On cognition, memory and learning

In engineering, mechanical sympathy means understanding how a system actually works, rather than treating it as a black box. Once you know its constraints and the way its parts interact, you can work with it instead of against it.

I've been thinking about what it would mean to apply that idea to learning.

We spend an enormous amount of time trying to get information into our heads, and surprisingly little time thinking about what happens to it once it's there. We talk about storing facts and retrieving memories as though the brain were a computer with files on disk. It's a convenient metaphor, but it obscures quite a lot.

Memory depends on networks of neurons whose activity and connections change with experience. Remembering involves reactivating and reconstructing patterns across those networks. If we understand even a little of that machinery, we can make better decisions about how to study, read, and practice.

Hardware

Start with a neuron. It's a cell specialized for receiving and transmitting signals.

Most neurons have a cell body, a branching set of extensions called dendrites, and an axon, which carries signals away from the cell. Dendrites receive input from other neurons. The cell integrates those inputs, and when its membrane reaches the right threshold, it can generate an electrical impulse called an action potential. That impulse travels along the axon.

Neurons communicate at junctions called synapses. At most synapses, an arriving electrical signal causes the release of chemical messengers called neurotransmitters. These bind to receptors on the receiving cell and change how likely it is to fire. Some inputs excite the receiving neuron; others inhibit it.

Not every input has the same influence. A neuron may receive thousands of synaptic connections, each with different strengths and effects. Whether it fires depends on the combination of inputs, their timing, and the neuron's own state.

The strengths of these connections can change. Repeated activity, particularly when neurons are active in particular patterns and under particular conditions, can make some synapses more or less effective. Receptors change, signaling processes change, and sometimes the physical structure of connections changes as well. This capacity for change is called synaptic plasticity. It's one of the mechanisms by which experience alters the brain.

A neuron by itself doesn't tell us much about a memory or an idea. The interesting behavior appears when enormous numbers of neurons interact. Representations of faces, words, movements, places, and concepts emerge from activity distributed across many cells and regions. There isn't necessarily one neuron, or one location, corresponding to a thing you know.

A few regions are worth distinguishing.

The cortex, the folded outer layer of the brain, is involved in perception, language, voluntary movement, planning, and much of what we know about the world. Its regions have different specialties, but they communicate extensively.

The hippocampus, deep in the medial temporal lobe, is particularly important in forming new memories of events and experiences. It helps associate details that are represented in different parts of the brain: a place, a conversation, a feeling, a sequence of events.

The prefrontal cortex contributes to planning, controlling attention, resisting distractions, and keeping information available while we work with it. The basal ganglia are involved in action selection, habits, and reinforcement learning. The amygdala helps process emotionally significant information. The thalamus plays a central role in relaying and coordinating signals, including sensory information.

These aren't independent modules with neat boundaries. They participate in interacting networks, and different kinds of learning draw on those networks differently. Learning a route, recognizing a face, remembering a birthday, and developing a reliable golf swing are all forms of learning, but they aren't the same biological task.

Thinking

A thought involves coordinated activity across neural networks.

Picture your childhood kitchen. You might see the layout, remember the light through a window, or hear a familiar voice. Those pieces aren't all represented in one place. Visual, auditory, emotional, and other systems contribute to what you experience as a single thought.

This distributed organization helps explain association. A song reminds you of someone; that person reminds you of a city; the city brings back a restaurant. Activity in one set of representations can make related representations more likely to become active.

Thinking also requires deciding what to keep active. Working memory lets us hold and manipulate a limited amount of information for a short time: the numbers in a calculation, the premise of an argument, the state of a program we're debugging. It depends on coordinated activity across several brain regions, including prefrontal and parietal areas.

It's limited, and interruptions matter. When I leave a difficult proof to check a message, I don't merely lose the seconds spent checking it. I often have to reconstruct what I was considering, which assumptions mattered, and where I was going. The interruption has displaced part of the context I was holding in mind.

That's a good reason to protect uninterrupted time. Not because distractions somehow erase what we've learned, but because complex thought depends on maintaining a fragile working context.

Remembering

Memory isn't a recording that gets played back.

Suppose you have dinner with a friend. The room, the conversation, the faces, and your emotional response involve different systems. The hippocampus helps bind aspects of that experience so they can be recalled together later.

When you remember the dinner, you're not reading a complete copy from storage. You're reconstructing an experience from information that can still be retrieved. Some details return vividly. Others are missing or altered. Things you've learned since then can influence the reconstruction too.

Memories also change with time. A process called consolidation helps stabilize them, and some memories gradually become less dependent on the hippocampus as connections across the cortex change. The details vary with the type of memory, and this isn't a simple transfer from one storage location to another.

Retrieval is just as important as storage. You've probably had the experience of being unable to remember a name, then hearing one clue and suddenly having it come back. The information wasn't necessarily destroyed. You couldn't get to it from the cues available at the time.

That distinction is useful when studying. Knowing something involves more than having encountered it or even recognizing it when you see it. You need to be able to retrieve it when the situation calls for it.

Learning

At the biological level, learning involves lasting changes to the way neural systems respond and interact. Synaptic plasticity is part of that story, alongside changes in larger networks, neural excitability, and other processes.

At the level we can actually control, the question is simpler: what kinds of experience make knowledge and skills easier to retrieve and use later?

Consider reading a chapter on linear algebra. While the book is open, every definition and explanation is available. Following the argument can feel like understanding it. Close the book, though, and try to explain why the argument works. Often the gaps become obvious immediately.

The difference is between recognition and recall. Recognition gives you the answer, or much of the context needed to produce it. Recall requires you to reconstruct the answer yourself.

This is why testing yourself is so useful. You discover what you can actually produce without help. Attempting retrieval can also improve later retention, especially when you check your answer and correct mistakes.

I find it helpful to treat studying less as consuming explanations and more as repeatedly trying to reproduce and use them. Read something carefully, put it away, and see what remains. Work through an example without looking at the solution. Explain the mechanism in plain language. Return to the source to find out exactly where your understanding broke down.

Repetition matters, but repeating the same experience isn't always the most effective approach. The way you engage with the material matters just as much as how many times you see it.

Sleep and consolidation

Sleep plays a role in stabilizing memories.

During sleep, the brain doesn't simply shut down. Patterns of activity associated with recent experiences can be reactivated, and interactions between the hippocampus and cortex are thought to contribute to memory consolidation. Different sleep stages appear to support different aspects of this process.

That doesn't mean a night's sleep automatically converts something you've read into knowledge. But it does mean sleep belongs in a theory of learning, not merely in a theory of recovery.

Cramming makes the distinction clear. It's possible to learn enough material in one sitting to perform well on an immediate test. Keeping it available weeks later is harder. Spaced practice and sleep give memories opportunities to stabilize and be retrieved again under different conditions.

There's also a limit to what another hour of tired, inattentive studying buys you. Sometimes the better choice is to stop and return with a rested brain.

Studying deliberately

A few techniques follow from these constraints. They're useful not because they're clever hacks, but because they change what the learner actually has to do.

Protect attention. A difficult idea may require holding several pieces in working memory at once. Give yourself enough uninterrupted time to build that context. When possible, keep the problem, notes, and necessary references together, and put unrelated distractions elsewhere.

Build chunks. Working memory can hold only so much at once, but familiar ideas can be organized into larger units. A beginner sees individual chess pieces and possible moves. An experienced player sees recognizable positions and relationships. A beginner reading code may track individual functions; an experienced engineer can reason about a whole subsystem. Those larger structures are learned through experience, not obtained by skipping the details.

Recall before reviewing. After finishing a section, close the book. Write down the argument, draw the system, or work the problem from memory. Then compare your attempt with the source. The missing steps tell you what needs attention.

Space your practice. Revisit material after some time has passed rather than exhausting all your repetitions in one sitting. The delay makes retrieval harder, which can help learning when the attempt is successful or followed by useful feedback. The right interval depends on the material and how well you know it.

Mix related problems. Ten nearly identical exercises can teach you how to execute a procedure. Mixing problem types also forces you to recognize which procedure applies. That's much closer to what happens outside a textbook, where nobody labels the next problem with the method you should use.

Make predictions. Before running code, predict what it will do. Before checking a derivation, decide what result you expect. Before turning the page, try to anticipate the next step of the argument. Then compare the prediction with reality. A surprising result gives you something specific to investigate.

None of this is a substitute for understanding. It's a way of discovering whether understanding is really there, and then improving it.

Reading

I've found it useful to distinguish different kinds of reading rather than treating speed as the main variable.

When approaching an unfamiliar subject, I often want a map before I want the details. I'll look at the headings, figures, introduction, and conclusion to see what the author is trying to establish. A quick survey makes the careful reading that follows easier to organize.

Other passages need to be read slowly. A mathematical definition might contain only one sentence, but every word matters. Rushing past it can make the next ten pages incomprehensible. Subvocalizing, rereading a sentence, or writing a small example may be exactly the right thing to do.

Sometimes I'm looking for a particular mechanism or claim. In that case, selective notes and underlining help me keep track of what I want to revisit. Marking every other sentence usually defeats the purpose.

The most important part often happens after I stop reading. Can I state the author's argument without looking? Can I give an example that isn't in the book? What assumption is doing the most work? What would change if the claim were false?

Writing is especially revealing here. An idea can feel completely clear while I'm reading it, then fall apart when I try to put it into my own words. The effort of finding a sequence of sentences exposes holes that passive reading lets me overlook.

A finger or pen can help guide the eyes through straightforward prose. It may reduce unnecessary backtracking, though going back is sometimes exactly what comprehension requires. I wouldn't try to eliminate it. The aim is to adjust the pace to the material: skim when building a map, move quickly through familiar territory, slow down around difficult arguments, and stop when something needs to be worked out.

Applying all of this

If I were to reduce all of this to a study routine, it would look something like this:

1. Survey the material so I know what I'm trying to understand.
2. Read or work through a small enough section to keep its parts in mind.
3. Close the source and reconstruct the explanation or solve a problem unaided.
4. Check the result and identify what I missed.
5. Return to it later, with enough time between attempts that I have to retrieve it again.
6. Mix in other problems so I learn when an idea applies, not just how to repeat it.
7. Get enough sleep and leave room for the learning to settle.

What interests me about all this is that learning stops looking like a question of how much information I can take in. The more useful question is what I can build, retain, and call upon when I need it.

That's where mechanical sympathy comes in. The brain has constraints, and it changes in response to what we repeatedly ask it to do. Understanding a little of how that happens changes how I want to spend my time with a book, a problem, or a new skill.