Neural Oscillation
Neural oscillations (also brain waves or cortical oscillations) are rhythmic fluctuations in the activity of neuronal populations. They emerge when many neurons synchronize their Neural Spiking, typically through the interplay of excitatory and inhibitory neurons, and are measured as periodic signals in EEG, MEG or local field potentials. Oscillations are thought to coordinate the timing of neural activity, gating when information is transmitted and integrated between brain areas.
Frequency Bands¶
Oscillations are grouped into frequency bands. Exact boundaries vary across the literature.
| Band | Frequency | Commonly associated with |
|---|---|---|
| Delta | ~1–4 Hz | Sleep; phrase-level and prosodic structure in speech |
| Theta | ~4–8 Hz | Memory, navigation; syllabic rate of speech |
| Alpha | ~8–12 Hz | Inhibition of task-irrelevant processing, attention, listening effort |
| Beta | ~13–30 Hz | Motor control, top-down signaling |
| Gamma | ~30–100 Hz | Local processing and feature binding; phonemic features in speech |
Oscillations in different bands interact through cross-frequency coupling, e.g. the phase of a slow theta oscillation modulating the amplitude of fast gamma activity.
Oscillations in Speech Processing¶
Speech is quasi-rhythmic: syllables occur at roughly theta rate, phonemic features at gamma time scales, and phrases at delta rate. Auditory cortical oscillations entrain to these rhythms, i.e. their phase aligns with the temporal structure of the speech envelope (Giraud & Poeppel, 2012). A proposed function is that theta phase segments continuous speech into syllable-sized chunks, while nested gamma activity samples the fine-grained phonemic information within each chunk. Stimulus detection also depends on oscillatory phase and amplitude, which suggests a rhythmic inhibitory gating of input.
In multi-talker scenes, cortical tracking is stronger for the attended speaker than for ignored speakers, which links oscillatory entrainment to selective attention in the Cocktail Party Problem. Findings relevant to the CHAIS project are collected in Existing Research. On the level of single neurons, Phase Locking describes spikes firing in phase with the stimulus waveform itself.
Relevance for Machine Learning¶
- Harmonic oscillators are a simple mathematical model of oscillatory dynamics. LinOSS uses a bank of forced harmonic oscillators as the hidden state of a state-space model.
- The Oscillatory SSM project asks whether oscillators that entrain to the input, with input-dependent phase and damping, improve phase reconstruction in speech enhancement.
- In Spiking Neural Networks, oscillations and phase coding can be modeled directly.