Human few-shot object learning benchmarks

This repository contains the few-shot object learning benchmarks described in "How well do rudimentary plasticity rules predict adult visual object learning?" (Lee and DiCarlo, 2023). It also lets you download the raw data and images from the experiments in the paper.

If you'd like to download the raw data and images without using the hobj library, you can do so at the OSF repository for this project.

Quickstart

Install

The hobj package works for Python >=3.12.

Install it from PyPI:

pip install hobj

If you prefer to use uv, you can install it with:

uv pip install hobj

Using hobj to compare a model against human learning data

The script below shows you how you can run evaluate a linear learning model – a model where new objects are learned by training a linear decoder using trial-and-error reward feedback – against humans.

All you need is a way to process a PIL.Image into a vector of image features (as an np.ndarray). There are ~18,000 256x256 images that you'd need to compute image features for.

import hobj
import numpy as np 

# Compute your features for the images 
my_image_features: dict[str, np.ndarray] = {}
for image_id in hobj.list_image_ids():
    image = hobj.load_image(image_id=image_id) # PIL.Image
    
    # Compute your features here:  
    my_image_features[image_id] = ... # replace right hand side with your image-computable model

# Assemble the learning model:
model = hobj.create_linear_learner(
    image_id_to_features=my_image_features,
    update_rule_name='Square', # "Square", "Perceptron", "Hinge", "MAE", "Exponential", "CE", "REINFORCE",
    alpha=1, # learning rate between [0, 1]
)

# Load the benchmark:
benchmark = hobj.MutatorHighVarBenchmark()  # or hobj.MutatorOneshotBenchmark()

# Score the model:
result = benchmark.score_model(model)

# Print its score and its CI:
print(result.msen, result.msen_CI95)

# You can also check out more granular statistics of the model's behavior, like its learning curves: 
# print(result.model_statistics)

For more details (e.g., how to load the raw behavioral data or images in Python), check out the Jupyter notebooks in examples/.

Contact

If you have any questions, need help, or experience a bug, please don't hesitate to email me (mil@mit.edu), or open an issue on this repo!

Changes to codebase since publication

This codebase was overhauled in 2026 to improve its accessibility, performance, and quality. Along the way, minor changes to the statistical analysis procedure were introduced, along with changes to the names of the original filenames (see changelist). To see the codebase at the time of publication, check out the repo with the v1 tag here.

Citation

@article{lee2023well,
  title={How well do rudimentary plasticity rules predict adult visual object learning?},
  author={Lee, Michael J and DiCarlo, James J},
  journal={PLOS Computational Biology},
  volume={19},
  number={12},
  pages={e1011713},
  year={2023},
  publisher={Public Library of Science San Francisco, CA USA}
}