# Outline
- MS2: WNN Verifier
- MS3: Benchmarking
- Outlook: MS4
---
## MS2: WNN Verifier
- Added features to the WNN CLI to:
1. Deploy the EVM Verifier
2. Submit proofs to the EVM Verifier
- Components are packed into a library for easy reuse
- [Deployed](https://sepolia.etherscan.io/address/0x956e9d9fe585b93926694285e29c70dd08caa9ad) & [Tested](https://sepolia.etherscan.io/tx/0x13dbb8dabf9f142c66718f4ada7feb02c2ed09b10d1aae3a2285d3d7a0067e91) Verifier for MNIST-Small on Sepolia
- Proof Verification Cost ~ 610,000 Gas
---
## MS3: Benchmarking
- Benchmarked against EZKL and Daniel Kang's ZKML
- Measured **Prooving Time**, **Accuracy** and **Proof Size**
- Overall:
- WNN is comparable to existing DNN-based approaches in the low-accuracy regime.
- In high accuracy, the WNN are loosing the battle drastically
---
## MS3: Benchmarking
<!--
Source for this chart:
https://docs.google.com/spreadsheets/d/1R-jxOnfU-FgiWcPe7X8Piegj0H-OTTTx8e3vwWGbMCQ/edit?usp=sharing
-->
![](https://hackmd.io/_uploads/BywEOgiOh.png)
---
### Verdict: WNNs might still have their place in situations where extremely fast proving time is essential and the task is simple, perhaps simpler than image recognition.
---
## Outlook: MS4
### Improve the accuracy/performance trade off for WNN
---
## Outlook: MS4
1. Hash Function Optimisation
2. Multi-Layer WNN
3. Lookup Folding
4. DNN to select permutation
---
### Thank you! :sheep:
{"title":"ZeroGravity MS2 & MS3 Review","breaks":"true","description":"View the slide with \"Slide Mode\".","contributors":"[{\"id\":\"57638455-1234-47cb-a685-1ef1d280e798\",\"add\":1479,\"del\":2291}]"}