Preliminary Pentachoron Experiments
This is a series of baseline tests to determine the potential capabilities of pentachoron shapes.
The earliest forms have shown high-yield expansion and crystalization potential for rapid learning, which allows them to manifest complex behavior in nearly no steps on CPU training. These earlier forms are low accuracy in comparison to the later forms, but I need to expose them nonetheless for full transparency. This structure evolved from one that did not work very well, into one that works predominantly accurately with oftentimes 80x smaller param counts.
The later tests not shown here are displaying unique behavioral traits akin to crystalized silicone; and I have begun experimenting heavily on crystalizing forms of these into extremely compact intelligent clusters in single epochs.
The accuracy suffers for these smaller variations; but after only a single epoch the variances are showing that not only can they diversify the entirety of MNIST at 99.95%, but they can contain the entire diversity of MNIST Fashion at 94.6% alone - and the combination of MNIST fashion + MNIST digit at 96.4%~~ I don't remember that exact one but it's > 96% accuracy.
The experiments are ongoing and the upscaled collectives are showing extensive potential with upscaled variants. Due to the rapid training speed of these geometric shapes, they can be rapidly trained and discarded at runtime - resulting in self learning and self crystalizing classification for high yield behavioral responses.
The more complex the latent the more complex they can learn the geometric structure in conjunction with other structures, while still retaining the condensed and compact shapes.
More yield soon
I have many notebooks and many trainings for this structure that have all yielded high accuracy, many of which contain many unique geometric loss methods and structural responses to those losses that are formed directly through goemetric constraints using collapsed and flattened images without required dimensional and relational behavior.