Inside a Black Box

It is an attempt to create a structure which is intermediate between a natural and a completely digital neural network making neuron links between neighbouring layers optical, neural cells – analog, and the structure inside the network – parallel.

The neural network consists of an input layer, two hidden layers, and an output layer. The input layer is represented by 122 neurons, the second and the third layers – by 122 and 84 neurons, and the output layer – by 50 neurons. The input layers consist only of controlled light-emitting diodes. The neurons of the hidden layers consist of the structure ‘phototransistor – operational amplifier – light-emitting diode’. The links between the layers are represented by optical multipliers with liquid-crystal light valve. The output layer is the structure “phototransistors – analog-to-digital converter”.

The network task is extremely simple: recognition of symbols in swarming of fruit flies. A picture from the camera is transformed into a line of 122 different voltages for each LED of the input layer. Then, the light passes through arrays of mutually perpendicular light conductors with liquid-crystal light valve where light multiplication of matrix by the weight of each link occurs. All the LEDs are linked to all the photodetectors of the next layer via the connection weight. The adjustment of the activation function is performed by the resistor in the feedback circuit of operational amplifier of each neuron. It turns on all the output LEDs of this layer which are connected to these amplifiers, then – the light conductors with liquid-crystal matrix of the next layer (optical multiplier), and so on.

Thus, all the computations inside the network occur simultaneously which approximates it to the natural one. 50 neurons of the last layer are read sequentially and each neuron is activated to its letter or number. The matrices of weights were defined beforehand in the process of the network learning in a simulator where a dataset of handwritten letters was used; the matrices don’t change during the work. On the monitor the following can be seen: what the camera sees (fruit flies), what the neural network sees (11×11 pixels fruit flies), and the result of work a text consisting of a sequence of recognized letters and numbers.

Exhibitions:

2019 / Pavlovo Koltushi / Pavlov Institute of Physiology,
Russian Academy of Sciences “THE NEW ANTHROPOLOGY”