An Algorithmic Analog for the Brain
Imagine a tool to unlock mysteries of the mind, power new technology, and expand treatment capacities. That's our hope for our new approach to artificial learning.
We are developing biologically plausible learning algorithms that learn complex patterns in a few-shot, unsupervised manner directly from sensory information. Our motivating goal is to create a learning system that accomplishes useful and challenging real-world tasks.
After more than a decade of iteration, Density Networks can separate the complex sounds in its environment into distinct audio stream outputs without training, test data, instruction, or audio preprocessing.
Complex systems theory proposes that incredibly complex systems often emerge from an underlying set of individually simple rules that govern the interactions of autonomous actors. In the brain, an individual neuron can influence and be influenced by neurons around it, but only through localized information exchanges. We see that learning, motivation, memory, motor control, and more emerge as neurons interact, but there is no explicit orchestration of these emergent phenomena.
We have accomplished this while adhering to mathematical and architectural constraints that were initially motivated by complex systems theory and have led us increasingly toward biological plausibility:
Entities are autonomous
Information is local
Experience is temporal
To replicate these emergent behaviors in Density Networks, we have incorporated observations and architecture from neuroscience research, including:
Distinct regions with distinct learning rules
Multiple interacting neurotransmitter types
Interwoven LTP, LTD, and STP mechanisms
AMPA and NMDA receptor dynamics
Neuromodulation via volume transmission
Silent synapses for developing neurons
Tripartite synapses
Multiplexing simultaneous environmental stimuli
Continuous signal from cochlear hair cells
Unlike the calculus, statistics, linear algebra, and multidimensional geometry that powers current deep learning algorithms, Density Network calculations are simple. Computational and behavioral complexity in a Density Network emerges from interactions between the individually simple calculations we define.
Addressing the “Cocktail Party Problem”
A newly instantiated Density Network achieves sound source separation within seconds of listening to an instrumental duet. We are finalizing additional learning behaviors that will enable the network to identify and follow human voices.
Today, a Density Network can hear. In the future, this biologically plausible learning approach is extensible to vision and touch, and eventually motion. These sensory inputs can be combined to achieve multimodal learning within a single, interconnected network, unlocking new possibilities for sensory-motor prosthetics and autonomous robotics.
We also hope there may be applications to neuroscience including diagnostics and treatment. Given the network is modeled directly from observable structures in the brain and all emergent behaviors can be tracked over time with complete transparency from the actor level up, researchers can explore how the network responds to changing variables.
We formed Cambrya two years ago to support the development of this technology. We are passionate about building cross-disciplinary relationships that can help accelerate this vision.