The scripts were ran in Python 3.13.3 with NumPy 2.2.6, SciPy 1.16.0, PyTorch 2.7.0, Scikit-Learn 1.6.1, Matplotlib 3.10.1, seaborn 0.13.2, Numba 0.61.2 and tqdm 4.67.1.
Song preprocessing (optional; see below) requires Autoencoded Vocal Analysis 0.3.1 (AVA) and Joblib 1.5.0.
This repository contains the code for simulating the models and generating the core figures in the paper Correctness is its own reward: bootstrapping error signals in self-guided reinforcement learning (bioRxiv).
The mathematical details are described in the paper. src contains the source code for the networks, the reinforcement learning environment, the sparse coding model, and other helper functions.
For the basic training and simulations of the models, check out examples in simulations, which also contains the code to generate the panels such as neuronal firing rates, and joint distributions of mean rates between correct and perturbed singing cases in the paper. Notice that the networks themselves are implemented to take any time series input that matches the dimensions, but the simulations to reproduce the figures in the paper uses sparse embeddings of real spectrograms of adult songs. To generate this embedding, follow the steps in realistic_auditory_processing or directly download the generated embeddings.
further_analysis contains the code to run and visualize the analysis of the learned recurrent weights, as well as the reinforcement learning experiments. robustness_tests contains the code to plot the cancellation quality over different parameters.
pearsonlab/spiking-ven is a spiking implementation of
the same cancellation computation, developed alongside the full song-circuit model. Where
the networks in src/models.py are rate-based (Wilson–Cowan WCNet / EINet with the
bilin_hebb_* plasticity rules), that repository uses leaky integrate-and-fire
excitatory/inhibitory populations with spike-timing-dependent plasticity on the
inhibitory-to-excitatory weights, driven by a sparse-coded spike train from a gammatone
cochleagram front end.
It reproduces the same qualitative result — the excitatory response to the bird's own song is cancelled while responses to novel or perturbed sound survive — and is quantified against Mandelblat-Cerf et al. (2014) using their own criterion: 50 ms noise bursts during singing, with spike counts in 150 ms windows compared per neuron across renditions. The modelled response has an onset latency of 32.7 ms (bird: 23 ± 12 ms), peaks 48 ms after onset (bird: ~40–60 ms), and sits on a 2.2 Hz singing baseline (bird: 1–10 Hz), with 56% of units responding (bird: 41%, 7/17). Response magnitude remains the open problem, ~2× too strong: 1.79 extra spikes per responder against the bird's ~0.8–1.0. The pipeline is fully seeded and reproducible end to end from the public song corpus in about twenty minutes.
It is a separate, independent implementation rather than a port of this code, and its dependency footprint is deliberately minimal (numpy and scipy for the core), so it can be reused without this repository or the circuit model.