networks (Priority: 4/5): Current research includes unsupervised generative models for images and audio, plus symmetry-aware graph neural nets for molecular property prediction and 3D rotations/permutations. Applications in Wireless Systems and Signal Processing (Priority: 4/5): The team is applying neural augmentation to MIMO detection, error correction decoding, channel estimation, and channel tracking, where learned modules improve classical algorithms. Combinatorial Optimization and Chip Design (Priority: 4/5): He highlighted using neural methods to augment solvers for place-and-route and traveling salesman-style problems, enabling learned guidance for complex optimization and adaptation to changing constraints. Federated Learning and Privacy-Preserving Compute (Priority: 5/5): Welling sees device-centric, privacy-aware federated learning as a major future compute paradigm, with data staying on devices and model training distributed across them, potentially using noise or homomorphic encryption. Low-Precision Efficient Learning (Priority: 3/5): At Qualcomm, another major line of work is training neural networks with very limited precision to maximize accuracy while minimizing power consumption on mobile devices.
Key Arguments: Many real-world systems are too complex to model fully with either hand-written equations or data-only neural nets; hybrid approaches are a practical middle ground. Neural augmentation works by keeping the classical algorithm intact and using a neural network to correct residual errors or subtle deviations, reducing data needs and improving transfer. Hybrid systems can outperform both pure classical and pure learned approaches across data regimes: classical wins with little data, learned wins with lots of data, hybrid adapts to both. Quantum mechanics is useful not only for quantum computers but also as a mathematical language that may enrich classical neural network design. Quantum computing may arrive as a hybrid classical-quantum system rather than a purely standalone replacement, with useful systems possible within roughly 10 years. Graph neural networks benefit from encoding symmetries such as node permutations and 3D rotations, especially for molecules and manifold-based data. In optimization problems like chip placement or TSP, a neural network can learn recurring solution patterns and guide a solver step-by-step. Federated learning is attractive because it keeps data on devices, addresses privacy concerns, and can potentially create a distributed data economy. Homomorphic encryption could let models aggregate encrypted updates without exposing individual device data, though computation remains expensive. Efficient low-precision inference is strategically important for Qualcomm because edge/mobile hardware must balance accuracy and energy use.
Data Points: Podcast ranking: Second most downloaded show of the year - Sam notes the prior interview with Max Welling was the podcast's second most popular episode of the year. Time since last appearance: Just over 1 year - Welling was last on the show in May of the previous year. Quantum timeline estimate: 10 years - Welling says he would not be surprised if reasonably useful quantum computing systems exist within a decade. Model regime comparison: Small data vs. lots of data - He contrasts classical, learned, and hybrid solutions across data availability regimes, though no numeric dataset size is given. Optimization scale: Millions to hundreds of millions of components - He describes chip place-and-route as involving extremely large numbers of components on a two-dimensional canvas. Privacy signal: Noise added to model updates - In federated learning, devices may add noise before sending updates to protect sensitive information.
Pivotal Quotes: "the best of both worlds" — Max Welling: Describing hybrid solutions that combine classical models and neural networks across different data regimes. "quantum mechanics is sort of a mathematical tool" — Max Welling: Explaining his view that quantum formalisms can be useful as a language for describing complex systems and neural networks. "I think this might be the future compute paradigm" — Max Welling: Referring to federated learning and device-centric distributed training that keeps data local and privacy-aware.
Implications: The conversation suggests near-term AI progress will come less from replacing classic methods and more from hybridizing them with neural networks. That has implications for edge AI, privacy-preserving learning, quantum research, and domain-specific applications like medicine, wireless, and chip design.