Deep learning with air bubbles

Glass identifies faces at the speed of light with zero power

12 July, 2070

7:00 pm

I am tired and exhausted as I walk up the curb towards my front door. Elon Musk’s funeral was sad, happy and then sad again. Almost everybody was stoned in tribute to Mr Musk, but thankfully I am starting to sober up.

Elon Musk (1978–2070) Inventor, Entrepreneur, Hyperlooper

I am slightly sick from the five minute hyperloop ride back home; the world around me is a carousel in slow motion. I really can’t remember my password or the entry code. I just hope I’m at the right house.

At the center of the door is an unassuming rectangle of translucent glass. I am happy to see it today; it knows who I am and will let me inside — no password or passcodes. It just needs my face, like my mother.

The back side of the glass has a panel with five faces, one of which is mine. Right now it must be lighting up — letting everyone know that I am at the front. Soon it will trigger a signal that will unlock the door and activate the air conditioning.

This, as you have surmised, is no ordinary glass. It is a smart glass that is capable of doing deep learning.

Inside the one millimeter thick body, air bubbles have been carefully arranged to mimic a deep neural network. But unlike neural networks from 2019, these are not arranged in discrete layers but in a sort of random jumble.

A light ray enters the glass and is scattered by a bubble to the right. A larger bubble bounces it towards the bottom left from where it gets kicked straight back to the front. Here a tiny bubble, the size of an amoeba , focuses it to the back once again.

Electromagnetic field from incoming light wave is modulated by inclusions in glass — air bubbles and non linear materials. From ‘Nanophotonic media for artificial neural inference’’

The principle of this optical medium has remained more or less the same as when it was first proposed in the Journal of Photonics Research by a team from the University of Wisconsin Madison and MIT in July 2019.

Their paper established the proof of concept for smart glasses with incorporated AI.

An air bubble is a linear scatterer — which means that it scatters light in some fixed proportion to what it receives.

Deep learning however requires nonlinear activation — there must be something that works as a switch. In traditional neural nets these are written into the code as ‘softmax’ functions, ‘RELU’, ‘tanh activations’ etc. Simply put, these activations work like beaver dams — they only transmit information once an internal threshold has been exceeded.

In order to add non linear behavior to the structure, the glass also has a smattering of non linear materials such as optical saturable absorbers and photonic crystals.

After traveling through a dozen air bubbles, a light ray hits one of the nonlinear spots. It is allowed to pass through only if the intensity exceeds a certain limit. Otherwise, that is the end of that train of light.

In the example that the team studied in 2019, the glass simulation learned to recognize handwritten digits. ( It’s interesting that we still use the MNIST dataset today!)

The light from the image of the digit on the left is channeled by the inclusions to the third spot on the right panel. Corresponding to ‘2’. From ‘Nanophotonic media for artificial neural inference’

An image of the digit 2 — shown above — is allowed to fall on the left side of the 2D glass. Individual rays are scattered and re-scattered until they start to collect in one place at the right side — next to the label for ‘2’.

The Wisconsin — Madison — MIT team could not have imagined how their glass is manufactured today.

Thanks to the enormous progress that we have made in the inverse design of photonic structures, these glasses can be made to a very high precision. We can dictate what bubbles go where — to within 10 nano meters!

I should stop staring at the glass and get into the house now. The lack of oxygen in this planet is dizzying, especially in the evening.

But for no reason I think of the logistic regression cross entropy function for this material. It is really the same as that for any other digit recognition algorithm. Except that now the gradient descent function optimizes the dielectric constant of the medium — not a fictitious number in a virtual neuron.

It is kind of amazing to think that in just fifty years, these analog neural networks are found everywhere.

Why wouldn’t they be? They use no power and work at the speed of light!The computation is built into their body — like a super smart river drainage system.

I am rambling now. I should seriously go inside.

But I hesitate. In a minute, I will see the rather unusual sight of my shadow cloning into two when Deimos rises in the Eastern sky.

In the distance I can see ten rocket boosters land in synchronized festivity. Musk is finally gone.

I open the door and go inside.

Photo by Jared Evans on Unsplash