[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

[Doc] Deep Learning & Debian Development

Hi fellow devs,

FYI, I wrote a lengthy documentation that covers many
sub-topics about "Deep Learning & Debian Development":


The topics in the document is associated to ~90% of my debian
activities. Here is the outline of the documentation, afterall
most people won't really read the lengthy texts:

====== BEGIN OUTLINE ===

1. Deep Neural Network

  A brief and not quite mathematical explanation on what
  deep neural network is and how it works. This section
  also mentions the core components of a neural network

2. Deep Learning Framework

  Relation between DL framework and BLAS.
  Performance is a big issue. 

  2.1 SIMD

    (1) Bump ISA Baseline for the whole system (SIMDebian)
    (2) Build software locally (DUPR, Gentoo)
    (3) GCC's FMV feature
    (4) ld.so's "Hardware capability" feature

  2.2 Hardware Acceleration

    (1) Nvidia CUDA. It is the dominating solution provider.
        But its incooperative product license makes
        everything boring in terms of volunteered work.
    (2) AMD's fully-opensource counterpart ROCm/HIP. not
        quite mature.

  2.3 Third-party software distributors

    Anaconda. Not restricted by incooperative non-free
    licenses. Has its own advantages.

  2.4 Deep learning framework implementations

    Taxonomy: first generation, second generation.

    First generation: features static graph, including
      Caffe, Theano, Torch(Lua), TensorFlow(v1)

    Second generation: features dynamic graph and automatic
      differentiation, including PyTorch, TensorFlow (v2,
      or eager-execution mode).

    Some practical packaging issues related to them.

    Julia community also tried to provide some deep learning

3. Deep learning applications

  3.1 Data & pre-trained neural networks

    you guys already know what I'm going to talk about in
    this section.

  3.2 Software freedom and DFSG


  3.3 Neural network reproducibility

    todo (can be partially found in ML-Policy)

  3.4 Neural network releases and security

    todo. Deep neural networks can be vulnerable, actually.
    There is not a "CVE" for deep learning. (it's not time)

4. Ethics


5. Preliminary conclutions


====== END OUTLINE ===

I didn't carefully polish section 3 because it fully overlaps
with the ML-Policy motivation. I still need some time to
sync the document and ML-Policy.

The link to the HTML document will always available as long
as people.d.o keeps online. My writing style is not suitable
for a wiki page.

[1] https://salsa.debian.org/lumin/people.d.o-lumin
    Source code is available here. The source code of the
    html files are the manually written HTML themselves.

Reply to: