Density Networks in Action

What happens when a Density Network listens to speech?

A standard method for analyzing brain activity is the spike raster plot , which shows how neurons fire in response to external stimuli or endogenous spontaneous activity. If a neuron fires consistently in the context of a specific environmental cue, we can infer this response is associated with the brain’s representation of that stimulus.

Artificial neurons in our hearing Density Network are designed to work like biological neurons -- this means we can use this same neuroscience method to observe the firing of Density Network neurons responding to a set of four voice clips featuring different speakers.

We see the network performing three incredibly important behaviors over the course of 33 seconds.

• Immediately recognize pitches consistently

• Continuously refine its representation of specific phonemes

• Use the stable pitch recognition and dynamic phoneme learning to recognize and appropriately categorize novel inputs

A Density Network is able to achieve this naturally, without training, test data, instruction, or audio preprocessing.

The below spike trains have highlighting boxes to show these behaviors in more detail.

Tonotopic Mapping

Hearing mammals have “tonotopic” areas in their brain, where pitches of different frequencies map to tone-specific neurons.

The top spike train demonstrates neural responses in the tonotopic region of the Density Network as it listens to a voice.

Each neuron is attuned to a specific frequency range, and the activation of these frequencies mirrors the changing intonations of the speaker's voice without training, labeling or explicit instruction.

Learning

We believe biological learning is fractalized, with abstractions being layered on top of the initial sensory input and on themselves.

The bottom spike train shows activity in the region of the Density Network responsible for learning phonemes.

Initially, the neurons in this region parse vocal sounds into broad phonetic categories (arrows on the left). As these generalized neurons stabilize, additional neurons step in to help break the phonemes into more specific phoneme representations (arrows on the right).

Recognizing

The most impressive part of biological learning is how quickly we are able to use new experiences to understand the world around us.

When a Density Network is given a novel stimulus, it continues to correctly identify the intonation of the new speaker. It also recognizes that the speaker's words contain phonemes that it has heard before from other voices, and accurately reactivates neurons that represent those phonemes.