MIT boffins cram ML training into microcontroller memory

Researchers claim to have developed techniques to enable the training of a machine learning model using less than a quarter of a megabyte of memory, making it suitable for operation in microcontrollers and other edge hardware with limited resources. The researchers at MIT and the MIT-IBM Watson AI Lab say they have found " …

[Anonymous Coward]

Why the focus on microcontrollers?

If it is true these algorithmic improvements can lower both training times and resource consumption by an exceptional amount while preserving accuracy, wouldn't that also be true for all sizes of models on all devices?

[Old Used Programmer]

256KB, you say?

That suggests that it would run on a Raspberry Pico, or other RP2040 board.

[Arthur the cat]

as little as 256KB of SRAM

Luxury!

When I wer't lad …

[IanRS]

Re: as little as 256KB of SRAM

... you had to hand carve the ferrite cores?

[Arthur the cat]

Re: as little as 256KB of SRAM

Ferrite? We used to dream of ferrite!

[heyrick]

I wonder what was the bigger impediment

The small amount of memory or the low clock speed?