Machine Learning Metallurgi – Assume atoms are functions. Then you could train the iron atoms with coal in an add and removal procedure. Add a large amount of coal then remove it. Assume some function still remain. // Per
Machine Learning Idea – Hyper Convolutional Layers. Train a model with ridiculus high settings on conv2D then use it as augmented label data for a bypass lower grade convolution after some middle layers. Wow HyperConvolution // Per
Made a conversion model with 300 hidden size and convolution as augmented third layer label data. Wow 99.04% on MNIST with only fc layers. World record? // Per Lindholm
Machine Learning Breakthrough! – Train a large Perceptron like model with convolution bypass. Wow got 99.05% MNIST. 2000,300 size batch 100. To train large matrices use convolution then bypass it. // Per
99.0% on a raspberry pi 4 2G 32GB zram (needed) MNIST
Super Intelligence Physics – What is the best walk ever. The transistor circuit walk. So the electron in the atom does something similar. It performs calculations mapping out a transistor circuit. // Per
Super Intelligence Physics – Ohm’s temperature function could lead to super low resistance. Measure the resistance, voltage, current and temperature for a circuit. Assume with a u,r,i = model(…) you can lower the temperature of the circuit given variation in the variables.
Salmonella vs Machine Learning Experiments – Insert an error like salmonella in the process. Classify or train on these inserted errors with a machine learning model(). Run the train model on food line without inserted errors and classify similar risks or problems. // Per
ovid19 Idea – Decision tree learning for cooking a treatment. Boil in several pots. Measure. Mix portions % of each. Decision tree cooking. Create a machine learning robot cooker for this 100+ pots. // Per #COVID19
Revolutionary Machine Learning Physics – What if every facet matter area is a matrix function that calculates the node point movement. So to move something matter has to calculate its new position. Since it also works and is more detailed. Why not this explanation. Mega Wow //Per
Machine Learning Physics – Sun physics – Assume everything in the sun belongs temperatures. So there must exist T = f()*g() force temperature acceleration laws. // Per
Machine Learning Physics – Gravity – Why has the earth such a big radius. My guess is that its because the correlation coeffient somewhere between -1..1. The error outside that range is another planet interfering with earths matter. To gravity belongs stochastic variables. //Per
Machine Learning Thinking Math – When solving a math problem with an optimizer. That is. With a machine learning model function. You can create a thinking process inbetween the sample feed. Could perhaps give insight into numerical problem solving. // Per
Here thinking means have another model2() with X2 and yt2 as output for the main model(…). Like a simulation where the input and output is unknown but adapted to overal problem loss function.
Machine Learning Insight – Adapt to a “fake” loss reduction. The model will adapt to the loss reduction to maintain clearity. // Per
Just try loss = F.mean_squared_error(y,yt) then loss = loss/(j+10) so its a reduction that came from nowhere. So maybe the reduction should be supervised aswell.