Top‑Down and Bottom‑Up outline
Behaviour reflects a negotiation between model‑driven expectations (top‑down) and data‑driven signals (bottom‑up).
Have you ever looked at one of those ambiguous figures---a picture that can be either a duck or a rabbit---and had it suddenly flip when someone mentioned one of the animals? Or misheard a lyric for years, then read the actual words and found you couldn't un-hear it? Or walked straight past the turn you meant to take because your mind was on something else, until a flashing light or an unexpected sound snapped you out of it?
In each case, what you perceived was the result of a negotiation between two streams of information: top-down (what you expect, what you want, what your goal is) and bottom-up (what's actually there in the raw sensory data). Behaviour emerges from that negotiation.
Top-down carries your model of the world---your goals, expectations, and prior knowledge. When top-down precision is high, the model fills in gaps, suppresses distractions, and keeps you on track so you can act with confidence and speed. Bottom-up carries the raw signals from your senses. When bottom-up precision is high, fresh data can move the model, letting you notice surprises and adapt.
The brain constantly weighs these two streams. Attention and control set the mix. Sometimes the model wins and you see the duck because someone said "duck." Sometimes the data wins and the flashing light pulls you off your planned route. Most of the time, it's somewhere in between: the model shapes what you notice, and what you notice subtly updates the model.
So what can you do? The lever is precision---how much weight the system gives to each stream. Tighten the model (raise top-down precision) when stability and focus matter: preload goals, prompts, or constraints at the cue, and the system will stay on rails. Loosen the model (raise bottom-up precision) when you need change, insight, or adaptation: reduce the salience of familiar patterns, add novelty, or shift the context so alternative groupings can surface.
Overdo it in either direction and you get problems. Too much top-down gives you tunnel vision---you miss what's actually there because the model is too strong. Too much bottom-up gives you noise and indecision---every stray input pulls you in a different direction. The art is in knowing which mode you need and designing the situation to support it.
Let's look at the neural architecture that helps us understand this heuristic better.
What neural architecture makes this happen?
These neural systems underpin this heuristic:
The neural hardware
The underlying neural systems that make this heuristic possible
Chunking & Binding
The first place top-down and bottom-up meet is in chunking and binding. Bottom-up regularities---things that are similar, close together, or form continuous lines---automatically bind features into units. But top-down expectations can bias which groupings 'pop out'. What you notice, before any explicit reasoning happens, is already a negotiated product of both streams.
This means that to reveal new patterns, you need to relax top-down constraints and tweak the inputs so different groupings become salient. To stabilise a pattern, strengthen the grouping cues you want---make the model tighter, the context clearer.
Key takeaway: to reveal new patterns, relax top‑down constraints and tweak inputs so different groupings become salient; to stabilise, strengthen grouping cues you want.
Task Sets
Task sets are the top-down configurations that filter what enters awareness and what actions are on the menu. When you load a task set---say, "navigate to the shops"---the system binds relevant features, rules, and responses for that goal. The set biases what you notice and suppresses what doesn't fit.
The catch is, if you don't preload the intended set at the cue, bottom-up distractors or the previous set can leak through. The negotiation tips towards whatever was already active or whatever is most salient. To get top-down to win the selection, you need the right set loaded before the cue arrives.
Key takeaway: preload the intended set at the cue so top‑down wins the selection; avoid unnecessary set switches to reduce leaks from bottom‑up distractors.
Working Memory & Control
Working memory is where you actively maintain and manipulate the few items that matter right now. It's costly, so you don't want to use it to run the whole behaviour. Instead, use it to set precision: which signals to weigh more heavily (top-down or bottom-up), and then let the trained routines run.
Think of working memory as the dial that adjusts the negotiation, not the engine that runs the whole show.
Key takeaway: use control briefly to set precision (what to weigh more/less) and then let trained routines run.
Hierarchical Control
The brain's control systems are layered: higher layers set goals and constraints, lower layers execute routines and resolve details. At each level, there's an arbitration process that sets the relative influence of expectations (top-down) and evidence (bottom-up).
Sometimes the problem is at the top---the goal is wrong or unclear---and shifting constraints there reshapes everything downstream. Sometimes the problem is at the bottom---the signals are too noisy or too rigid---and you need to adjust there instead. Choosing the right level matters.
Key takeaway: choose the level to adjust---sometimes shifting constraints at the top is cleaner than fighting signals at the bottom (and vice versa).
Interoception & Affect
Your body's state tunes the negotiation. Stress and arousal bias which signals get amplified, changing how quickly top-down yields to bottom-up or holds firm. When you're stressed, the system tends to tighten the model---top-down wins, focus narrows, and you stick to the plan. When you're calm, the system broadens---bottom-up signals get more weight, you sample more evidence, and alternative interpretations can surface.
This is why state management matters. Regulating your body's state is regulating the balance between streams.
Key takeaway: regulate state to set the negotiation---calm broadens evidence sampling; urgency tightens top‑down.
Neural Pathways
Learned mappings make familiar interpretations fast and sticky. Expectations (top-down) accelerate the route you've practiced most. But surprising inputs (bottom-up) can push the system onto an alternative path, if they're strong enough.
Practice determines which interpretation or action is the path of least resistance. To shift the balance, you can either change the inputs (strengthen bottom-up) or rehearse alternatives (build new top-down routes).
Key takeaway: practice determines which interpretation/action is cheap; change inputs or rehearse alternatives to shift defaults.
Networks vs Regions
It's not one spot in the brain that sets the balance between top-down and bottom-up; it's the configuration of networks. Control networks, salience networks, and sensory networks coordinate moment-to-moment which signals dominate. The task and environment recruit different configurations, which is why changing the situation can change the balance more effectively than trying to 'will' it.
Key takeaway: adjust the task and environment to recruit the right network configuration for the desired balance.
Sources
- analects/making-meaning-in-the-brain.md