index.html

---
layout: default
title: Elegant ML
---

<style>
/* Header */

.jumbotron {
  margin-bottom: 0;
  position: relative;
  border-bottom: 1px thick;
}

.jumbotron .container {
  max-width: 750px;
  box-shadow: 2px;
  position: relative;
  padding:none;
}

.jumbotron .buttons {
  padding: 10px;
}

@media (min-width: 576px){
  .jumbotron .container{
      padding: calc(1rem + 30px) 15px;
  }
}

.jumbotron hr {
  border-top: 1px solid rgb(100, 150, 150);
}

.jumbotron h1 {
  font-family: lato;
  font-style: italic;
}

/* Description */
.jumbotron {
  padding: 0;
}

.jumbotron p {
  font-size: 1.1em;
  margin-bottom: 1.5rem;
}

.btn-primary {
  color: rgb(30, 50, 50);
  border-color: rgb(150, 200, 200);
  background: none;
  border-width: 2px;
}

.jumbotron .btn-primary:hover {
  color: rgb(0, 0, 0);
  border-color: rgb(30, 80, 80);
  background: none;
  border-width: 2px;
}

/*Features*/
.features {
  padding-top: 1em;
}

.feature {
  padding-top: 1em;
  padding-bottom: 2rem;
  letter-spacing: 0.1px;
}

.paper {
  padding: 1em;
}

.feature h5, .paper h5 {
  font-style: italic;
  padding-bottom: 0.5em;
}

.friends{
  background: #fff;
}
.friends *{
  text-align: center;
  padding:10px;
}
.friends h3{
  margin-top: 50px;
}
.friends ul{
  display: flex;
  -webkit-justify-content: space-evenly;
          justify-content: space-evenly;
  flex-wrap: wrap;
  list-style: none;
  align-items: center;
}
.friends ul li img{
  max-width: 200px;
  max-height: 100px;
  height: auto;
}
</style>

<!-- Header -->
<div class="jumbotron jumbotron-fluid no-pad">
  <div class="container" style="text-align:center">
    <img src="logo.png" style="max-width:100%;padding-bottom:20px"  />
    <h2>The <i>Elegant</i> Machine Learning Stack</h2>
    <div class="buttons">
      <a class="btn btn-primary btn-lg" href="Flux.jl" role="button"><i class="fas fa-download"></i> Try It Out</a>
      <a class="btn btn-primary btn-lg" href="https://github.com/FluxML/Flux.jl" role="button"><i class="fas fa-star"></i> GitHub</a>
    </div>
  </div>
</div>

<!-- Main Content -->
<div class="features" style="background:white;padding-top:2em;">
  <div class="container">
    <div class="row">
      <div class="col-md-12" style="text-align:center;color:rgb(100,130,130)">
        <h2>Features & Ecosystem</h2>
      </div>
    </div>
    <div class="row">
      <div class="col-md feature">
        <h5>Compiled Eager Code</h5>
        <p>
          Flux provides a single, intuitive way to define models, just like mathematical notation.
          Julia transparently <a href="https://julialang.org/blog/2018/12/ml-language-compiler">compiles your code</a>, optimising and fusing kernels for the GPU, for the best performance.
        </p>
      </div>

      <div class="col-md feature">
        <h5>Differentiable Programming</h5>
        <p>
          Existing Julia libraries are differentiable and can be incorporated directly into Flux models.
          Cutting edge models such as <a href="https://julialang.org/blog/2019/01/fluxdiffeq">Neural ODEs</a> are first class, and <a href="https://github.com/FluxML/Zygote.jl">Zygote</a> enables overhead-free gradients.
        </p>
      </div>

      <div class="col-md feature">
        <h5>First-class GPU support</h5>
        <p>
          GPU kernels can be written directly in Julia via <a href="https://github.com/JuliaGPU/CUDAnative.jl">CUDAnative</a>.
          Flux is uniquely hackable and any part can be tweaked, from GPU code to custom gradients and layers.
        </p>
      </div>
    </div>
    <div class="row">
      <div class="col-md feature">
        <h5>The Model Zoo</h5>
        <p>
          A <a href="https://github.com/FluxML/model-zoo">rich collection</a> of Flux scripts to learn from, or tweak to your own data.
          Trained Flux models can be used from <a href="https://github.com/JuliaText/TextAnalysis.jl">TextAnalysis</a> or <a href="https://github.com/FluxML/Metalhead.jl">Metalhead</a>.
        </p>
      </div>

      <div class="col-md feature">
        <h5>ONNX</h5>
        <p>
          <a href="https://github.com/FluxML/ONNX.jl">Import</a> trained ONNX models as Flux scripts, for high-quality inference or for transfer learning.
        </p>
      </div>

      <div class="col-md feature">
        <h5>FluxJS</h5>
        <p>
          <a href="https://github.com/FluxML/FluxJS.jl">Export</a> your models to JavaScript for the browser, or see our <a href="experiments/">experiments</a> page for demos.
      </div>
    </div>

    <div class="row">
      <div class="col-md feature">
        <h5>Probabalistic Programming</h5>
        <p>
          The <a href="https://github.com/TuringLang/Turing.jl">Turing.jl</a> and <a href="https://github.com/willtebbutt/Stheno.jl">Stheno</a> libraries enables probabalistic programming, bayesian inference and Gaussian processes on top of Flux.
        </p>
      </div>

      <div class="col-md feature">
        <h5>TPUs & Colab</h5>
        <p>
          Flux models can be <a href="https://github.com/JuliaTPU/XLA.jl">compiled to TPUs</a> for cloud supercomputing, and run from Google Colab notebooks.
        </p>
      </div>
    </div>
  </div>
</div>

<div class="publications" style="background:rgb(230,250,250)">
  <div class="container features">
    <div class="row">
      <div class="col-md-12" style="text-align:center;color:rgb(100,130,130)">
        <h2>Publications Highlights</h2>
      </div>
    </div>
    <div class="row">
      <div class="paper col-md">
        <h5>Policy Guided Monte Carlo: Reinforcement Learning Markov Chain Dynamics</h5>
        <p>
          We introduce <i>Policy Guided Monte Carlo</i> (PGMC), a computational framework using reinforcement learning to improve Markov chain Monte Carlo (MCMC) sampling.
          <a href="https://arxiv.org/abs/1808.09095">arXiv</a>
        </p>
      </div>

      <div class="paper col-md">
        <h5>Machine Learning and System Identification for Estimation in Physical Systems</h5>
        <p>
          We draw inspiration from both classical system identification and modern machine learning in order to solve estimation problems for real-world, physical systems.
          <a href="https://lucris.lub.lu.se/ws/files/55399444/phdthesis.pdf">link</a>
        </p>
      </div>
    </div>

    <div class="row">
      <div class="paper col-md">
        <h5>Don't Unroll Adjoint: Differentiating SSA-Form Programs</h5>
        <p>
          This paper presents reverse-mode algorithmic differentiation (AD) based on source code transformation, in particular of the Static Single Assignment (SSA) form used by modern compilers.
          <a href="https://arxiv.org/abs/1810.07951">arXiv</a>
        </p>
      </div>

      <div class="paper col-md">
        <h5>Automatic Full Compilation of Julia Programs and ML Models to Cloud TPUs</h5>
        <p>
          We describe a method and implementation for offloading suitable sections of Julia programs to TPUs via the Google XLA compiler.
          <a href="https://arxiv.org/abs/1810.09868">arXiv</a>
        </p>
      </div>
    </div>
  </div>
</div>

<div style="background:white;padding-top:2em;padding-bottom:1em;text-align:center;">
  <div class="container">
    <h2>Blog</h2>
    {% for post in site.posts %}
    <p>
      <a href="{{post.url}}">{{post.title}}</a>
    </p>
    {% endfor %}
  </div>
</div>

<div class="friends">
  <div class="container">
    <h3>Researchers, users and developers of Flux</h3>
    <ul>
      <li><img src="./assets/images/ucl-logo.png"></li>
      <li><img src="./assets/images/uoft_logo.png"></li>
      <li><img src="./assets/images/juliac-logo.png"></li>
      <li><img src="./assets/images/alan-turing.jpg"></li>
      <li><img src="./assets/images/edinburgh.png"></li>
      <li><img src="./assets/images/rai-logo.png"></li>
      <li><img src="./assets/images/washington.jpg"></li>
      <li><img src="./assets/images/cambridge.jpg"></li>
    </ul>
  </div>
</div>