Reinforcement learning is a type of machine learning that involves training an agent to take actions that maximize a reward signal.

Description
Implement a complete framework for building reinforcement learning models in pure Rust.

General Formula

Describe the proposal or feature that this issue is tracking.

Issues

• #89
• #90
• #91
• #92
• #93

Pull Requests

Layer normalization directly estimates the normalization statistics from the summed inputs to the neurons in a hidden layer.

Model

Configuration

Parameters

The implementation can leverage the ParamsBase created for linear models.

• Bias (optional)
• Weight

A neural network is a set of layers that work together in-order to make a prediction.

Traits

• Backward
• Compile
• Forward
• Train

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

This trait should unify all numeric operators implemented in the crate and streamline several additional processes.

The transformer model was introduced in 2017 with the release of the Attention is all you need paper. Over the last few years, the popularity of the Transformer model has grown exponentially due to its abilities in natural language processing (NLP) workloads.

Transformers are mathematically simply, primarily relying on dot-product / matrix multiplication algorithms.

Issues

• #68
• #69
• #82
• #83
• #84

Pull Requests

Resources

• Attention is all you need
• The Annotated Transformer

• Create a set of traits that adequately describe the behaviors of a layer in a neural network
• Implement a Layer structure

Currently, concision is far too clunky and needs to be cleaned up before any additional work can take place.

TruncatedNormal is similar to Normal distributions, however, any values exceeding two times the standard deviation are discarded and regenerated.

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

References

• Improving neural networks by preventing co-adaptation of feature detectors

cnc::linear::Params

A tensor is defined to be an n-dimensional array.

• Reduced complexity with relaxed dimensional typing

References

Scaled Dot-Product Attention is a mechanism most extensively leveraged in the popular Transformer model.

Resources

• Attention is All You Need

Issues

• #86

Pull Requests

• [ ]

Resources

• Efficiently Modeling Long Sequences with Structured State Spaces
• The Annotated S4

Due to the lack of generalization present in Rust, it is important to consider exactly what any given model is operating on. Standardizing this would provide the system with the intended flexibility and may even have performance benefits.

Resources

• PyTorch Module

Formula

LecunNormal initializes objects using a truncated normal distribution that is centered at 0 and has a standard deviation of 1/n where n is the number of inputs.

Issues

• #87

Pull Requests

• [ ]

Resources

• Kolmogorov-Arnold Networks (KAN)

In the paper, the Annotated S4, they leverage the scan function from jax. While Rust has a well-developed scan feature present in iterators, it seems to differ in execution than the one from jax.

Development of this macro officially kicks off a long-standing desire to create a DSL of-sorts for building powerful, agile, and scalable neural networks.

The casual convolution function currently considers two array's of equal dimensions, however, the generated kernel is defined to be a one-dimensional array by the 'kernel_dplr' function.

Currently the rely on the wrong dimension, thus altering the shape.

Use either uuid::Uuid or another Copyable type for ids.

Create functions that consider Real-Valued inputs.

Updates

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

This is a tracking issue for initialization primitives.

Issues

• #71
• #72
• #78

Custom Distributions

• #73
• #75
• #76
• #77

Artificial neurons are given two free-parameters, bias and weights, that are to be learned through training. The bias describes an array of values that don't require any input to have an affect.

XavierNormal

Each weight is drawn from a normal distribution where the mean $\left({\mu}\right)$ is set to 0 and the standard deviation $\left({\sigma}\right)$ is computed as follows:

XavierUniform

Each weight is drawn from a uniform distribution in $\left[{-x, x}\right]$

Resources

• Understanding the difficulty of training deep feedforward neural networks

Methods

• lecun_normal
• normal
• stdnorm
• uniform