In this work two methods that are widely used to peek into the inner workings of artificial neural networks (ANN) - information theory and perturbation experiments– are compared. Both were applied to a positive control ANN and their results were contrasted. Their results were not complementary as expected. Information theory identified that information is redundant across the hidden nodes while the perturbation experiment found false positives: -nodes with little information and high impact on the output-. Each method by itself seems to be insufficient to understand how the system works and a new method should be developed.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:du-41909 |
| Date | January 2022 |
| Creators | Sella, Magdalena |
| Publisher | Högskolan Dalarna, Institutionen för information och teknik |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0101 seconds