When we started @PropheticAI one of the biggest concerns raised by investors was why there weren't more people doing consumer tFUS - Fred get's it.
"Ultrasound neuromodulation has the ingredients necessary for a near-term mainstream consumer product."
We agree.
We’re pleased to be recognized by the San Francisco Design Week Awards. The Halo received an honorable mention for the Emerging Technology category last month.
@SFDesignWeek awards celebrate designers whose work contribute towards a positive future for society.
https://t.co/fTkRGYGvl5
Interesting hypothesis on Lucid Dreaming by Neuroscientist J. Allan Hobson.
The dorsolateral PFC is brain region we’ve performed neuromodulation on the most during waking-state.
Right now, we are working on REM sleep detection (EOG) so we can stimulate the dlPFC during REM.
Is there a name for this UI pattern, where it "stacks" as you go through nested elements? This example is from https://t.co/NLLbyNadAa. I'm hoping someone out there has some code editor that can do this 👀
Here’s a look at our full closed-loop system:
EEG -> Morpheus -> tFUS 🔄
Given an initial brain state, the transformer generates spatiotemporal sequence of neuromodulation to increase gamma frequencies.
This system adapts as the brain changes adjusting the tFUS targets in real time. Each click you hear is a token generated activating a different tFUS channel.
Prophetic has successfully engineered a focused ultrasound system and has started neurostimulation.
On April 16th, Prophetic's co-founders @EricWollberg & @weslouis_ were the first people to receive stimulation from an internally designed and manufactured ultrasound transducer.
Tomorrow there will be an announcement.
On Friday, we will be posting an expanded vision for Prophetic.
And every day thereafter will be dedicated to realizing that vision.
At @PropheticAI, we were training our transformer encoder in isolation. It was generalizing around 30-40% accuracy for 6 classes.
This paper on grokking (https://t.co/fl30URpr7b) discusses creating two optimizers and using a weight decay. We did this and found it drastically improved the models ability to generalize to above 80% accuracy.
Then we decided to train up to 50,000 epochs and found with two optimizers with weight decay the model pushed past overfitting and learned previously unattainable patterns achieving ~95% accuracy.
Amazing work by @KoyaSaito.