What Happened
Deep learning has a strange secret at its core: the field's most powerful tools work far better than anyone can fully explain. On June 24, that uneasy truth pushed its way back into the conversation when a 2026 paper from the École Polytechnique Fédérale de Lausanne (EPFL) climbed to the top of Hacker News and set off a fresh round of debate among researchers and engineers. Its title poses the question bluntly: "The Puzzling Success of Overparameterization: Lottery Tickets or Escape Dimensions?"
The paper, written by Flavio Martinelli, Johanni Brea, and Wulfram Gerstner of EPFL's Laboratory of Computational Neuroscience, takes aim at one of the most repeated explanations in modern machine learning. For nearly a decade, practitioners have leaned on the "lottery ticket" story to explain why bigger networks tend to train better. The EPFL group argues that the metaphor, however memorable, quietly misleads us about what is actually going on.
The lottery ticket hypothesis, introduced by Jonathan Frankle and Michael Carbin in 2019, observed that a large trained network usually contains a much smaller subnetwork that — given the right initial weights — can be trained on its own to match the full model's accuracy. The intuitive leap that followed was seductive: a wide network succeeds because it holds many such winning "tickets," and the more tickets you buy, the more likely one of them wins.
Why It Matters
Martinelli and his colleagues point out that the lottery framing smuggles in three assumptions borrowed from a real lottery: that a winning ticket guarantees a win, that your odds grow with the number of tickets, and that the tickets are independent of one another. None of the three, they argue, survives contact with how neural networks actually behave.
Their alternative reframes the question in terms of geometry. When a network has more parameters than training examples, the set of perfect solutions is not a handful of isolated points but a vast, high-dimensional surface threaded through the loss landscape. Each extra neuron adds another dimension to that landscape — and crucially, that new dimension often provides an escape route out of the bad local minima where a narrower network would get stuck.
Stack up enough of these "escape dimensions," the authors say, and the traps that plague small networks dissolve into gentle passes. Training stops being a gamble over which subnetwork you happened to draw and becomes something closer to a near-inevitable slide downhill. Width helps not because it buys more lottery tickets, but because it reshapes the terrain itself.
Reaction
The Hacker News thread that revived the paper drew hundreds of comments, with the usual mix of enthusiasm and skepticism. Many readers welcomed a cleaner mental model; others wondered whether "escape dimensions" is itself just a new metaphor dressed in more rigorous clothing.
That tension is part of why the debate resonates. The lottery analogy survived because it is easy to teach, even if it nudges people toward conclusions the math does not support — for instance, the idea that a winning subnetwork can be understood in isolation from the network around it. The EPFL authors show that subnetworks are deeply entangled, sharing weights and influencing one another, which is exactly where the lottery picture quietly falls apart.
For working engineers, the appeal is practical as much as philosophical. A sharper intuition about why scale helps could inform decisions about how wide to build a model, when pruning will pay off, and how to read the strange smoothness of large-network training.
What's Next
If the escape-dimensions view holds up, it carries direct implications for one of deep learning's most active subfields: model pruning, the art of shrinking a trained network without losing accuracy. The lottery framing suggested you might find and train winning subnetworks from scratch; the geometric view warns that those subnetworks lean on the full landscape they were carved from.
The work also feeds a broader research current that treats the loss landscape as an object worth studying in its own right, not merely a surface to descend. Tools that map its curvature, count its dimensions, and trace the manifolds of good solutions are becoming central to understanding why some architectures train smoothly and others stall.
None of this rewrites how today's largest models are built overnight. But it sharpens the questions researchers will bring to the next generation of architectures, where the cost of width — in energy, money, and compute — makes knowing why it works more than an academic concern.
Closing Thoughts
There is something quietly humbling about a field this powerful still arguing over its own foundations. The systems reshaping science, medicine, and daily work rest on training dynamics we are only beginning to describe in honest terms.
What makes the EPFL paper worth sitting with is not that it hands down a final answer, but that it insists on a more careful one. Trading a comforting metaphor for a harder, more accurate picture is rarely glamorous work. Yet that is often how understanding actually advances — not by buying more lottery tickets, but by learning to see the shape of the ground beneath our feet.
한글 요약
EPFL 계산신경과학 연구실의 플라비오 마르티넬리, 요한니 브레아, 불프람 게르스트너 연구진의 논문이 6월 24일 해커뉴스에서 다시 주목받으며, 큰 신경망이 왜 더 잘 학습되는지를 설명해 온 '로또 티켓' 비유가 오해를 부른다는 주장이 화제가 됐다. 2019년 제기된 로또 티켓 가설은 거대한 학습된 신경망 안에, 따로 떼어내 학습해도 전체 성능을 내는 작은 부분망이 존재한다는 관찰이었다. 여기서 '신경망이 클수록 당첨 티켓을 더 많이 가지므로 잘 학습된다'는 직관이 굳어졌지만, 연구진은 이 비유가 실제 신경망의 동작과 맞지 않는다고 지적한다.
대신 연구진은 '탈출 차원(escape dimensions)'이라는 기하학적 설명을 제시한다. 파라미터가 데이터보다 많을 때 완벽한 해는 몇 개의 점이 아니라 손실 지형 위에 펼쳐진 고차원 곡면을 이루며, 뉴런을 하나 더할 때마다 지형에 새로운 차원이 생겨 좁은 신경망이 갇히던 나쁜 국소 최소점에서 빠져나갈 통로가 열린다는 것이다. 너비가 도움이 되는 이유는 더 많은 '복권'을 사기 때문이 아니라 지형 자체를 바꾸기 때문이라는 해석이다.
이 관점이 맞다면 모델 가지치기(pruning)나 아키텍처 설계 방식에 직접적 함의가 있다. 부분망은 전체 지형에서 떼어내 독립적으로 이해할 수 없을 만큼 서로 얽혀 있기 때문이다. 딥러닝이 과학·의료·일상을 바꾸는 지금도 그 학습 원리는 여전히 논쟁 중이며, 편안한 비유 대신 더 정확한 그림을 택하려는 시도가 이해를 한 걸음 진전시킨다는 점에서 의미가 있다.