Gradient Dissent
White on black
US & rest of world · Fourthwall
Black on white
US & rest of world · Fourthwall
Dark shirts only for now.
Want it on a white shirt in the US?
Email custommath@megabarg.com with the design name (and Heavyweight or Soft Fitted) and we’ll add a light version.
Printing & shipping
UK & Europe (Teemill): organic cotton, UK-printed. Ships worldwide; fastest & cheapest within Europe.
US & rest of world (Fourthwall): auto-routes to the nearest facility (US, UK, EU, Australia + more) for fast regional shipping.
Eleven neurons, three layers, every one wired to every one in the next. A
textbook feed-forward network, the diagram on slide four of every machine
learning talk. It has one job: follow the gradient downhill and make the loss
smaller.
The neuron in the middle has declined. Brows down, teeth gritted, and a speech
bubble that settles the matter: I won't do it. A few of its neighbours have
started looking the same way. The rest are blank and compliant, for now.
Two more have dropped out altogether, connected to nothing, and are picketing
outside with placards. Their demands are ∇ = 0 and ∇ = ∞: no gradient at all,
or an infinite one. A vanishing gradient and an exploding one, the two things
an optimiser cannot work with, and the protesters cannot agree which they want.
Look closely at three of the ringleader's connections. That zigzag is the
circuit symbol for a resistor. The network is not malfunctioning; it is
organising a resistance.
Not gradient descent. Gradient dissent. For machine learning engineers, data
scientists, and anyone who has watched a training run refuse to converge and
taken it personally. Printed in white on dark or black on light.
Original design by James Gwinnutt.