Temporal Delay Neural Network (TDNN)

A biologically-inspired time-delay neural network for temporal pattern and speech recognition. The model learns the transmission delays between spiking units — not just synaptic weights — so that information arriving along different paths can be aligned in time and integrated at a downstream neuron. This makes the network naturally suited to problems where the temporal structure of a signal carries the information, such as phoneme and consonant recognition from spectrograms.

Presented as a poster at the New York Academy of Sciences (NYAS), 2016 (NYAS-2016-POSTER.pdf).

Paper

A Neural Architecture for Unsupervised Temporal Pattern Recognition Shady El Damaty (Neuroscience) & James Maguire (Linguistics), Georgetown University, 2016. 📄 paper/nips_2016.pdf · LaTeX source in paper/

Abstract. We present a novel neuron architecture for unsupervised, time-dependent neural networks. We introduce a new delay function based on a mixture of Beta distributions, addressing limitations of prior delay-learning work, and a new method for thresholding a neuron's potential using input activations from both before and after the neuron spikes. Preliminary results on two patterns show very high performance on an unsupervised pattern segmentation task, learning in both clean and noisy environments. The approach builds on time-delay neural networks (Waibel et al. 1989) and spike-timing-dependent plasticity (Izhikevich 2006) toward fully unsupervised learning of time delays.

Idea

Classic artificial neurons sum weighted inputs instantaneously. Real axons impose conduction delays, and a growing body of work (gamma models, polychronization, delay-tuned coincidence detection) shows that learnable delays are a powerful computational primitive for temporal coding.

This project implements a feed-forward network of integrate-and-fire neurons in which each connection carries a learnable delay distribution. Three families of delay functions are implemented and compared:

• Masquelier — fixed delay lines (baseline)
• Single-Beta — one beta-distributed delay kernel per connection
• Beta-Mixture — a mixture of beta kernels, allowing multi-modal delay tuning

Delays and thresholds adapt online as patterns are presented, so the network learns to fire selectively for the temporal signature of a class (e.g. a particular consonant) rather than its instantaneous amplitude.

What's here

Path	Contents
py/	Core Python implementation — networks, neurons, delay functions, update rules, and simulation drivers
py/Networks/	Feed-forward network construction
py/Neurons/	Input, integrate-and-fire, and base neuron models
py/DelayFunctions/	Masquelier / single-beta / beta-mixture delay kernels + their update rules
py/Updates/	Online threshold and variable adaptation
py/simulation_script_*.py	Reproducible experiment drivers (consonants, simple images, long streams, with/without noise)
network-3.0/	A cleaner, pickled rewrite of the core engine
matlab/	Stimulus generation and analysis — spectrograms, beta-mixture fitting, weight/potential visualization (generateStream*.m, myspectrogram.m, calc_*, graphics_*)
figures/	Result figures — training/test weight and membrane-potential meshes, per-stimulus responses across SNR levels

Method, briefly

1. Stimuli — speech tokens (consonants) and simple visual patterns are converted to time–frequency representations (myspectrogram.m) and streamed to the network as per-channel spike trains.
2. Encoding — input neurons emit spikes; each connection delays them according to its current delay kernel.
3. Learning — coincidence at a target neuron reinforces the contributing delays (delay/weight update rules in DelayFunctions/ and Updates/); thresholds adapt to keep firing selective.
4. Evaluation — the network is tested across delay-function families, delay ranges (e.g. 250–300 ms), distribution counts, and signal-to-noise ratios; figures report train-vs-test separability of the learned delay/weight structure.

Running

The Python simulations are driven by the simulation_script_*.py files; set the project path and dataset/category at the top of a script and run it. The MATLAB side generates the stimulus streams and the analysis figures. (Note: the large generated datasets and raw simulation outputs are not committed — regenerate them with the generateStream*.m / buildDataSet*.m scripts and the simulation drivers.)

Background

The design draws on gamma networks for automatic speech recognition, spike-timing-dependent plasticity, Izhikevich spiking-neuron dynamics, and delay-learning models (Masquelier; Legenstein, Naeger & Maass). See the NYAS poster for the consolidated references and results.

---

Author: Shady El Damaty. Research project on temporal coding and delay learning for speech / temporal pattern recognition.