Research fields

Computer science and AI

Computer science and AI labs run experiments that can be rerun, because every run says how it was made. Train and evaluate models in notebooks, keep the code commit, parameters and metrics of every run, and work under quotas and a monthly cost ceiling.

LabSpace for computer scientists

• Train and evaluate a classifier

  Split the data, train a model with scikit-learn and read the held-out accuracy and confusion matrix, all in one notebook cell you can run again.

  1. Load the dataset
  2. Train a gradient-boosted classifier
  3. Score it on held-out data
  4. Plot the confusion matrix

• Compare runs across seeds

  Run the same training with several seeds and plot the spread, so a result in the paper is not a lucky run.

  1. Train the model with five seeds
  2. Collect validation accuracy per epoch
  3. Plot the mean and spread
  4. Report the final accuracy with its range

• Keep the experiment in one place

  An experiment holds its notebooks, statistics, reports and notes on one canvas, so the lab can see what was tried and what came of it.

  1. Open the experiment from the lab
  2. Add a notebook to the canvas
  3. Add Stats and a report beside it
  4. Switch to the view the lab shares

The tools behind it

• Notebooks: Colab-style notebooks with a browser runtime or your own computer
• Results ledger: Code commit, parameters, metrics, dataset version and environment image per run
• Compute: Cloud resources under quotas and a monthly ceiling
• GitHub: Connect repositories so runs point at real commits