Benith Inc.

Benith Research

Benith Research is an experimental machine-learning effort focused on discovering neural architectures through evolutionary search over primitive operations. The current system builds and evaluates graphs composed from reductions, projections, local window extraction, normalization, and tensor transforms, then reruns promising candidates for validation.

What This Is

This is a small independent research effort under Benith Inc. The goal is to develop a search process that can repeatedly discover multiple strong model families, let them compete in the same runs, and transfer that process to stronger sequence and language-modeling tasks.

Current Results

On MNIST, the system has already produced search-discovered models that outperform a conventional LeCun-style reference. These runs were trained on the full MNIST training set and evaluated on the full 10,000-example MNIST test set.

Reference LeCun-style baseline Test accuracy: 6170 / 10000 Loss: 1.84442 Observed test duration: 4.18696s
Main discovered winner Validated search result Test accuracy: 7607 / 10000 Loss: 1.72265 Observed test duration: 1.78694s
Separate strong family Im2Col / reduce chain Test accuracy: 7624 / 10000 Loss: 1.74111 Observed test duration: 1.79837s
Compact validated family Low-memory discovered model Test accuracy: 6666 / 10000 Loss: 1.80351 Observed test duration: 0.78948s

Why More Compute

The bottleneck is experiment volume and validation throughput. More GPU access would directly increase the number of candidates we can evaluate, the number of promising models we can rerun to full validation, and the speed at which we can test improved mutation and selection strategies.

Contact

This page is a lightweight public summary for program applications and technical outreach. Benith Inc. is the legal entity behind the work. Contact: timprepscius@gmail.com.