Authors: Flavio Martinelli, Berfin Simsek, Wulfram Gerstner, Johanni Brea
Published on: April 25, 2023
Impact Score: 8.12
Arxiv code: Arxiv:2304.12794
Summary
- What is new: The ‘Expand-and-Cluster’ method uniquely identifies neural network parameters despite known challenges.
- Why this is important: Identifying neural network parameters just from input-output mappings is difficult due to permutation, overparameterisation, and activation function symmetries.
- What the research proposes: A two-phase ‘Expand-and-Cluster’ method that first expands the student networks to imitate the target network, then uses clustering to identify shared weight vectors.
- Results: Successful recovery of parameters and size of shallow and deep networks with less than 10% neuron number overhead, plus an ‘ease-of-identifiability’ analysis of 150 synthetic problems.
Technical Details
Technological frameworks used: nan
Models used: Expand-and-Cluster method, shallow and deep neural networks
Data used: 150 synthetic problems
Potential Impact
AI development and analytics services, companies relying on proprietary neural network architectures
Want to implement this idea in a business?
We have generated a startup concept here: NeuraMatch.
Leave a Reply