@BetterClient – Hack Club Scrapbook

Wed, Jan 1

Day 11 of making my first AI project: Today I worked on the web UI of the project (give feedback please). I also added a small trainer for smaller models that is more accurate, I will try to fix the other one as this one is VERY slow. I also worked on adding a digit recognition model using the MNIST dataset. (the box on the bottom left pops up when you click info)

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/89f83f15-a884-4a81-9a87-608eef2fb5db-image.png

Tue, Dec 31, 2024

Day i forgot of writing ai yea: new years btw Today I spent almost 4 hours trying to figure out why the language model was either Quitting immediately or saying ksi over and over again, and I've decided to do other models for now. Also the new training seems to have broken the isbiggerthan100 model I also spent a bit of time on making a webui for the project too

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/c888eb8d-ce52-43d7-8c7a-ecfed3cc1af2-img-20241231-wa0017.jpg

Mon, Dec 30, 2024

Day 9 of writing a neural network (AI) from scratch (no dependencies): I tried to improve the model but failed miserably. The model only returns "etail" for some reason. And the model takes 30 minutes to train

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/c03fc169-5735-4324-8c83-8ba378aa80d6-image.png

Sun, Dec 29, 2024

Day 8 of writing a neural network (AI) from scratch (no dependencies): Today I added things to the model to allow it to use different activation functions (GeLu or sigmoid) And I improved the very small language model I was making! Currently the model is kinda broken, so this is all I have

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/5ec5706b-ceb9-4a76-b82a-6351dcf648b9-image.png

Sat, Dec 28, 2024

Day 7 of writing a neural network (AI) from scratch (no dependencies) Today I was away from home, but I got a bit of work done from my phone: I started the very long process of porting my code to teavm code I also finished the new trainer (hopefully)

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/4078602a-db58-48c9-b86d-5b2d945e35e5-screenshot_20241228-232801_chrome.jpg

Fri, Dec 27, 2024

Day 6 of writing a neural network (AI) from scratch (no dependencies): Today I worked on re-doing training as the old one was just too slow. The training isn't finished yet so I can't show anything: Hopefully tomorrow I have some good news about this. However I can show you how accurate the new training is already! (thats -800 delta in a single train call, thats already a 10x improvement)

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/afdff1b5-9887-4553-9ea0-b362c9795ee0-image.png

Thu, Dec 26, 2024

Day 5 of writing a neural network (AI) from scratch (no dependencies): Today I worked on creating new models for the network, and objectified it a little too

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/af56bc63-8a30-4d91-ab26-a5cdf32191a2-image.png

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/fbf763c7-b945-41ec-9f06-01f1bac8984d-image.png

Wed, Dec 25, 2024

Day 4 of writing a neural network (AI) from scratch (no dependencies): Today I worked on optimizing the network's precision: I made it so neurons will turn themselves OFF from training for a bit while they aren't having an effect on the cost. This really helps the delta for every epoch: So you can probably get away with 50 epochs This also improves training time a bit. I measured around 15 minutes(from 20 minutes) I also worked on making the network more object oriented (see image 1) And see other images for current precision. Tomorrow is the day I try to make an actual model.

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/a80007cb-f09f-4676-8c65-c1d15271e73a-whatsapp_image_2024-12-25_at_21.50.35_4c8355c7.jpg

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/2d81d841-9d2f-4556-89b9-c3972da0e71d-image.png

Tue, Dec 24, 2024

Day 3 of writing a neural network (AI) from scratch (no dependencies): Today I worked on optimizing the network's training. Here is some information on the optimizations:. Optimization #1: Before training a layer of the network, all other layer's outputs are pre calculated. Optimization #GMM0VHBBM|: Changed the order of training Now here is how much they improved training time: (all tests are done with 120 parameters & 500 training samples) Before optimization: Training crashes after 1 hour (out of memory) Optimization #1: 36 minutes Optimization #GMM0VHBBM|: 20 minutes (and improves precision by 81.76%) I am probably going to keep doing optimization tomorrow.

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/18831d46-0bd2-4a3b-9eed-3464d699867b-image.png

Mon, Dec 23, 2024

Day 2 of coding a neural network from scratch (no dependencies): I got a very basic slow training implementation working. The model in the images are a model that tries to guess whether an input number is bigger or less than 100 The model uses a sigmoid activation function. The description the texts are in the top of the images. The outputs in the images are rounded to 2 decimal points for viewing experience. My goal for tomorrow: Make training faster.

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/b9d3aef9-0b6b-4fcc-9daa-a98c300474fa-image.png

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/907e20c2-324d-45aa-88ea-4b81aa51e14e-image.png

https://scrapbook-into-the-redwoods.s3.us-east-1.amazonaws.com/786b1502-e2a7-4ed2-a0b0-c28f7b0e078c-image.png

Sun, Dec 22, 2024

Day 1 of coding a neural network from scratch (no libraries): I got a basic network implementation working, i forcefully set biases and weights. The network returns input * 10 right now, I'm going to hopefully get training working soon.