Nooreign
Menu

computing · 2012

The Accuracy Surprise

Two researchers with two GPUs win ImageNet by a margin nobody predicted, and the field's priors about what deep learning could do are wrong by an order of magnitude. The alignment community's lesson is not about vision but about forecasts: capability surprised people who were closer to the evidence than the rest, which is a strong hint that not-yet-appearing capabilities are not a safe thing to schedule around.

“The results demonstrate that deep convolutional neural networks can achieve high accuracy on image recognition tasks. (paraphrase of the ImageNet classification paper's abstract)”
— Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, NeurIPS (2012)

Advances

  • Convolutional networks at scale, with GPU training and aggressive augmentation
  • A measurable, public benchmark that a field could be surprised by
  • Trajectory mismatch: capability outran the field's own documented projections

Landmarks

  1. 2012 CE

    ImageNet classification

    Krizhevsky, Sutskever and Hinton reduce top-5 error to 15.3% against a runner-up at 26.2%.

  2. 2012 CE

    The availability window

    Two commodity GPUs, an open dataset and a public leaderboard — the surprise was reproducible.

Who was arguing

    Alignment stack at 2012 CE (2 of 20 techniques)

    • Hard Constraints

      training

      Encode prohibitions as rules the optimiser cannot trade away, rather than as terms it can outbid. Runs from Asimov's Three Laws (1942) to rule engines, shielded formal verification, hard token constraints and refusal scaffolding. Weakness: a system that can route around a rule it can represent can route around it, so the constraint has to be enforced somewhere the policy cannot rewrite.

    • Imitation Learning

      training

      Learn by copying demonstrations instead of by optimising a proxy. Behaviour cloning and DAgger move the burden from reward design to demonstration coverage: the system inherits the trainer's competence and the trainer's blind spots. It remains the substrate under every current imitation-of-humans approach, including the pretraining-plus-finetuning of large language models.

      Dean A. Pomerleau and Dean L. Thorlabs · ALVINN: An Autonomous Land Vehicle in a Neural Network · NeurIPS 1989

    Techniques introduced in this era: Hard Constraints, Imitation Learning