Embedded Systems

Extension of our neural network framework HANNAH

Bach­e­lor’s The­sis / Mas­ter’s The­sis / Stu­dent Re­search Pro­ject

Ab­stract

The chair has its own frame­work called HAN­NAH (Hard­ware Ac­cel­er­a­tor and Neural Net­work seArchH) for sen­sor pro­cess­ing tasks (voice ac­tiv­ity de­tec­tion, key­word spot­ting, human ac­tiv­ity de­tec­tion, atrial fib­ril­la­tion) using dif­fer­ent neural net­works (TC-Resnet, Sinc­Net, Branchynet, Wavenet, LSTMs,…). HAN­NAH has the abil­ity to ex­tract dif­fer­ent fea­tures (Spec­tro­gram, MFCC, Mel Fea­tures), quan­ti­za­tion (weights, bias, ac­ti­va­tion) using Ner­vana Dis­tiller, ad­vanced noise han­dling and many more. The frame­work is built on Py­Torch, Py­Torch Light­ning and Ner­vana Dis­tiller. For the train­ing of the neural net­works we have a clus­ter with 160 Geforce GTX1080Ti or some local ma­chines equipped with Tesla P100.

Cur­rent top­ics:

  • Im­ple­men­ta­tion and analy­sis of new fea­tures pre­pro­cess­ings
  • Hard­ware-Fea­tures: In HAN­NAH we are using float­ing point for fea­ture ex­trac­tion but in our hard­ware ac­cel­er­a­tor a fixed point Fourier Trans­for­ma­tion is used. This leads to a dif­fer­ent be­hav­iour of the same neural net­work. The task would be to im­ple­ment the fea­ture ex­trac­tion in HAN­NAH like it is in hard­ware.
  • Im­ple­men­ta­tion and analy­sis of new neural net­works.
  • In­te­grate our Ul­tra­Trail de­ploy­ment in TVM
  • GUI for easy con­fig­u­ra­tion and vi­su­laza­tion of the re­sults.

We also wel­come your own ideas.

Re­quire­ments

  • You should have basic knowl­edge of Python
  • Knowl­edge of neural net­works, Py­Torch, quan­ti­za­tion and sig­nal pro­cess­ing is ben­e­fi­cial but not nec­es­sary.

Con­tact

Bring­mann, Oliver

Gerum, Christoph