What is it about?
Deep learning and quantum computing both attracted enormous attention around 2019, and combining them seemed natural. This paper examined whether the combination is as straightforward as the enthusiasm suggested. The core tension: deep learning works by training networks with many parameters on large datasets, exploiting redundancy and overparameterization. Quantum computing operates under strict physical constraints -- decoherence limits circuit depth, qubit counts are small, and measurement collapses superpositions. These constraints directly conflict with the requirements of deep networks. We analyzed specific barriers: the limited connectivity of qubit architectures, the overhead of encoding classical data into quantum states, and the difficulty of extracting gradients from quantum circuits. The paper argued that hybrid classical-quantum approaches, rather than fully quantum neural networks, were the more productive research direction.
Featured Image
Photo by BoliviaInteligente on Unsplash
Why is it important?
During the quantum machine learning hype of 2018-2019, many proposals assumed that quantum speedups for linear algebra would translate directly into speedups for deep learning. This paper identified the physical reasons why that translation is not straightforward. The constraints we identified -- limited qubit connectivity, measurement overhead, decoherence-limited circuit depth -- remain relevant. They apply to current NISQ hardware and will persist until fault-tolerant quantum computers become available. The paper's recommendation to focus on hybrid approaches (classical preprocessing, quantum subroutines for specific operations) anticipated the direction the field has since taken.
Perspectives
I wrote this paper from a desire to think carefully about quantum computing and machine learning together, rather than assuming the combination would automatically work. It followed directly from my earlier quantum computing work on distributed architectures. The argument is essentially about resource budgets: quantum systems have strict limits on width (qubits), depth (before decoherence), and measurement (collapsing states). Deep learning consumes all three resources liberally. Reconciling these budgets requires architectural compromises that reduce the potential quantum advantage. The paper is a position piece, not a proof of impossibility. Quantum machine learning may well become practical, but the path runs through hybrid architectures, not through direct implementation of classical deep networks on quantum hardware.
Rohit Goswami
University of Iceland
Read the Original
This page is a summary of: Qubit Network Barriers to Deep Learning, December 2019, Institute of Electrical & Electronics Engineers (IEEE),
DOI: 10.1109/wrap47485.2019.9013687.
You can read the full text:
Contributors
The following have contributed to this page
