Ph.D. in Mathematics, So far, I am posting animations of the gradient ascent process for word embeddings, graph embeddings, hypergraph embeddings, dimensionality reductions, and other machine learning algorithm. All fitness functions featured in these animations are my own. Some of the content here is cross posted on TikTok; since TikTok is good for shorter videos, my videos will tend to be short as well. Mathematical descriptions of the algorithms involved here can be found at circcashcore.org or at github.com/jvanname

I am available to consult your business on the use of AI (at a rate between 60 USD/hr and 100 USD/hr), so feel free to contact me if you want to use AI more safely and effectively. You can also support my work on AI safety and interpretability research at https://github.com/sponsors/jvanname.

Spectral radii dimensionality reductions training without gradient ascent and inheriting symmetry
2:38
Spectral radii dimensionality reductions training without gradient ascent and inheriting symmetry
Producing spectral radii dimensionality reductions without gradient ascent
4:03
Producing spectral radii dimensionality reductions without gradient ascent
Spectra from spectral radius dimensionality reductions inheriting properties of training data
2:30
Spectra from spectral radius dimensionality reductions inheriting properties of training data
Spectra from spectral radius dimensionality reductions inheriting properties of training data
2:10
Spectra from spectral radius dimensionality reductions inheriting properties of training data
More discrete optimization for L2 spectral radius dimensionality reductions of tuples of matrices
1:33
More discrete optimization for L2 spectral radius dimensionality reductions of tuples of matrices
Discrete optimization for L2 spectral radius dimensionality reductions of tuples of matrices
1:54
Discrete optimization for L2 spectral radius dimensionality reductions of tuples of matrices
The spectrum of a 600 by 600 complex matrix as it is being tropicalized.
1:20
The spectrum of a 600 by 600 complex matrix as it is being tropicalized.
Eigenvalues of sparse random matrices
2:25
Eigenvalues of sparse random matrices
Singular values from multilinear models with matrix outputs training on random data
2:51
Singular values from multilinear models with matrix outputs training on random data
Accuracy plot while training completely positive superoperators to compute non-linear functions.
2:00
Accuracy plot while training completely positive superoperators to compute non-linear functions.
Spectra of traceless completely positive superoperators being trained on symmetric data
2:11
Spectra of traceless completely positive superoperators being trained on symmetric data
Spectra during the training of traceless complex symmetric matrix valued- word embeddings
2:33
Spectra during the training of traceless complex symmetric matrix valued- word embeddings
Training a real vector to maximize orthogonality with Hermitian matrices
1:11
Training a real vector to maximize orthogonality with Hermitian matrices
Evolutionary computation produces the same -1,1-vector x with Ax positive each time.
2:14
Evolutionary computation produces the same -1,1-vector x with Ax positive each time.
Computing graphs establishing lower bounds for the Ramsey number R(5,5)
6:23
Computing graphs establishing lower bounds for the Ramsey number R(5,5)
Spectra from real matrices being trained on complex matrices
1:33
Spectra from real matrices being trained on complex matrices
Spectra from complex anti-symmetrizations of Jacobians of the octonions
11:55
Spectra from complex anti-symmetrizations of Jacobians of the octonions
Spectra from complex symmetrizations of Jacobians of octonions
5:12
Spectra from complex symmetrizations of Jacobians of octonions
Training a transformation of more complex matrices into Hermitian matrices
1:10
Training a transformation of more complex matrices into Hermitian matrices
Training a transformation of complex matrices into Hermitian matrices
0:50
Training a transformation of complex matrices into Hermitian matrices
Illegible adversarial examples of handwritten digits during training
2:57
Illegible adversarial examples of handwritten digits during training
Training an AI by maximizing a discontinuous fitness function multiple times with similar results
2:40
Training an AI by maximizing a discontinuous fitness function multiple times with similar results
Starfish shaped spectra of superoperators of matrix dimensionality reductions during training
1:20
Starfish shaped spectra of superoperators of matrix dimensionality reductions during training
Mercedez-Benz shaped spectra of superoperators of matrix dimensionality reductions during training
1:20
Mercedez-Benz shaped spectra of superoperators of matrix dimensionality reductions during training
Triangular spectrum of a superoperator of a matrix dimensionality reduction during training
1:20
Triangular spectrum of a superoperator of a matrix dimensionality reduction during training
Recovering -1,1 matrices from the phases of their Fourier transforms using inverse transforms
3:18
Recovering -1,1 matrices from the phases of their Fourier transforms using inverse transforms
Training a binary classifier 7 times and getting the same thing 6/7 times
1:30
Training a binary classifier 7 times and getting the same thing 6/7 times
An AI training with a discontinuous fitness function produces the same model 5/7 times.
1:43
An AI training with a discontinuous fitness function produces the same model 5/7 times.
Neural networks learn the sine function using evolutionary computation.
2:00
Neural networks learn the sine function using evolutionary computation.
Images of handwritten digits being recovered from the Fourier transform phases
6:52
Images of handwritten digits being recovered from the Fourier transform phases