Jason Wang

UCL neuroscience student building models, signal analyses and scientific software.

SHH and BMP concentration contours within model tissue; exterior diffusion values omitted — **Model Morpher** Press Play, then drag the bar to move between a dentate gyrus and a whole mouse-brain surface, blended through their distance fields. Geometric interpolation, not anatomical correspondence or development.
Open the recording

Research, figures and data

Selected research

01G27, a g.tec BR41N.IO team projectEEG imagery decoding

When in the trial does EEG imagery become decodable?

Team figure: later training trials correct (%) against decision time after the trial trigger, 0 to 8 s, where each time is the end of a causal 1 s window. A bold line stays near the 50% chance level until about 3 s, then rises to around 80%, with fainter lines around it. Marked points: 82/120 at +3.5 s in blue and 99/120 at +4.25 s in orange. Bands label get ready, instructed imagery, and feedback may be on, with its start assumed, not recorded. — Team figure from the G27 presentation, slide 6, unchanged. Highlight bands are a separate overlay.
Full size

How to read this figure · 5 steps

01What happens in one trial

From 0–2 s the participant gets ready. At 2 s the instruction appears (documented in the protocol), followed by 2–3.5 s of instructed imagery. From 3.5 s feedback may be on — that start time is assumed, not recorded in the data.
02Decision time & the 50% chance line

At any decision time on the horizontal axis, a decoder sees only the 1 s of EEG ending at that moment. Each point along the curve is a separate decoder fitted to that shifted 1 s window. The horizontal dotted line at 50% is chance.
03Near chance until ~3 s, then ~80%

Scores stay near the 50% chance line until about 3 s, then rise to around 80% once the 1 s window covers the instructed imagery period and beyond.
04Why +4.25 s beats +3.5 s (and its limit)

Two points are marked: 82/120 at +3.5 s and 99/120 at +4.25 s. Because the +4.25 s window spans 3.25–4.25 s, it overlaps the assumed 3.5 s feedback start by 0.75 s — so the later gain may partly reflect feedback.

Limit: both points were scored on already-inspected later training trials, making this a timing comparison rather than a held-out test gain.
05Calibration result & my role

In a separate check using 60 early calibration trials per session, choosing between CSP and Riemannian decoders scored 75/120 against 82/120 for fixed CSP — selection did not help.

Team project across 3 participants and 6 sessions. I contributed evaluation pull requests and obtained and relayed the organisers’ scoring rules.

Calibration did not pick a better decoder

75/120selector, choosing CSP or Riemannian

82/120fixed CSP

With 60 early calibration trials per session.

Marked in the figure

82/120at +3.5 s

99/120at +4.25 s

3.25–4.25 s overlaps assumed 3.5 s feedback start by 0.75 s; timing comparison on inspected trials, not a held-out gain.

View the team presentation

Team figure: a topology fit from 0 to 1 for 12 recording files, participants P1 to P3, pre and post, training and test, under two candidate electrode layouts. The layout in the bundled paper fits near 0.9 for every file except P1 post, where the layout drawn in montage.png fits better. — **Which electrode layout fits the recordings?** Higher means nearer electrodes carry more similar signal: a negative rank correlation used as a topology-fit proxy, not accuracy or probability. The mapping is inferred and left–right orientation is unresolved. 12 recording files from 3 participants. Team figure, slide 8, unchanged.
Full size

Evaluation design: early calibration informs a choice of CSP or Riemannian decoder, compared with fixed CSP on later trials at one fixed decision time. — **How the decoders were compared.** Design schematic, with no measured accuracy values. This was a retrospective comparison using previously inspected later trials.

Full G27 account: method, my part and limits

Question. Our team tested whether early EEG calibration data could select between common spatial patterns (CSP) and a Riemannian decoder for each session, and how decoding depends on when in the trial the decision is made.

Calibration. With 60 early calibration trials per session, the selector scored 75/120 against 82/120 for fixed CSP. Selection did not beat the fixed decoder.

Timing. Each point in the timing figure is a separate fit on a causal one-second window ending at that decision time, shifted along the trial and scored on later training trials that we had already inspected. The marked points are 82/120 at +3.5 s and 99/120 at +4.25 s. Feedback onset was assumed, not recorded, and the trials were not held out, so the difference is a timing comparison, not a held-out gain. The data cover 3 participants and 6 sessions.

Electrode layout. The electrode mapping was inferred and its left–right orientation is unresolved. The layout figure compares two candidate layouts for each of 12 recording files from 3 participants, not 12 independent participants. Its fit is a negative rank correlation, not accuracy or probability.

My part. I shared evaluation work through pull requests and obtained and relayed the organisers’ scoring rules, distinguishing their peak-time score from our fixed-time comparison. The figures are the G27 team’s, shown unchanged from the team presentation.

Team presentation · Figure sources and checksums (JSON)

02MorphBrainDevelopmental modelling

Contours of SHH and BMP within the model tissue, beside the same computed state's dorsal, intermediate and ventral positional identities. — The same computed state, shown as tissue fields and a positional readout. Exterior diffusion values are omitted from this view. Inspect the figure

Move a signal.
Change a pattern.

A workbench for exploring how molecular signals assign regions in a neural tissue model.

My contribution: I developed reaction–diffusion components and compared SHH/BMP signalling conditions.
Try the model: Move or remove signalling sources and compare the regions assigned by the model.

The folding aim remains open. The public demo shows illustrative patterning, not a biologically validated growth-to-folding model.

Explore MorphBrain Open the model

Study notes

Fly feeding circuit: study, figures & replay

Fly feeding circuit

Does disconnecting Roundup, a circuit neuron, reduce activity in motor neuron MN9?

Selected model connections around Roundup and motor neuron MN9. Roundup's outgoing connections are marked as cut; its incoming connections remain. — Selected model connections; schematic layout. The experiment cuts all Roundup outputs; only selected links are shown. Incoming connections remain. This is wiring, not activity flow.

Fifteen paired simulated MN9 spike counts at recurrent gains 0.8, 1.0 and 1.2. Each matched seed has fewer spikes after Roundup’s outgoing weights are zeroed. — MN9 motor-neuron spikes over one simulated second. Each line joins a matched input seed; five seeds at each connection strength.

Connection-strength study

I proposed and directed a study of whether removing Roundup’s outputs still reduces MN9 activity when connection strength changes.

In a 127,400-neuron implementation of a published fly model, setting Roundup’s outgoing weights to zero reduced MN9 spiking in all 15 paired comparisons: three connection strengths and five input seeds.

The result concerns one motor neuron and input regime in a computational model; it does not establish a change in real feeding behaviour.

Explore the circuit comparison

Methods & sources

Thirty one-second simulations compare Roundup’s outgoing connections kept or cut at three recurrent gains and five matched input seeds.

Within each seed, both conditions at all three gains shared candidate and accepted sensory inputs. A refractory-eligibility filter rejected some candidates; accepted inputs matched across these runs. The shared-input design and an execution correction were adopted after initial failures; the design was not preregistered.

The five seeds represent alternative model inputs, not animals. The comparison covers one neuron intervention, motor output and input regime. The source paper corrects left–right labels in its dataset; genetic-driver coverage and historical cell-ID lineage are not established here.

Shiu et al. (2024), source paper · Pinned source model

Download aggregate results (CSV) · Evidence and source hashes (JSON)

Network selection and sources · Original full results figure · Video and chart provenance · Separate body-replay source. The chart shows the 15 matched simulation pairs. The video covers −0.4 to 2.0 simulated seconds, using a separate motor archive and an illustrative movement mapping.

Explore the 3D fly

Replay simulated motor spikes alongside whole-fly and mouth views. Rotate the fly or step through the activity in the interactive replay. The recorded replay is illustrative movement from a separate simulation archive, not the paired comparison.

Play the recorded replay Explore the 3D replay

Input timing

Does timing change the effect of cutting more connections?

I proposed and directed a comparison of cutting only Roundup’s connection to MN9 with cutting all its outputs. Across five artificial input templates, shifted timing increased, decreased or left unchanged the difference between the cuts.

No consistent direction across these five templates. This is a simulation result from one fixed model and measurement window.

Explore the input-timing study

Separating neural signals: study preview

Separating neural signals

Can stimulus and choice be separated when errors are rare?

When stimulus and choice usually agree, can the analysis tell their signals apart? In this synthetic study, a component can be strongly labelled as choice-related while reconstructing the known choice signal poorly. With no error trials, stimulus and choice cannot be separated by this fit.

Six trial mixes · five noise seeds · synthetic neural activity

Supporting synthetic methods exploration developed with AI assistance.

Explore the rare-error experiment

Five synthetic noise-seed results at one error per stimulus: mean choice attribution 97.63%, with mean normalized reconstruction error 0.1000. Each plot shows all five seeds, on separate scales; ranges are not confidence intervals.

One error per stimulus. Explore the other trial mixes.

EEG source modelling

UCL

Undergraduate Research Assistant

I work with Professor Lorenzo Fabrizi in UCL’s Medawar Pain and Somatosensory Labs. My work focuses on EEG source modelling: relating models of brain activity to signals measured at the scalp.

About & contact

I study neuroscience at UCL and build tools for exploring brains and data. I’m particularly interested in simulation, machine learning and making scientific ideas easier to inspect.

I’m interested in paid internships, scientific software work and research opportunities in brain modelling, computational neuroscience and machine learning.

jason.wang.25@ucl.ac.uk

Request my CV GitHub ↗LinkedIn ↗

Jason Wang

Model Morpher

MorphBrain

Can measurements distinguish diffusion from clearance?

Morphliner

Selected research