References: Where deep learning breaks

Source material

Source curriculum (structural mirror, cited as further study):
• MIT 6.S191, "Introduction to Deep Learning", Lecture 6: "New Frontiers" (limitations)
  Instructors: Alexander Amini and Ava Amini (MIT)
  Course page: https://introtodeeplearning.com
  Code and labs: https://github.com/aamini/introtodeeplearning
  License: MIT (slides, code, and labs); videos are YouTube standard
  Required attribution: "© Alexander Amini and Ava Amini, MIT 6.S191:
    Introduction to Deep Learning, IntroToDeepLearning.com"
This lesson mirrors the limitations material of the new-frontiers lecture, kept
to technical failure modes. Clawdemy's lessons are original prose that follows
the pedagogical arc of this course. We do not reproduce or transcribe the
lectures; we cite them as the recommended companion. Course materials are used
under their MIT license with the attribution above; all rights to the original
videos remain with the creators.

Watch this next

• MIT 6.S191, Lecture 6: New Frontiers by Alexander and Ava Amini. The lecture this lesson mirrors. It surveys where deep learning struggles and where the field is heading, with the instructors’ own framing of the limitations covered here. The natural companion for the honest end of the tour.

Going deeper

A short, durable list. Each link is a specific next step, not a generic pile.

• “Explaining and Harnessing Adversarial Examples” (Goodfellow, Shlens, and Szegedy, 2014). The paper behind the panda-to-gibbon example in this lesson. It shows how a tiny, human-invisible perturbation flips a confident classification, and argues why networks are vulnerable. The primary source if you want to see brittleness demonstrated and explained.

• The MIT 6.S191 software labs. Some labs let you probe model behavior directly, including how models respond to inputs unlike their training data. MIT-licensed; a hands-on way to feel the brittleness this lesson describes rather than just read about it.

Adjacent topics

Where this connects inside the track.

• Learning by trial and reward (lesson 8). The previous lesson already met one of these limits up close: reinforcement learning’s sample-inefficiency and brittleness. This lesson generalizes the honesty across all of deep learning.

• Seeing the field whole (lesson 10). The final lesson assembles the whole track, the capabilities and these limits together, into one map, and points you toward where to go next. Knowing the limits is what makes that map trustworthy rather than a highlight reel.