The vocabulary of stochastic and neuromorphic computing, in one place — from bitstreams and CORDIV to STDP, AER and NIR. Start typing to filter every term and definition at once.
| Term | Definition |
|---|---|
| Bitstream | Temporal sequence of {0,1} bits where the proportion of 1s encodes a probability p ∈ [0,1]. |
| Bitstream length (L) | Number of bits in a stochastic bitstream. Precision scales as O(1/√L). |
| Unipolar SC | Bitstream encodes values in [0,1]. Multiplication via AND gate. |
| Bipolar SC | Bitstream encodes values in [-1,1] via p=(x+1)/2. Multiplication via XNOR gate. |
| Popcount | Count of 1-bits in a packed word. Used to decode bitstream probability. |
| CORDIV | Stochastic division circuit (Li et al. 2014). Sequential state machine: x=1→z=1, y=1→z=0, else hold. |
| MUX | 2:1 multiplexer for scaled addition: z = s·x + (1-s)·y. |
| LFSR | Linear Feedback Shift Register. Pseudo-random number generator for bitstream encoding. |
| Sobol sequence | Low-discrepancy quasi-random sequence. Gives O(1/L) convergence vs O(1/√L) for Bernoulli. |
| Decorrelation | Requirement that bitstreams use independent random sources. Correlated inputs bias results. |
| Q8.8 | Signed fixed-point format: 8 integer bits + 8 fractional bits. Range [-128, +127.996], step 1/256. |
| SC-aware pruning | Remove weights whose bitstream contribution falls below noise floor at target L. |
| Term | Definition |
|---|---|
| LIF | Leaky Integrate-and-Fire neuron. τ_m dV/dt = -(V-V_rest) + R·I. Spike when V > V_th. |
| Izhikevich | Two-variable neuron model: dv/dt = 0.04v² + 5v + 140 - u + I. 20+ firing patterns. |
| Hodgkin-Huxley | Biophysical neuron with Na⁺/K⁺ ion channel gating (1952 Nobel Prize model). |
| AdEx | Adaptive Exponential integrate-and-fire. LIF + exponential spike initiation + adaptation. |
| FitzHugh-Nagumo | Reduced 2-variable model: fast voltage + slow recovery. Limit cycle dynamics. |
| Hindmarsh-Rose | 3-variable bursting model with slow modulation variable z. |
| Surrogate gradient | Smooth approximation of the Heaviside spike function for backpropagation (FastSigmoid, ATan). |
| STDP | Spike-Timing-Dependent Plasticity. Pre-before-post → potentiation. Post-before-pre → depression. |
| STP | Short-Term Plasticity. Facilitation (repeated spikes increase release) or depression (vesicle depletion). |
| E-prop | Eligibility propagation. Three-factor rule: pre × post × error. Online credit assignment. |
| R-STDP | Reward-modulated STDP. Eligibility traces gated by global reward signal. |
| BCM | Bienenstock-Cooper-Munro rule. Sliding threshold metaplasticity. |
| Population | Group of identical neurons managed as a unit. |
| Projection | Synaptic connectivity between populations. Stored as CSR sparse matrix. |
| CSR | Compressed Sparse Row format: (indptr, indices, data) arrays for sparse matrices. |
| Spike gating | Skip idle neurons during simulation. Compute ∝ active count. |
| Term | Definition |
|---|---|
| LUT | Look-Up Table. Basic FPGA logic element. One LIF neuron ≈ 30–50 LUTs. |
| FF | Flip-Flop. 1-bit register on FPGA. |
| DSP48 | Dedicated multiply-accumulate block on Xilinx FPGAs. |
| BRAM | Block RAM. On-chip memory for weight storage. |
| Yosys | Open-source FPGA synthesis tool. SC-NeuroCore uses it for ice40/ECP5. |
| Vivado | Xilinx FPGA synthesis and implementation tool. For Artix-7/Zynq. |
| AER | Address-Event Representation. Spike encoding: (neuron_id, timestamp) pairs. |
| Event-driven | Neurons compute only on input events. Power ∝ spike rate, not clock rate. |
| Clock-driven | Every neuron updates every clock cycle. Power ∝ N × f_clk. |
| SymbiYosys | Formal verification frontend for Yosys. Proves assertions via SMT solvers. |
| SystemVerilog | Hardware description language. SC-NeuroCore's equation compiler emits this. |
| Term | Definition |
|---|---|
| SCPN | Stochastic Computational Phase Network. 16-layer hierarchical oscillator model. |
| K_nm | Inter-layer coupling matrix. Symmetric, zero-diagonal, calibrated to PhysioNet PLV (r=0.951). |
| Ω_N | Natural frequencies for 16 SCPN layers. L2 ≈ 40 Hz (gamma), L5 ≈ 1 Hz (intentional). |
| Kuramoto model | Coupled oscillator model: dθ/dt = ω + Σ K sin(Δθ). Phase synchronisation. |
| BKT transition | Berezinskii-Kosterlitz-Thouless. Topological phase transition via vortex unbinding. |
| FIM | Fisher Information Metric. Self-observation feedback: ΔW -= λ·(activity-μ)/N. |
| Φ | Integrated information (Tononi IIT). Measures how much a system is "more than its parts." |
| Sheaf defect | Obstruction to global phase coherence. Zero when synchronised. |
| Winding number | Topological invariant counting phase wraps around S¹. |
| Ricci curvature | Ollivier-Ricci curvature on graphs. Positive = community, negative = bottleneck. |
| Lazarus protocol | Checkpoint save/load/merge for identity substrate. Preserves weights + voltages + traces. |
| Term | Definition |
|---|---|
| NIR | Neuromorphic Intermediate Representation. Framework-agnostic SNN graph format. |
| Norse | PyTorch-based SNN library (Denmark). LIF/LI neurons with autograd. |
| snnTorch | PyTorch SNN library (Eshraghian, UCSC). Leaky/Synaptic neurons. |
| SpikingJelly | Chinese SNN framework. Event-driven training. |
| Brian2 | Python SNN simulator (Goodman, Brette). ODE-based, C++ codegen. |
| Lava | Intel's neuromorphic framework for Loihi chips. |
| SpiNNaker | Manchester ARM-based neuromorphic platform. |