Reproducible AI: How and Why?

Most LLMs are not reproducible because the underlying deep neural networks are not. Because that's something LLM creators don't care about. We do, and ours are reproducible, including our GenAI systems that use GAN.

With traditional AI, you save the weights, which represent the model. It takes a massive amount of space, and you can't reproduce the weights if you don't save them. You could if you had saved the 3-4 seeds that generates them, along with the other hyperparameters. But in practice, AI creators don't even specify the seeds, using default ones that change each time you re-run the steps to create the model. It does not help with fine-tuning, especially for models sensitive to seeds.

Yet, all you have to do is allow the user to specify the seeds of the random number generators involved, be it from PyTorch, Pandas, NumPy, Random, GPU, base Python, and whatever source of randomness (libraries) you use to create your model. First, you need a good random generator you have full control over. Better than numpy.random, which in addition is subject to updates that can lead to previous seeds not working as expected, depending on library version. See our random number generator, with infinite period and one line of code, faster and better than what's in Python and elsewhere. With more flaws found in numpy.random, using new tests.

?? Read the full article, with access to Python code on GitHub, and case studies.

A word from our sponsor: for a very fast, versatile database that can handle large AI applications in real time and works with various architectures (JSON, graph, vector, SQL and so on) more efficiently than in their native environment, I invite you to attend this event.

This hands-on workshop is for developers and AI professionals, featuring state-of-the-art technology, case studies, code-share, and live demos. Recording and GitHub material will be available to registrants who cannot attend the free 60-min session.