Biomimicry: Emulating Neurons with Hardware

 

What is Biomimicry?

While the tech industry is often focused on cutting edge innovations, one cannot help question- Why reinvent the wheel? Hence why programmers turned to systems which occur in nature to create computer systems. This process is known as Biomimicry, IE using biological structures and processes to inspire the design of technology such as machine learning and artificial intelligence.

Just as biological evolution is a process of iterative development, so too does repetitive design gradually optimize man-made solutions.

Comparison of human brain v. Fruit Fly, image credit: Louise Crosby / University of Sheffield

Comparison of human brain v. Fruit Fly, image credit: Louise Crosby / University of Sheffield

The Networks of the Brain: Fruit Fly v. Honey Bee v. Human

The mammalian nervous system is complex and diverse, with millions of heterogeneous neuron types, sparsely connected in a complex topology. The current generation of machine learning capabilities are on the order of a fruit fly brain. A fruit fly brain has 100,000 neurons, while a honeybee has 1,000,000 neurons (ten times as many). By comparison, a human brain has ~ 100,000,000,000 (one hundred billion) neurons.

Neural Networks

Artificial neural networks (ANN) are inspired by the biological neural networks of animal brains, which are then simulated in computers.

In order to be comparable to biological systems, the ANN needs the following:

  • Support for data input and output

  • Some kind of memory

  • A capacity to improve its internal state (learn)

  • A way to abstract useful data

  • Tolerance for both stochastic and deterministic phenomena

  • The ability to form concepts.

Technology has not matured to create these full suite of characteristics within ANNs.

The Layers of an ANN

A fitness function for an ANN is analogous to an evolutionary trait in the brain determining survival fitness. In an ANN, neurons are grouped into layers, and degrees of interconnection are specified as weighing factors.

An artificial neural network is a resonating construct that matches data inputs to patterns in its own data structures. That is, an artificial neural network is a self-reinforcing (self-tuning) resonator and comparator.

Image credit: Michael A. nielsen, 'neural networks and deep learning'

Image credit: Michael A. nielsen, 'neural networks and deep learning'

The Movement of Data through ANN

Lowest level neurons are connected to sensors receiving data from the outside world. High level neuron layers are hidden from direct contact with outside data, and are used to process data from other neurons. Patterns in the neural network mimic patterns in the outside data.

The network structure propagates weights through the hierarchy to tune itself to recognizes patterns in the data. Correlations among images or other data gradually accumulate as recognizable patterns that can be detected in new images. Tuning the neural network to resonate with sympathetic vibrations in the data patterns is an iterative process. Data that matches patterns is said to be recognized. The overall process involves setting up initial conditions, creating connections, establishing weights, and determining patterns.

In short, an artificial neural network mimics the brain, while also being a resonator that mimics the data.

By bohdan shmorhay

Bohdan Shmorhay