Return to Colloquia & Seminar listing
PhD Exit Seminar: inverse neural network problem
Special Events| Speaker: | Qianhui Stella Wan |
| Location: | ZOOM |
| Start time: | Wed, Sep 11 2024, 4:00PM |
We explored the inverse problem of parameter recovery in black box neural networks. Specifically, we investigated a layer-by-layer approach for learning parameters in single-hidden-layer feedforward neural networks, with focus on rectified/monotone and binary activation functions. With input data under Gaussian distribution assumption, we can successfully leverage matrix factorizations and other optimization schemes to split the networks as layers. In addition to providing theoretical recovery guarantees for parameter recovery, we also explored the algorithm design and tested them onsmall-scale networks. I will also discuss the future directions of this work in the talk.
Join Zoom Meeting https://ucdavis.zoom.us/j/2057581328?omn=95818987120 Meeting ID: 205 758 1328