There is no machine learning in the normal sense in the Naim streamers. But yes the firmware code is optimised for best sound performance/balance as determined by Naim before it is released.
Usually beta releases are not optimised, and so there can be an audible shift between beta and release firmware with the latter being optimised.
The code execution sound performance optimisation is undertaken for each device, as the electronics and digital underlay varies between models.
As far as noise, when digital logic gates are operated, such as in buffers, CPU, shift registers and memory small amounts of electro magnetic noise pulses are created. This is natural physics due to the rapid change of states. When aggregated up over a digital system this creates a sequence of noise bursts, almost like a rhythm, as if you were shaking out a rhythm with some maracas. (But extremely quickly). By undertaking certain machine operations at precise times, the induced rhythm of noise, or shaped noise can be optimised to interact with the audio and clocking subsystems in a way that modulates any induced artefacts to act, as as I have been told by Naim, to almost act as tone controls.