Full explainability and interpretability deep-learning analysis platform

Tensorleap allows data scientists to fully understand their neural networks, ensure model performance, design the perfect dataset, and get on the same page as their team

How Tensorleap Works

The platform runs mathematical operations on each node in the neural network's computational graph in order to extract indicators from all model feature maps and evaluate their contributions.

Tensorleap’s algorithms then construct the most informative latent space to explain the model’s interpretation, find clusters of samples, and more.

KEY FUNCTIONALITY & BENEFITS

Reliability

Guided Error Analysis

Detect true root causes and understand which and why specific experiments are needed

Clustering:
Automatically group failing samples with identical root causes including metrics to evaluate the model’s quality and generalization

Clean-up:
Identify repetitive, redundant, ambiguous, and mislabeled data to understand which samples should be removed

Labeling prioritization:
Find the best samples for labeling to improve variance and achieve required representation of certain data that's causing model failures

Objectivity

Dataset Architecture

Design unbiased and balanced datasets to achieve practical distribution

Scoring:
Measure the quality of your datasets in regards to variance, density, entropy, and balance

Sample analysis:
Explore a model’s interpretation of each sample and identify main reasons for the prediction

Visualization engine:
View the model’s performance and breakdown across all sample characteristics and drill down into any specific population

Validity

Deep Unit Testing

Ensure a problem was solved and no regression took place while scanning for unknown issues

Split test sets:
Create numerous unit tests by either clustering using model features or automatically running cross-search on sample characteristics, and then validate them all simultaneously

Guided selection:
Drill down on specific sample groups to define and track specific tests using a visualization dashboard

Automatic scan:
Run unsupervised analysis utilizing a model's features to search for suspicious clusters and anomalies, without relying on user’s understanding of model’s behavior

Efficiency

DNN Collaboration

Confidently approve completed work understanding how and why it will affect model performance

Pull requests:
Enable developers to quickly review changes with their related tests and understand the impact on model’s interpretation

Issues:
Automatically track and document the entire issue-resolution process including tests, references, related samples, and more

Reporting:
Quickly generate a summary of individual work completed as well as collaborative effort to solve a specific issue

Tensoreleap is a fully Kubernetes-based solution which runs on any cloud or on-prem infrastructure.
It allows for quick integration to any framework of your models and datasets and supports any type of structured or unstructured data (images, text, time-series, tabular, etc.)

Use Tensorleap on your own model and data!