Neural Network

This is a network containing four layers:
Input(14 * 14 = 196),
Layer1(16),
Layer2(16),
and Output(10)

Therefore, there are 3594 params in total.

This network is as rough as what was built half a century ago, the training process is simple with nothing else than trials and adjustments.

Current progress:
Loss 0.4066,
Accuracy (On training dataset) 75.0%

Scratch?!

Yes, the recognizer supports running on Scratch, giving the most classical and primitive coding experience and running efficiency same as what pioneers have experienced 50 years ago... :|

The trainer works on Scratch as well, theoretically, but it's just too slow for any human to wait.

Building projects on Scratch sometimes makes people feel a bit retro... and magical.

Extras

This project is just a messy demo drafted occasionally one weekend at which I watched But what is a neural network? | Chapter 1, Deep learning and decided to build a neural network myself.

It was like unboxing some random Christmas presents when seeing params being trained...

=￣ω￣=

Try: Portal to ccw.site!

Codes: On GitHub!

1998

What if back to 1998? The "Internet" of the world is down due to an accidental natural disaster, putting the network which used to link every corner of this planet in serious danger.

Digital の pioneer "goose", who was entrusted with the crisis, offered to fall into the vortex of "time and space", then slowly woke up in one of the "1998"s in the stream of history......

Growth is also lost:

Installing drivers (gaining new weapons / abilities) takes up storage space (reduces the maximum health point) The goose, you, have to choose between damage power and fault tolerance...

But we'll make sure it is possible to beat this game with only basic weapons... (probably?)

More about

This game is an entry of CCW 2024 Scratch Game Jam. Only 14 days were given to build the entire game, requiring a start from building the basics, including coding the physics engine.

It was a great time.
:O

Play: Portal to ccw.site!