Reproduce “Function regression using the forward forward training and inferring paradigm” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shivam Padmani, Akshay Joshi
Researchers at the Indian Institute of Science adapted Geoffrey Hinton's Forward-Forward algorithm for function regression, demonstrating its capability to approximate 1D, 2D, and 3D functions by recasting regression as a binary classification task. The methodology successfully yields predictions with confidence intervals, highlighting a peculiar "inversion of goodness" phenomenon and providing initial insights into its application with Kolmogorov Arnold Networks.
Key insights
- Function regression can be effectively performed using the backpropagation-free Forward-Forward algorithm by transforming the regression problem into a specific binary classification setup.
- A peculiar 'inversion of goodness' was observed during inference, where points near the true function value yielded higher goodness for the 'out-tolerance' label, a phenomenon that could be reliably leveraged for prediction.
- The choice of hyperparameters like tolerance (`tol`) and the ratio of in-tolerance to out-tolerance training points significantly influences the accuracy and uncertainty of the regression results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.