Dataflow Matrix Machines (since 2017)

• which can fluently modify their own weights, connectivity patterns, and size;
• and can serve as a programming platform, resulting in a programming formalism with continuously deformable programs.

2015-2017

Timeline and details: Partial Inconsistency and Vector Semantics of Programming Languages.

Reference paper: Michael Bukatin and Jon Anthony. Dataflow Matrix Machines and V-values: a Bridge between Programs and Neural Nets. In "K + K = 120" Festschrift, 2017.

2018

The slides for my talk at IBM AI Systems Day 2018, Cambridge, MA, October 3, 2018: Dataflow Matrix Machines and V-values: a Bridge between Programs and Neural Nets (joint work with Jon Anthony).

(The file with slides https://researcher.watson.ibm.com/researcher/files/us-lmandel/aisys18-bukatin.pdf has moved. This file is referenced in a number of PDFs on this page. Use https://web.archive.org/web/20220305051310/https://researcher.watson.ibm.com/researcher/files/us-lmandel/aisys18-bukatin.pdf also known as https://tinyurl.com/ibm-aisys18-bukatin)

DMM technical report 11-2018. Dataflow matrix machines: recent experiments and notes for next steps. Preprint, November 2018.

2019

Regularization in intrinsically sparse networks. An experimental study, February-March 2019.

A white paper on dataflow matrix machines: dmm-white-paper-2019.pdf (2 pages, November 2019).

2019 research notes on dataflow matrix machines.

An interdisciplinary and collaborative research agenda related to dataflow matrix machines: Version 1 (4 pages, December 2019).

2020

Using streams of probabilistic samples in neural machines: dmm-probabilistic-samples.pdf (January 2020).

Synergy between AI-generating algorithms and dataflow matrix machines: synergy-dmm-ai-ga.pdf (March 2020).

An interdisciplinary and collaborative research agenda related to dataflow matrix machines: dmm-collaborative-research-agenda.pdf (7 pages, August 2020; updated April 2022).

Higher-order neuromorphic computations with linear streams. Extended abstract and slides for my CCC 2020 talk (September 3, 2020).

2021

• interplay between DMMs and Transformers
• interplay between DMMs and Differentiable Programming

Towards practical use of dataflow matrix machines: towards-practical-dmms.pdf (March 2021; updated April 2022).

Multiplying monochrome images as matrices: A*B and softmax. A virtual poster presented at JuliaCon 2021 (July 2021)

2022

• Gradients with respect to values stored inside trees work (Julia, JAX, April 2022)
• Gradients through variadic DMMs with dictionary flows work (Julia, May 2022)
• DMM training by gradient method works for some classes of variadic DMMs with dictionary flows (Julia, June 2022)
• A primer on performing program synthesis and DMM synthesis via a neural architecture search algorithm based on differentiable programming works (Julia, June 2022)

• The June primer on program synthesis and DMM synthesis via a neural architecture search algorithm based on differentiable programming is open-sourced (September 2022, see https://github.com/anhinga/2022-notes)

2023

• interplay between DMMs and Transformers
• interplay between DMMs and Differentiable Programming
• interplay between DMMs and conventional algorithmic tasks
• exploring issues relevant to AI safety research (see https://github.com/anhinga/2023-notes)

Exploring synthesis of flexible neural machines with Zygote.jl (Materials for a JuliaCon 2023 talk, July 2023)

2024

• Pondering invariant properties of self-modifying systems (see https://github.com/anhinga/2024-notes)