Method and apparatus with neural network processing

An apparatus with neural network processing, the apparatus comprising:

• one or more processors configured to:
• determine whether to refresh a population based on at least two of a time, a size of the population, a non-convergence of the population, and a local optimization of the population;
• determine, in response to the determination to refresh the population, whether a portion of the population satisfies convergence criteria, the population comprising comprises a plurality of instances each having a different mixed-precision quantization applied to layers of a neural network;
• generate, in response to the determination that the portion satisfies the convergence criteria, a new instance by determining a precision for each corresponding layer of the new instance based on an average of precisions of the plurality of instances comprised in the portion;
• update the population by adding the new instance to the population where the updated population comprises instances that represent preferred trade-offs among a plurality of objectives for the neural network;
• select from the updated population, based on at least one objective among the plurality of objectives, an optimized instance to which an optimized mixed-precision quantization of the neural network is applied for the neural network;
• and
• perform an inference based on the neural network using the optimized instance selected from the updated population, the optimized mixed-precision quantization enables the at least one objective during the inference.

A processor-implemented method with neural network processing, the method comprising:

• determining whether to refresh a population based on at least two of a time, a size of the population, a non-convergence of the population, and a local optimization of the population;
• determining, in response to the determination to refresh the population, based on a convergence criterion comprising a dispersion of a predetermined threshold, a plurality of instances having converging features among the population comprising a plurality of instances to which different mixed-precision quantizations are applied for a neural network;
• generating a new instance based on a representative value determined based on precisions of the determined plurality of instances corresponding to a same layer of the neural network;
• updating the population by adding the new instance to the population, the updated population comprises instances that represent preferred trade-offs among a plurality of objectives for the neural network;
• and
• performing an inference based on the neural network using an instance selected from the updated population.