AI Glossary / AI and machine learning / Advanced
Three-factor learning
Definition
Three-factor learning is a family of synaptic learning rules in which changes depend on presynaptic activity, postsynaptic activity and an additional modulatory signal. The third factor can influence whether or how local neural activity produces a lasting connection change.
What it means
A basic Hebbian idea links learning to the activity of connected neurons. Three-factor rules add another signal that can represent reward-related information, error or a modulatory condition, depending on the model.
This helps address temporal credit assignment: an action may occur before its outcome is known. A temporary eligibility trace can mark recently active synapses, allowing a later signal to influence the relevant connections. The trace does not necessarily change the weight by itself.
The term describes a broad modeling framework, not one universal equation or a complete established account of biological learning. Different brain regions, signals and experiments involve different mechanisms. Computational usefulness and biological evidence should be discussed with the scope of the particular rule or study in view.
How it works
Local pre- and postsynaptic activity generates an eligibility signal according to the model's timing and activity rules. That signal can decay over time, retaining a temporary record of which connections may deserve an update.
A third factor then modulates the effect on synaptic strength. In a simplified mathematical form, a weight change can be proportional to an eligibility trace multiplied by a modulatory signal. The sign and magnitude of the signal can determine whether a connection strengthens or weakens.
Not every modulatory signal is truly global, and not every synapse receives the same biological influence. Some models use additional factors or more complex dynamics. Experiments and simulations test how timing, signal delivery and synaptic state affect learning. A convenient three-term formula should not be mistaken for a literal universal recipe used identically throughout the brain.
Three practical examples
1. A computational-neuroscience model combines local activity traces with a delayed reward-related signal to learn from an outcome.
2. A neuromorphic research prototype explores local connection updates modulated by an error or reinforcement signal.
3. A laboratory study examines whether prior synaptic activity changes how a later neuromodulatory signal affects plasticity.
Think of it this way
Think of placing temporary markers beside recent actions, then using later feedback to decide which marked actions should be reinforced. The limit is that biological eligibility traces are specific mechanisms with their own timing and conditions, not conscious notes or a complete explanation of learning.
Worked example
Consider a toy rule: weight change equals learning rate times eligibility times modulation. Let the learning rate be 0.1, eligibility be 0.4 and modulation be 0.5. The change is 0.1 × 0.4 × 0.5 = 0.02.
If the modulation is -0.5 instead, the same eligibility produces a change of -0.02. If the trace has decayed to 0.1 before the positive signal arrives, the change is only 0.005.
This example illustrates how a delayed third factor can interact with a local trace. It is not a full biological model, and actual rules may include bounds, nonlinearities and additional state variables.
Illustrative case study
A fictional neuroscience teaching group builds a simulation in which a reward arrives after an action. Without a trace, the model struggles to connect the outcome with the earlier activity. Adding a decaying eligibility variable lets the delayed signal affect recently active connections.
The group then makes the delay much longer. Learning weakens because the trace largely disappears before feedback arrives. Students compare different time constants and discuss what evidence would be needed to connect the toy model to a biological system.
The case illustrates both the purpose and the limit of eligibility traces: they can bridge a delay, but their duration and specificity matter.
Common misconceptions and mistakes
- Three-factor learning means three neural-network layers. It refers to influences on a learning rule, not layer count.
- The third factor is always one global reward signal. Its interpretation and delivery depend on the model and biology.
- An eligibility trace is already a permanent weight change. It can be a temporary state awaiting modulation.
Business use
Three-factor learning is most relevant to neuroscience, biologically inspired AI and neuromorphic research. It offers ideas for local learning and delayed feedback, but it is not a general guarantee of efficient business-model training.
For a practical research proposal, ask which signals are available, how the trace is maintained and what objective the system learns. Evaluate timing, noise and credit assignment explicitly. Distinguish a biologically motivated design from a biologically validated mechanism, and compare performance with suitable conventional learning methods for the intended task.
People also ask
What are the three factors?
Typically presynaptic activity, postsynaptic activity and an additional modulatory signal, though formulations vary.
Why use an eligibility trace?
It preserves a temporary record of local activity so later feedback can influence the relevant connection.
Is every rule a reinforcement-learning rule?
No. Third factors can represent different forms of modulation or error, depending on the model.
Check your understanding
Why might a very delayed reward produce little learning in a model with a rapidly decaying eligibility trace?
Answer: The trace identifying the earlier relevant activity may have weakened before the modulatory signal arrives.
Related terms
- Hebbian learning
- Eligibility trace
- Reinforcement learning
- synaptic plasticity
- neuromodulation
- temporal credit assignment
- spike-timing-dependent plasticity
Learn next
Read Hebbian learning and eligibility traces. Vary the delay between local activity and feedback in a toy model, and identify which assumptions determine whether the earlier event still receives credit.
More in AI and machine learning
- Test set - A test set is a collection of examples reserved for assessing a model after its development choices have been made.
- Text classification - Text classification assigns one or more categories to a piece of text.
- Text span - A text span is a specific continuous portion of text identified by its start and end positions.
- Threshold (for decision trees) - A threshold in a decision tree is a value used in a comparison at a branching point.
- Time travel (for features) - Time travel for features is the ability to retrieve historical versions of model-input values.
- Timestep - A timestep is one position in a sequential calculation.
Sources
Gerstner and colleagues, Eligibility Traces and Plasticity on Behavioral Time Scales (2018): https://www.frontiersin.org/journals/neural-circuits/articles/10.3389/fncir.2018.00053/full . Independent review read sections 2.2–2.3 on eligibility flags, third-factor modulation and timing. Frémaux and Gerstner (2016): https://www.frontiersin.org/journals/neural-circuits/articles/10.3389/fncir.2015.00085/full . Consulted modulation and eligibility discussion. PMC copies returned browser challenges; accessible publisher texts support the explanation. The arithmetic and fictional teaching case are original; no individual experiment is reproduced.
Previous: Thread pool Next: Three-layer architecture