← Research
Research poster · 2025
Synaptic performance of biological neurons
Omer Faizan · M.S. Data Science, Montclair State University
1 Question
Real synapses change with timing, recent activity and noise. Which of these behaviours do artificial spiking models leave out, and does it matter for learning?
2 Why it matters
The STDP rules used in machine learning are simplified biology. The missing pieces might explain why deeper STDP layers learn so little.
~86 Bneurons in the human brain
~1 msone spike, a ~100 mV swing
<1%of cortical neurons strongly active at once
3 Timing: STDP
Weight change vs. spike timing (shape after Bi & Poo, 1998)
input before output input after output
4 Recent activity: depression
Response to a 20 Hz spike train (Tsodyks–Markram model, illustrative parameters)
5 Reference points
| Property | Typical |
|---|---|
| STDP time constants | ~10 ms |
| Depression window (neocortex) | often wider than potentiation |
| Brain power budget | ~20 W |
6 Biology vs. my model
biology: typical published ranges · my model: projected
| Behaviour | Biology | My model |
|---|---|---|
| STDP timing window | ~10–20 ms | 16 steps |
| Depression, 2nd spike | −40 to −60% | −45% |
| Recovery time | ~0.5–1 s | 0.5 s |
| Facilitation | +20 to +50% | +25% |
| Spike-time jitter | ~1–3 ms | 2 ms |
7 What I expect
- Short-term dynamics change which features a spiking layer learns.
- A little biological noise helps more than it hurts.