Podcast Benchmark Documentation¶
A benchmarking framework for neural decoding from podcast listening data.
Decoding Tasks¶
- Brain --> perceived word decoding Translate brain signals to perceived words, comparing performance to previously published results.
- Audio Reconstruction Reconstruct podcast audio envelope from brain signal (Regression)
- Sentence Onset Detection Classify (binary) segments of brain data as containing the beginning of a sentence or not
- Content/Non-Content Words Classification (Binary classification)
- Part of Speech Classification (Multiclass classification)
- LLM Surprise Predict how likely the perceived word is given it's context (Regression)
- LLM Decoding Encode brain data as vector input to language models (GPT-2) for direct brain-to-text generation
Table of Contents¶
- Quickstart - Get up and running quickly
- Onboarding a New Model - Step-by-step guide to adding your own decoding model
- Adding a New Task - How to implement custom decoding tasks
- Configuration Guide - Understanding and configuring experiments
- Task Reference - Complete reference for all available tasks
- Baseline Results - Performance benchmarks for all tasks
- Registry API Reference - Registry decorators and function signatures
Overview¶
This framework provides a flexible system for: - Training neural decoding models on iEEG data - Comparing different model architectures - Evaluating performance across multiple metrics - Running systematic hyperparameter searches
For long updates and discussions, see this notebook.