Multi-Objective NAS
An architecture search that optimises accuracy alone will cheerfully hand you a network too slow for the phone it was meant to run on. Multi-objective NAS searches with the other costs written into the objective instead of checked afterwards.
Trading accuracy against cost
The second objective is whatever actually constrains deployment: measured latency on the target device, parameter count, FLOPs, memory footprint, energy per inference. Two ways of combining them are common. A scalarised reward folds the objectives into one number — MnasNet multiplies accuracy by a latency penalty, tuned so the search is rewarded for staying near a target runtime. True Pareto methods instead keep a population of architectures and return the whole frontier: every design for which no other is better on both axes. Evolutionary search suits that second approach, since a population is already what it maintains.
Details that decide whether it works
- Measure, do not model — FLOPs correlate poorly with wall-clock latency, so the leading methods time candidates on the real hardware or on a lookup table built from it.
- One frontier, many products — a single Pareto run supplies a phone model and a server model without searching twice.
- Hardware specificity — a frontier found for one accelerator does not transfer to another.
Why it matters
Most of the networks shipping on consumer devices come out of searches like this, because the constraint is not negotiable: a design that misses the latency budget is not a worse option, it is not an option. Putting efficiency into the objective also turns a hardware limit into a visible design choice, priced in accuracy.
Multi-objective NAS searches neural architectures against accuracy and deployment cost together — latency on the target device, parameter count, FLOPs, or energy — rather than filtering for efficiency after the fact. It either scalarises the objectives into a single reward, as MnasNet does, or returns a Pareto frontier of designs that cannot be improved on one axis without losing on another. It matters because a model that misses its latency budget cannot ship.
Buying a car, nobody optimises one number. You want speed, fuel economy and a sensible price, and improving one usually costs you another. What you really want is the set of honest best trade-offs, so you can pick the one that suits you.
Multi-objective NAS designs AI models that way. Instead of hunting only for the most accurate network, it searches for networks that are accurate and also fast enough, small enough and cheap enough to run on the device they are meant for — a phone, a watch, a doorbell camera. The result is a menu of good compromises rather than a single winner.