Skip to content

Cheatsheet: Where deep learning breaks

Every prior lesson: "here's what deep learning can do."
This one: "here's what it can't, and why."
Each limitation falls out of HOW these systems work, not from a missing patch.
LimitationWhat it isWhy (connects to)Why it matters
Data hungerneeds lots of labeled datalearning purely from examples (L1’s data pillar)poor fit when data is scarce; small data → memorizing
Brittlenessconfident failure on inputs unlike trainingpattern-matching, not understandingadversarial + out-of-distribution inputs break it
Data-slant (bias)mirrors imbalances in its datalearns whatever the data containsworse on under-represented cases
No guarantees / opacitycan’t certify correctness; hard to explainknowledge is in opaque weights (T11)confidence ≠ correctness; “black box”

Adversarial example: a panda image, correctly classified. Add a tiny, human-invisible noise → the network says “gibbon,” confidently. It keyed on fragile statistics, not the concept. Same fragility on out-of-distribution inputs (unusual angles, unseen topics): it answers anyway, often confidently.

A model is more accurate on what it saw more of. Over-represent some cases and under-represent others → uneven performance. This is a statistical property of learning from data, not intent. More data fixes it only if the data is more representative.

A text generator can produce fluent, confident, fabricated output (a citation to a paper that does not exist), because it produces what looks like its training data, not what is checked against truth. Convincing ≠ correct.

  • Invented citation = hallucination.
  • Weird confidence on a strange photo = out-of-distribution failure.
  • Smooth on common requests, stumbles on rare ones = data hunger + slant.
  • Rules of thumb: be skeptical of confident output (especially on novel/rare inputs); verify anything stated as fact; never read confidence as a guarantee.

Researchers work on all four and things improve, but they are deep properties of learning from data, not quick bugs. Treat “this is solved” claims with the same skepticism.

  • “These are temporary bugs.” No. Consequences of the approach; improved by research, not patched away.
  • “Trust the model’s confidence.” No. Confidently wrong is a thing (panda→gibbon, invented citation).
  • “More data fixes bias.” Only if it is more representative data.
  • “A fluent answer is a true answer.” No. Generative models optimize for plausible, not verified.

Everything in this track is real, powerful, and bounded; understanding deep learning means seeing why the same machinery produces both the wins and the failures.