This page was generated from docs/devices/xylo-overview.ipynb. Interactive online version: Binder badge

🐝 Overview of the Xylo family

Xylo is a family of spiking neural network ASICs, for efficient simulation of spiking leaky integrate-and-fire neurons with exponential input synapses. Xylo is highly configurable, and supports individual synaptic and membrane time-constants, thresholds and biases for each neuron. Xylo supports arbitrary network architectures, including recurrent networks, residual spiking networks, and more.

Xylo is currently available in several versions with varying HW support and front-ends.

[1]:
# - Image display
from IPython.display import Image

Image("images/xylo_network-architecture.png")
[1]:
../_images/devices_xylo-overview_2_0.png

This figure shows the overall logical architecture of Xylo.

  • Synchronous time-stepped architecture with a global time-step dt

  • Up to 16 event-based input channels, supoorting up to 15 input events per time-step per channel

  • 8-bit input expansion weights \(W_{in}\)

  • Up to \(N_{hid} = 1000\) digital LIF hidden neurons (see below for neuron model)

    • two available input synapses

    • bit-shift decay on synaptic and membrane potentials

    • Independently chosen time constants and thresholds per neuron

    • Subtractive reset, with up to 31 events generated per time-step per neuron

  • 8-bit recurrent weights \(W_{rec}\) for connecting the hidden population

  • 1 output alias supported for each hidden neuron -> copies the output events to another neuron

  • 8-bit readout weights \(W_{out}\)

  • Up to \(N_{out} = 8\) digital LIF readout neurons (Xylo v1, Xylo Audio v2)

    • same neuron model as hidden neurons

    • only one available input synapse

    • only one event generated per time-step per neuron

[2]:
Image("images/xylo_neuron-model.png")
[2]:
../_images/devices_xylo-overview_4_0.png

The figure above shows the digital neuron model and parameters supported on Xylo (Xylo v1 and Xylo Audio v2)

  • Each of the two input synapses may receive events from up to 16 pre-synaptic neurons

  • Up to 15 input spikes are supported per time step for each input channel

  • Reset during event generation is performed by subtraction

  • Multiple events can be generated on each time-step, if the threshold is exceeded multiple times. Up to 31 spikes can be generated on each time-step

  • Each hidden layer neuron supports two synaptic input states. Output layer neurons support only one synaptic input state

  • Synaptic and neuron states undergo bit-shift decay as an approximation to exponential decay

[3]:
# - Switch off warnings
import warnings

warnings.filterwarnings("ignore")

# - Useful imports
import sys
!{sys.executable} -m pip install --quiet matplotlib
import matplotlib.pyplot as plt

%matplotlib inline
plt.rcParams["figure.figsize"] = [12, 4]
plt.rcParams["figure.dpi"] = 300

try:
    from rich import print
except:
    pass

import numpy as np

Bit-shift decay

Xylo devices simulate exponential decay using an efficient bit-shift-subtraction technique. Consequently, time constants are expressed as β€œbit-shift decay” parameters (β€œdash” parameters) which are a function of the fixed simulation time-step. The equation for that conversion is:

\(dash = [\log_2(\tau / dt)]\)

Hence, the dash parameter for a synapse with a 2 ms time constant, with a simulation resolution of 1 ms, is 1.

But what is done with this bitshift of 1? Let’s compare exponential decay with bitshift decay. The bit-shifting decay method is illustrated by the function bitshift() below.

[4]:
def bitshift(value: int, dash: int) -> int:
    # - Bit-shift decay
    new_value = value - (value >> dash)

    # - Linear decay below dash-driven change
    if new_value == value:
        new_value -= 1

    return new_value

As you can see, the bit-shifting decay is accomplished by the line

\(v' = v - (v >> dash)\)

where \(>>\) is the right-bit-shift operator; \(v\) is the current value and \(v'\) is the new value after the bit-shifting step. Since \(v\) is an integer, for low values of \(v\), \(v >> dash = 0\). We therefore perform a linear decay for low values of \(v\).

[5]:
def plot_tau_dash(tau: float, dt_: float = 1e-3, simtime: float = None):
    if simtime is None:
        simtime = tau * 10.0

    # - Compute dash, and the exponential propagator per time-step
    dash = np.round(np.log2(tau / dt_)).astype(int)
    exp_propagator = np.exp(-dt_ / tau)

    # - Compute exponential and dash decay curves over time
    t_ = 0
    v_tau = [1000]
    v_dash = [1000]
    while t_ < simtime:
        v_tau.append(v_tau[-1] * exp_propagator)
        v_dash.append(bitshift(v_dash[-1], dash))
        t_ += dt_

    # - Plot the two curves for comparison
    plt.plot(
        np.arange(0, len(v_tau)) * dt_ * 1e3,
        v_tau,
        label=f"exponential decay, $\\tau$={int(tau * 1e3)}ms",
    )
    plt.plot(
        np.arange(0, len(v_dash)) * dt_ * 1e3,
        v_dash,
        "--",
        label=f"bitshift decay, dash={dash}",
    )
    plt.legend()
    plt.xlabel("Time (ms)")


# - Plot examples for tau = 16ms and tau = 25 ms
plt.figure()
plt.subplot(1, 2, 1)
plot_tau_dash(16e-3)

plt.subplot(1, 2, 2)
plot_tau_dash(25e-3)
../_images/devices_xylo-overview_11_0.png

For powers of two, the approximation is very close. But since this is integer arithmetic, the approximation will not always be perfect. We can see this with the example of \(\tau = 25\)ms.

Next steps

For information about setting up a Xylo simulation and deploying to a Xylo HDK, see 🐝⚑️ Quick-start with Xylo