Skip to content

Neural Networks overview

NeuroSeq is a neuronal sequencer: it controls the input to a neural network and reads out its output, whether the network is digital, biological or neuromorphic. For any network, input is split into controlled (what NeuroSeq delivers) and uncontrolled (everything else), and output into observed (what is recorded) and unobserved. The same input protocol and the same analysis run on every substrate.

Input arrays have one row per neuron. On biological networks, neuron ID matching links each row to a real cell through a Cellpose segmentation (File → Cell segmentation, ID Mapping Options).

Terminology

  • Import: loading files of any kind into a project. File → Import data has an Input submenu (e.g. SysCon2 .seq, patch-clamp command waveforms) and an Output submenu (calcium imaging, patch-clamp recordings).
  • Input: the arrays that drive the network, one row per neuron (input arrays, shown in the Input panel and chosen with the Input data selector). They are designed with static or dynamic input design and imported or exported as files.
  • Output: what is read out from the network: traces, spikes and synapses, simulated or recorded (selected with Neuron data, Spike data, Synapse data).
  • Stimulation: the physical delivery of input on a rig, e.g. holographic optogenetic stimulation through SysCon2, configured in the Stimulation setup pane. A stimulation artefact is laser light picked up by the imaging.

Supported substrates

These are the integrations in use today. They are starting points for new substrates, third-party software and APIs rather than a fixed list.

Substrate Network Input Readout
Digital Brian2 Input arrays drive the built-in model Simulated traces, spikes, synapses
Digital External simulators Input arrays passed to your Python simulate(...) The Experiment your backend returns
Biological Holographic stimulation Two-photon holographic optogenetics on HOLO4D: export .seq, upload and run through SysCon2 ScanImage calcium imaging, started or armed through its MATLAB API: ΔF/F and deconvolved spikes per Cellpose ROI
Biological Patch clamp Export protocols for the acquisition software Import patch-clamp data
Neuromorphic BrainScaleS-2 PyNN runs on the remote chip through a container helper: constant current per neuron, or spikes from a virtual source population; chip runs 1000x faster Spikes of all driven neurons and the membrane voltage of one neuron

Workflow

  1. Design input with the static input editor or a dynamic input routine (rheobase, STDP).
  2. Deliver it: run a simulation, export a file, or send it to a connected device from the Stimulation setup pane (View → Show Stimulation setup).
  3. Read out: the result is added to the project as a data group of traces, spikes or imaging data.
  4. Analyse and adapt. Dynamic input routines close this loop automatically on Brian2 and SysCon2 + ScanImage (rheobase also on BrainScaleS-2).

Interaction mechanisms

How NeuroSeq talks to a network depends on what its software and hardware offer:

  • File export and import: export a protocol, run it in the other application, then import the readout. Existing protocols can be imported where the format is supported.
  • Direct API: configure and run a simulator or device through its API and get the response back.
  • Combined: generate protocol files, upload and run them through a control API, and import the recordings automatically, as with SysCon2 holographic stimulation plus ScanImage acquisition.

A file-based connection is useful on its own or as one step of an automated experiment. New backends can be added through the custom simulator interface.