reverse-pred 0.1.6

Library to run Reverse Predictivity

Reverse Predictivity

A lightweight, modular Python library for computing bidirectional alignment between artificial neural network (ANN) representations and primate inferior temporal (IT) cortex responses.

This package accompanies the preprint:

Muzellec & Kar (2025). Reverse Predictivity: Going Beyond One-Way Mapping to Compare Artificial Neural Network Models and Brains. bioRxiv. https://www.biorxiv.org/content/10.1101/2025.08.08.669382v1

Reverse predictivity complements traditional forward neural predictivity by asking the reciprocal question:

How well do neural responses predict ANN activations?

Together, the forward and reverse metrics provide a more complete picture of representational similarity between brains and models.

---

🧠 Library Overview

This library contains four core mapping modules:

Module	Mapping Direction	Question Answered
model_to_monkey.py	Model → Monkey	How well do ANN features predict neural responses? (forward predictivity)
monkey_to_model.py	Monkey → Model	How well do IT neurons predict ANN unit activations? (reverse predictivity)
monkey_to_monkey.py	Monkey A → Monkey B	How consistent are neural populations across animals? (biological upper bound)
model_to_model.py	Model A → Model B	How aligned are representations across models or layers?

All functions compute explained variance (EV) using repeated linear mappings and save EV arrays to disk.

---

🔧 Installation

We recommend using a clean environment:

conda create -n reverse_pred python=3.10 -y
conda activate reverse_pred

Install required Python packages:

pip install numpy scipy scikit-learn matplotlib

Install this library:

pip install reverse_pred

---

🚀 Usage

Each mapping function takes:

• model_features: (n_images × n_units) array
• rates: (n_images × n_neurons × n_repeats) array
• out_dir: output directory for saving EV results
• reps: number of cross-validated fits (default: 20)
• model_type: Choice of regressor models among ridge, linear, lasso, elasticnet, pls. Default=ridge

---

1. Forward Predictivity

Module: model_to_monkey.py
Function: compute_model_to_monkey

from reverse_predictivity.model_to_monkey import compute_model_to_monkey
import numpy as np

model_features = np.load("features/resnet50_itlayer.npy")
rates = np.load("data/it_rates.npy")

compute_model_to_monkey(
    model_features=model_features,
    rates=rates,
    out_dir="results/model_to_monkey/resnet50",
    reps=20,
    out_name='forward_ev'
)

Output:

results/model_to_monkey/resnet50/forward_ev.npy

---

2. Reverse Predictivity

Module: monkey_to_model.py
Function: compute_monkey_to_model

from reverse_predictivity.monkey_to_model import compute_monkey_to_model
import numpy as np

model_features = np.load("features/resnet50_itlayer.npy")
rates = np.load("data/it_rates.npy")

compute_monkey_to_model(
    model_features=model_features,
    rates=rates,
    out_dir="results/monkey_to_model/resnet50",
    max_n=None,
    reps=20,
    out_name='reverse_ev'
)

Parameters:

max_n: can be set to subsample n number of neurons within the neural population. Each repetition will be done using a different sampling.

Output:

results/monkey_to_model/resnet50/reverse_ev.npy

---

3. Neural–Neural Consistency

Module: monkey_to_monkey.py
Function: compute_monkey_to_monkey

from reverse_predictivity.monkey_to_monkey import compute_monkey_to_monkey
import numpy as np

ratesA = np.load("data/monkeyA_rates.npy")
ratesB = np.load("data/monkeyB_rates.npy")

compute_monkey_to_monkey(
    rates_predictor=ratesA,
    rates_predicted=ratesB,
    out_dir="results/monkey_to_monkey/",
    reps=20,
    max_n=None,
    name_predicted="monkeyB",
    name_predictor="monkeyA",
)

Parameters:

max_n: can be set to subsample n number of predictor neurons. Each repetition will be done using a different sampling.

Output:

results/monkey_to_monkey/monkeyA_to_monkeyB_ev.npy

---

4. Model–Model Alignment

Module: model_to_model.py
Function: compute_model_to_model

from reverse_predictivity.model_to_model import compute_model_to_model
import numpy as np

modelA = np.load("features/resnet50_itlayer.npy")
modelB = np.load("features/convnext_itlayer.npy")

compute_model_to_model(
    model_features_predictor=modelA,
    model_features_predicted=modelB,
    out_dir="results/model_to_model/resnet_to_convnext",
    reps=20,
    name_predicted="convnext",
    name_predictor="resnet50",
)

Output:

results/model_to_model/resnet_to_convnext/resnet50_to_convnext_ev.npy

---

📊 Downstream Analysis

After generating EV results:

1. Load the saved .npy files.
2. Compare forward vs reverse predictivity.
3. Compare model–monkey EV to monkey–monkey EV.
4. Compare model–model EV across architectures.

import numpy as np
import matplotlib.pyplot as plt

fwd = np.load("results/model_to_monkey/resnet50/forward_ev.npy")
rev = np.load("results/monkey_to_model/resnet50/reverse_ev.npy")

plt.hist(fwd, bins=30, alpha=0.6, label="Forward")
plt.hist(rev, bins=30, alpha=0.6, label="Reverse")
plt.legend()
plt.xlabel("Explained Variance")
plt.ylabel("Count")
plt.show()

---

📌 Citation

If you use this library, please cite:

@article{muzellec_kar_2025_reversepredictivity,
  title   = {Reverse Predictivity: Going Beyond One-Way Mapping to Compare Artificial Neural Network Models and Brains},
  author  = {Muzellec, Sabine and Kar, Kohitij},
  journal = {bioRxiv},
  year    = {2025}
}

---

📜 License

MIT License — see LICENSE.