Learning path
from first bitstream to proven silicon
One route through the whole field, in eight levels. Each builds on the last, and each names three things: what to learn, what to run, and the checkpoint that tells you it stuck. Times assume comfort with Python and basic linear algebra. Read across, or jump to your level.
The route
0
What is stochastic computing?~30 min
Understand why a stream of random bits can stand in for a number, and how that makes a multiplier one gate.
Checkpoint: you can explain why P(A and B) = P(A)·P(B) for independent Bernoulli streams.
1
Core primitives~1 h
Encode values into streams and spikes, build a neuron, connect a synapse, and read a firing rate back out.
Checkpoint: you can encode a value, feed it through a synapse into a neuron, and recover the rate.
2
Networks & layers~2 h
Compose neurons into dense, convolutional and recurrent layers, and run multi-step simulations.
Checkpoint: you can build a two-layer network, run it for a few hundred steps, and plot the spike raster.
3
Learning & plasticity~2 h
Train with spike-timing plasticity (STDP), reward-modulated three-factor rules, and surrogate gradients.
Checkpoint: you can train a network to classify a simple pattern with STDP or a surrogate gradient.
4
Advanced architectures~2 h
Reach for convolutional, recurrent, reservoir, hyperdimensional and attention layers, and cortical microcircuits.
Checkpoint: you can classify MNIST with a spiking network, or run a cortical-column or HDC model end to end.
5
Performance & acceleration~1 h
Move to the compiled multi-language engine, packed-bitstream vectorisation and the GPU path for throughput.
Checkpoint: you can run the accelerated engine from Python and measure its speed-up over the pure-NumPy path on your own machine.
6
Hardware deployment~2 h
Lower a neuron design onto an FPGA: fixed-point conversion, generated Verilog, co-simulation, synthesis and formal proof.
Checkpoint: you can synthesise a LIF neuron with Yosys, verify it against the Python model, and read its resource report.
7
Research & extensionopen-ended
Add your own neuron models, wire them into the compiler, and situate the work against the wider field.
Checkpoint: you can add a custom neuron model, connect it to the compiler, and generate HDL for it.