Research map / Architectures and efficiency / Neural architecture search
AutoML-Zero: Evolving Machine Learning Algorithms From Scratch
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evolve the full algorithm from basic mathematical operations | AutoML-Zero · CIFAR-10 binary classification test | accuracy: 82.22±0.17% -> 84.06±0.10% | Helped |
| Evolve the full algorithm from basic mathematical operations | AutoML-Zero · SVHN binary classification | accuracy: 85.14% -> 88.12% | Helped |
| Evolve the full algorithm from basic mathematical operations | AutoML-Zero · Downsampled ImageNet binary classification | accuracy: 78.44% -> 80.78% | Helped |
| Evolve the full algorithm from basic mathematical operations | AutoML-Zero · Fashion MNIST binary classification | accuracy: 98.21% -> 98.60% | Helped |
| Use evolutionary search instead of random search | AutoML-Zero · Linear regression tasks | acceptable-algorithm success rate: 5 times | Helped |
| Add noise to the inputs (ablation) | Best evolved algorithm · CIFAR-10 binary classification | accuracy | Helped |
| Use bilinear multiplicative interactions (ablation) | Best evolved algorithm · CIFAR-10 binary classification | accuracy | Helped |
| Normalize the weight gradient to a unit vector (ablation) | Best evolved algorithm · CIFAR-10 binary classification | accuracy | Helped |
| Accumulate training weights for inference (ablation) | Best evolved algorithm · CIFAR-10 binary classification | accuracy | Helped |
| Use noisy-ReLU adaptation | Evolved neural network · Few-example CIFAR-10 tasks | frequency of noisy-ReLU emergence: expt: 8/30, control: 0/30, p<0.0005 | Helped |
| Use learning-rate decay | Evolved neural network · Fast-training CIFAR-10 tasks | frequency of learning-rate decay emergence: expt: 30/30, control: 3/30, p<10^{-14} | Helped |
| Use transformed mean of the weight matrix as the learning rate | Evolved neural network · 10-class CIFAR-10 tasks | frequency of transformed-mean learning-rate emergence: expt: 24/30, control: 0/30, p<10^{-11} | Helped |
Closest papers by meaning
- Searching for Machine Learning Pipelines Using a Context-Free Grammar (AAAI 2021 · similarity 0.88)
- Large-Scale Evolution of Image Classifiers (ICML 2017 · similarity 0.88)
- ADELA: Accelerating Evolutionary Design of Machine Learning Pipelines with the Accompanying Surrogate Model (AAAI 2025 · similarity 0.87)
- Runtime Analysis of Evolutionary NAS for Multiclass Classification (ICML 2025 · similarity 0.85)
- AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML (ICML 2025 · similarity 0.84)
- AutoAL: Automated Active Learning with Differentiable Query Strategy Search (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.