Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a conference talk examining quantum acceleration methods for deep learning inference in multilayer convolutional neural networks. Discover how fault-tolerant Quantum Processing Units (QPUs) can potentially deliver exponential speedups for pre-trained networks that output probability distributions over classes or tokens. Learn about the integration challenges of quantum computing into classical deep-learning pipelines and understand which network architectures can benefit from quantum acceleration under different Quantum Random Access Memory (QRAM) assumptions. Examine fundamental residual blocks composed of regularized multi-filter 2D convolutions, sigmoid activations, skip-connections, and layer normalizations that echo ResNet structure. Analyze three distinct computational regimes: regime 1 where both input tensors and network weights utilize QRAM achieving O(polylog(N)^d) inference cost, regime 2 maintaining only weights in QRAM while requiring O(N) input loading cost yet enabling quartic speedups for shallow bilinear networks, and regime 3 operating without QRAM where quantum advantages must exploit high-dimensional latent feature transforms. Understand the end-to-end complexity bounds that account for both input and memory assumptions, with practical applications including multi-turn LLM chat scenarios with repeated calls to slowly-changing inputs. Gain insights into why known dequantization methods may not apply to certain quantum acceleration approaches and explore the limitations and opportunities for quantum speedup in terms of input dimensions and network parameters.
