Near-Term Feasibility and NISQ Constraints
Today's noisy, intermediate-scale quantum devices limit qubit count, circuit depth, and reliability, which sets hard bounds on what quantum machine learning can do now.
The NISQ regime
Near-term quantum machine learning runs on noisy intermediate-scale quantum (NISQ) devices: tens to a few hundred physical qubits, no error correction, and two-qubit gate error rates around a fraction of a percent. Every gate adds error, so useful circuits must be shallow. These constraints, not algorithmic ideas, are the binding limit on the field today.
What the constraints imply
- Shallow circuits only: decoherence caps depth, so deep expressive ansatze cannot run faithfully.
- Limited qubits: problem sizes stay small, and error correction is not yet available to extend coherence.
- Restricted connectivity: two-qubit gates exist only between adjacent qubits, forcing costly swaps for distant interactions.
- Statistical readout: every value comes from many shots, so precision is bought with runtime.
Why hybrid is the near-term answer
Because full quantum algorithms need error-corrected depth that NISQ lacks, the field relies on hybrid variational methods that keep the quantum circuit shallow and push optimization to classical hardware. Kernel methods go further, using the device only to estimate overlaps. Both patterns are responses to hardware limits, not the algorithms one would choose with a fault-tolerant machine.
The training gauntlet
Even within depth limits, training must survive barren plateaus, noise-induced flattening of the landscape, and shot noise in every gradient. Error mitigation techniques such as zero-noise extrapolation reduce bias without full correction, at the cost of extra circuit runs. These measures help but do not remove the ceiling that noise imposes.
An honest near-term outlook
The realistic near-term role of QML is research and small demonstrations: exploring encodings, ansatze, and training methods, and probing quantum-native data where the case is strongest. Broad, practical advantage over classical machine learning on classical data awaits either much better error rates or fault tolerance. Claims that today's NISQ devices already beat classical methods on real tasks should be read against the constraints listed here and against strong classical baselines, as discussed in quantum advantage.