Software

This section lists various open-source software packages available from the Evolutionary AI team. These are available for research purposes only; for commercial use, contact info@evolution.ml. Unless otherwise specified, the packages are provided "as-is", i.e. as concrete implementations of technologies described in our research papers.

RHEA: Unlocking the Potential of Global Human Expertise (Code, Paper)

Realizing Human Expertise through AI (RHEA) unlocks the latent potential in diverse human-developed solutions by injecting them into an evolutionary process.

Semantic Density: Uncertainty Quantification for LLMs (Code, Paper)

Semantic density provides a response-wise uncertainty/confidence score for large language models (LLMs). It is off-the-shelf for any existing LLMs, and it works for free-form generation tasks.

LMX: Language Model Crossover (Code, Paper)

Language Model Crossover (LMX) uses large language models as the engine of evolution, driving recombination and variation for any task where solutions are representable as text.

SEPX: solving the permutation problem in NAS (Code, Paper)

Shortest Edit Path Crossover (SEPX) directly recombines architectures in the original graph space, overcoming the permutation problem in traditional evolutionary neural architecture search (NAS). Its advantage over Reinforcement Learning (RL) and other methods is proved theoretically and verified empirically.

AQuaSurF: Efficient Activation Function Search (Code, Paper)

AQuaSurF uses a surrogate modeling approach to quickly discover new activation functions that improve performance on a variety of tasks.

Act-Bench: Activation Function Benchmark Datasets (Code, Paper)

The Act-Bench-CNN, Act-Bench-ResNet, and Act-Bench-ViT benchmark datasets contain precomputed training results from 2,913 unique activation functions, and can be used to quickly benchmark new AutoML algorithms.

AutoInit: Intelligent Network Initialization (Code, Paper)

Initializing deep learning networks automatically so that learning will be more robust and effective. AutoInit is maintained and can be used as part of Tensorflow experiments.

TaylorGLO: Loss-Function Metalearning (Code, Paper)

TaylorGLO evolves loss-functions for a given architecture and task through multivariate Taylor polynomial parameterization, resulting in automatic regularization.

TOM: Multitask Embeddings (Code, Paper)

The Traveling Observer Model (TOM) implements deep multi-task learning through spatial variable embeddings.

XPRIZE Pandemic Response Challenge (Code, Paper)

In this competition, predictors were developed to forecast number of cases and prescriptors to recommend non-pharmaceutical interventions to cope with the COVID-19 pandemic. The package includes sample predictors and prescriptors and their evaluation code (see the competition tech page for details).

RED: Misclassification Detection (Code, Paper)

Residual-based error detection (RED) extends RIO to classification. It detects misclassification errors and distinguishes them from out-of-distribution inputs and adversarial inputs.

RIO: Modeling Uncertainty (Code, Paper)

Residual estimation with an input/output kernel (RIO) quantifies confidence in point-prediction models and improves their accuracy.

MUiR: Diverse Multitasking (Code, Paper)

Modular Universal Reparameterization (MUiR) makes deep multitask learning possible across diverse domains such as language, vision, genetics, etc.