Project summary

Factity is a not for profit organization with a mission to make ai systems harmless, humane and honest. We have built several open source software tools that are aiding in tackling the problem of ai safety. From benchmarking model capabilities across domains, OCR based ai safety tools, Documenting key questions in ai safety and gathering public perception of ai etc. We have built and deployed several data sets, packages, projects etc. 

We are seeking a total of $ 20000 to further our effort in our most recent endeavour. We must raise at least a total of $ 12000 to continue further with the following project. 

Note: - The funding allocated will be directed particularly towards building a library and further documenting the under appreciated field of lipschitz neural networks.

Links 

Project docs - https://factity.github.io/fairness/draftnormal

About me - https://factity.github.io/fairness/love

Key problem we are tackling:

We believe that normalization is so fundamental in the training of any ML based models that significant efforts should be directed towards optimizing at each and every step of the training process. We believe the below presented approach will be promising and can have significant impact on pretraining of llms. It also makes the models robust against adversarial inputs.

Normalization is one of the important aspects of training a transformer based neural network. If the weights, gradients become too large or too small the training becomes almost impossible and the accuracy steadily decreases. The vanishing or exploding gradients can be tackled using a particular set of approaches one of the promising approaches is to map the weight matrices i.e, the internal weights onto manifolds and essentially constraining them and using retraction steps while optimizing to map them back essentially constraining the weights to the manifold. I have explored some of the important approaches related to different promising manifolds and distance metrics I intend to develop and expand upon libraries that essentially establish pypi packages for the effort

What are this project's goals? How will you achieve them?

• Studying different suitable manifolds their and advantages and disadvantages for different networks 

• Exploring different distance metrics like Geodesic distance – general Riemannian manifolds; Chordal distance – Grassmann manifold, Procrustes distance – Stiefel manifold, Bures–Wasserstein distance, Log‑Euclidean distance 

• Exploring different ways to make the entire network lipschitz 

• Developing a library or expanding existing libraries

• Developing experiments to test and compare different manifolds 

There is a significant gap in the literature both academic and open. We believe that research related to ai should be open and should have exposure from the wider scientific community 

How will this funding be used?

• Maintaining and developing the libraries 

• Marketing and Distributing the software 

• Experimental costs for cloud infrastructure 

• Rent and stipend 

• Improvements and bug fixes 

Who is on your team? What's your track record on similar projects?

https://mukullight.github.io/mukul-portfolio/

What are the most likely causes and outcomes if this project fails?

We believe the negative implications of the following work will be minimal as the field is growing fast. The work has far reaching implications in Maths, scientific models, Machine learning etc. 

This work is invaluable to large scale research labs. 

The possible outcomes of a failure include still having an repository but very few resources allocated towards maintenance. We are very motivated and believe the problem is important we are happy to spend our own time to ensure the software and the community expands and grows  

How much money have you raised in the last 12 months, and from where?

The following filed of research is very new and not yet explored we did not raise funds yet