Portfolio
Here I present a few of the applications I’ve created in years past. We see visualization tools, compilers, tensor libraries and Deep Learning demos. I truly seek to understand the concepts and principles beneath the mathematics and industry tools. Everything here is built from scratch.
Claude
Claude is a concrete representation of three-dimensional space. Complete with a virtual camera and built atop WebGL, Claude is robust and efficient – able to render over 10M pixels in real time. I built Claude due to the lack of competitive python visualization libraries: Matplotlib was choking on 100K pixels. Here we see a visualization of a 10MB image in pixel space:
Big-Box
Big-Box is an efficient NumPy implementation in pure JavaScript. I needed a backend for Claude. It supports complex numbers and quaternions. Here we see key benchmarks against industry leaders:
| Random |
|
| big-box: |
89.044ms |
| numjs: |
98.539ms |
| mathjs: |
710.838ms |
| Ones |
|
| big-box: |
5.786ms |
| mathjs: |
49.439ms |
| numjs: |
51.870ms |
| Zeros |
|
| big-box: |
1.285ms |
| numjs: |
37.166ms |
| mathjs: |
60.393ms |
| Range |
|
| big-box: |
14.712ms |
| mathjs: |
159.187ms |
| numjs: |
172.807ms |
| Literal |
|
| big-box: |
0.087ms |
| numjs: |
0.215ms |
| mathjs: |
1.141ms |
| Sum |
|
| numjs: |
22.922ms |
| big-box: |
24.788ms |
| mathjs: |
661.154ms |
| Mean |
|
| numjs: |
19.135ms |
| big-box: |
25.306ms |
| mathjs: |
336.405ms |
| Min |
|
| big-box: |
20.660ms |
| numjs: |
30.452ms |
| mathjs: |
158.234ms |
| Max |
|
| big-box: |
22.281ms |
| numjs: |
26.227ms |
| mathjs: |
157.730ms |
| Add |
|
| big-box: |
15.892ms |
| numjs: |
29.694ms |
| mathjs: |
665.261ms |
| Subtract |
|
| big-box: |
13.244ms |
| numjs: |
20.207ms |
| mathjs: |
704.024ms |
| Multiply |
|
| big-box: |
15.469ms |
| numjs: |
19.704ms |
| mathjs: |
370.649ms |
| Divide |
|
| big-box: |
15.184ms |
| numjs: |
19.585ms |
| mathjs: |
654.055ms |
Deep Demos
Machine Learning over the past decade has exploded. Lagging behind, however, is the ability to learn from small or unlabeled datasets. I believe breakthroughs will arise as more inspiration is lifted from biology. This has been an evident trend. Neural Networks are more brain-like than their predecessors (SVMs, Logistic Units, etc.), and they peform better. Convolutional Networks are more eye-like: better performance. Our ears execute the fourier transform, so let’s build some ear models on top of it. Below is a visualization of the top 100 fourier coefficients over time for simple utterances:
