← Reference notes

License: CC BY 2.0 Belgium. Source: https://archive.fosdem.org/2026/schedule/event/ZRFZHL-deep-learning-demystified/

Summary

Jens Mönig (Snap!, with SAP and Berkeley), speaking at FOSDEM on 1 February 2026, says schools are teaching students to prompt chatbots and not how a neural net works. His classroom aims are a general-purpose learner, a clear line between similar and the same, a tiny algorithm students can see, and room for emergence. The tiny algorithm is a sales-slip: multiply quantities by prices and add them up (a dot product), then pass the total through a step or a smooth step. He credits the step function and that notation to older mathematics, then Frank Rosenblatt's perceptron. The live demo is the sonar set that separates mines from rocks, which a person cannot sort by eye. Random weights land near a coin flip. One learning pass nudges each weight by a fraction of the error. On stage the single perceptron climbs to about 78 percent and stalls. He uses Marvin Minsky and Seymour Papert for the reason: one layer only separates problems a straight cut can divide. A hidden layer plus backpropagation fixes that. The forward pass pipes the sample through the hidden layer and then the output layer. Learning sends the error backward, first from the output, then from the hidden layer, which is one extra delta on the same update rule. With a hidden layer of a few units the demo moves past 90 percent. He then draws six scribbles in Snap!, lets a block-sized network learn them, and uses the learned weights as a new block that names a fresh sketch and plays a matching animation. In questions he says "deep" here means more than one layer, images usually want a convolutional net, and the classroom uses the block on student data rather than asking every student to wire every layer. The whole demo runs locally inside Snap!. Percentages are what the live demo showed, not a published benchmark. Captions garble a few names. This note follows the talk, not a textbook.

Why it is here

Notes on agents and assistants assume a neural net. This lecture is a computing-basics account of the perceptron, the hidden layer, and backpropagation, taught as blocks a student can run, not as a prompt recipe.

@misc{iil_note_ry2ap94f,
  author = {{iinventedlight Library}},
  title = {{Deep Learning Demystified: Having Fun with Neural Networks in Snap!}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-10-18},
  url = {https://iinventedlight.com/library/reference/talks/fosdem-2026-deep-learning-demystified/},
  note = {Permanent ID iil:note:ry2ap94f. License: CC-BY-2.0. Summary of https://archive.fosdem.org/2026/schedule/event/ZRFZHL-deep-learning-demystified/.},
}

For language models and agents: start at /llms.txt. This page in Markdown: /library/reference/talks/fosdem-2026-deep-learning-demystified.md.