Axon [Release 1.0] : Functional DL Library (+ Secret Multimodal Learning AI?)

Resources Community Resources
, , , ,
#1

Sorry to annoy you! But could you give us a like for this post? We would appreciate it!

Usage Preview

Project Axiom Details

“Project Axiom” (or “Axiom”, for short) is a secret experimental model architecture developed by Aqwam (me) under the DataPredict ecosystem.

It uses a new multimodal model that combines with text, environment state and deep reinforcement learning. However, it does not use images or audio as its environmental state input, but instead uses raw numerical values from the environment.

Generally, when text is involved, it uses “Large Language Model” (LLM) and this can be computationally expensive.

Fortunately, Axiom uses very different model architecture compared to LLM and tend to be smaller in size in exchange for more complex mathematics to build the model architecture. The Axiom model contains autoencoder-like parts that allows it to be smaller in size despite it being able to generate coherent and logical texts. It also allows the model to run on CPU and without GPU. It also runs natively on Roblox’s Luau ecosystem.

Current Prototype Specs:

  • 30K model parameters.

  • 5K context window.

Overview

DataPredictAxonIconSmall

DataPredict Axon is a functional deep learning library with dynamic computational graph capabilities for rapid prototyping.

Documentation And Tutorials: Welcome to Aqwam’s DataPredict™ Axon Library! | DataPredict-Axon

Features

  • Take advantage of automatic differentiation and dynamic computational graph for rapid prototyping.

  • Build singular models that are interconnected between servers and clients through distributed training.

  • Build models that handles multi-dimensional tensor inputs and outputs to solve any demands of your projects.

  • Dive into user-friendly API designed for you to learn in a couple of minutes.

  • Built for research-grade applications.

Use Cases

  • Self-Learning AIs (such as enemies, pets and companions)

  • In-Game Recommendation System

  • Image Moderation

  • Image Generation

  • Player Action Prediction System

  • Furniture/Parts Placement Generation (if you have some land plot system like in Lumber Tycoon or the Work At Pizza Place)

  • Personalized Item For Player (for example, if you have an item that wants to adapt to player’s usage of item)

Preview Codes 

local ServerScriptService = game.ServerScriptService

local TensorL = require(ServerScriptService.TensorL)

local DataPredictAxon = require(ServerScriptService.DataPredictAxon)

local ADTensor = DataPredictAxon.AutomaticDifferentiationTensor

local PaddingLayer = DataPredictAxon.PaddingLayers

local inputTensor = ADTensor.createRandomNormalTensor{{10, 2}}

local weightTensor = ADTensor.createRandomNormalTensor{{2, 4}}

local targetTensor = ADTensor.createRandomNormalTensor{{10, 4}}

local WeightContainer = DataPredictAxon.WeightContainer.new{

	{weightTensor, 1}, 

}

local function model(inputTensorPlaceHolder, weightTensorPlaceHolder)
	
	local tensor3 = inputTensorPlaceHolder:dotProduct{WeightTensorPlaceHolder}

	local finalTensor = DataPredictAxon.ActivationFunctionLayers.LeakyRectifiedLinearUnit{tensor4}

	local costValue = DataPredictAxon.CostFunctions.FastMeanSquaredError{finalTensor, targetTensor}
	
	return costValue
	
end

for i = 1, 100000 do

	local costValue = model(inputTensor, weightTensor)

	print(costValue)
	
	costValue:differentiate{}

	WeightContainer:gradientDescent{}

	costValue:destroy{true}
	
	task.wait()
	
end
Performance Benchmark: DataPredict Axon Vs DataPredict Neural

  • For single data point training, DataPredict Axon is up to 100x more faster than DataPredict Neural.

  • For batched data, DataPredict Neural is up to 900x more faster than DataPredict Axon.

Other Libraries From Me

  • DataPredict (General Purpose Machine Learning And Deep Learning Library For Roblox)

  • DataPredictNeural (Static Computational Graph Deep Learning Library For Roblox)

Plugins And Scripts That Complements With This Library

  • Parallel Luau (For speeding up the self-learning AIs’ training process through multi-threading.)

  • StepPhysics Plugin API (For speeding up the self-learning AIs’ training process through speeding up the environment.)

  • Graph Module (For plotting graphs.)

FAQs

  1. Can i build transformers with it?

    • Yes.

  2. Can I build computational graphs with it?

    • Yes.

  3. Can I build graph neural networks with it?

    • Yes.

  4. Can I build convolutional neural networks with it?

    • Yes.

Planned Updates:

  • Deep Reinforcement learning + Recurrent Deep Reinforcement Learning.

  • Patenting Project Axiom and make improvements to the model architecture. Then release the codes under this library.

  • Currently, busy with my Masters research thesis, expect there will be no update for few months. For now, I’m releasing it for everyone to enjoy my library.

With Project Axiom lurking around in the dark corner, would you use this library and others that falls under the DataPredict ecosystem?

  • Yes
  • Probably
  • DataPredict Axon Only

0 voters

You want me to release the Project Axiom code for your own use?

  • Yes
  • No

0 voters

7 Likes