Better AI Interpretability
Improve your models with minimal time/compute/data cost: identify where models are weak and do targeted fine-tuning/retraining
Knowledge discovery: find patterns in data that you didn’t know existed
Predict failure modes: prevent embarrassing or dangerous mistakes before deployment
Demonstrate safety/coherence: show that the features your model uses to make predictions are sensible and unbiased
Powerful and opaque systems are changing our world.
Humans need to understand them better – so we can learn from them, so we can ensure they don't inherit our biases, and so we can make them safe.
Leap Labs is a research-driven interpretability startup. We develop state of the art interpretability techniques that make it possible for humans to understand complex AI systems in new ways.
ML developers work hard to train reliable, accurate models, spending enormous amounts of time on data prep, model training, testing, and troubleshooting – but most of the time we’re working in the dark, with our understanding of what our models are learning limited to a few quantitative metrics and evaluations on a fixed dataset. Even if a model’s performance on a test set is good, there’s no guarantee that all possible failure modes and edge cases are contained in that test set.
We can’t be sure our models won’t fail in the real world, because we don’t understand them.
Enter interpretability – the science of understanding what neural networks have learned. Existing interpretability methods are mostly academic in nature and not up to the task of interpreting today’s increasingly powerful and complex models. Besides, most of these methods are data-dependent, meaning that they provide explanations based on data that we already have – this doesn’t tell us much about what a model has learned overall.
Leap Labs, 10min.mp4We need better interpretability.
We need methods work for any kind of model, so that we don’t have to trade-off performance for interpretability, and can make informed choices between architectures based on what we want a model to learn.
We need to be able to do this no matter what type of data we’re working with and what problem we’re trying to solve – from cancer detectors to chat bots.
We need to extract information directly from the model, so what we learn isn’t limited to the data we already have.
We need both global and local interpretability, so teams can understand what a model has learned overall, as well as why an individual prediction is made.
We need all of this to have low engineering overhead – so it’s easy to integrate throughout the entire model development process.
This is exactly what we're building at Leap Labs.
Find out more