Exam Priority Notice
BMI content (Section 7) was covered exclusively in lecture and is NOT in the textbook. The instructor flagged BMI questions as high-priority on the exam. Pay special attention to Fetz (1969), cosine tuning, population coding, BMI types, neural drift, and ethics.
Section 1
What Is Skill Memory?
A skill is an ability that improves over time through practice. Skill memories are procedural memories — implicit memory for how to do something. They differ from declarative memories in three key ways:
| Dimension | Skill Memory (Procedural) | Declarative Memory |
|---|---|---|
| Communication | Hard to verbalize — you demonstrate rather than explain | Flexible verbal communication in many formats |
| Awareness | Often acquired and retrieved without conscious awareness (implicit) | Requires conscious access |
| Acquisition Speed | Requires repeated practice and feedback | Can form from a single exposure (episodic) |
| Brain Systems | Basal ganglia, cerebellum, motor cortex | Hippocampus and medial temporal lobe |
The Bicycle Paradox
You can RIDE a bicycle without being able to SAY what muscles keep you balanced. You DO it without being able to SAY it. This is the core paradox of skill memory and captures the implicit/explicit dissociation.
Section 2
Types of Skills
| Dimension | Type A | Type B |
|---|---|---|
| Content | Perceptual-Motor — dancing, tennis, driving | Cognitive — chess, poker, mental arithmetic |
| Environment | Closed — predefined sequences, stable environment (choreographed ballet, dart at fixed target) | Open — adapted in real time to changing environment (basketball defense, catching a frisbee) |
Lecture Examples — Know These
- Piano etude = closed motor. Jazz improvisation = open motor.
- Routine surgery = closed motor (for an expert). Poker = open cognitive.
- Chess grandmaster = approaches closed perceptual-motor — pattern recognition triggers near-automatic responses.
Key Insight: Cognitive Skills Become Motor Programs
With enough practice, even cognitive skills become motor programs. A grandmaster "sees" the right move before consciously "thinking" it. This is the transition to the autonomous stage (Fitts). The boundary between cognitive and motor dissolves at expertise.
Section 3
The Science of Practice
3a. Feedback (Knowledge of Results)
Thorndike's blindfolded line-drawing experiment: Both groups drew equal numbers of lines, but only the group receiving feedback about accuracy improved. Repetition alone is NOT enough — feedback is essential.
Maps onto reinforcement learning: without a prediction error signal (TD error), the policy does not update.
3b. Power Law of Learning
Universal finding: early practice yields large improvements; additional practice yields progressively smaller gains. Rate decreases as a power function. Applies across species and across both motor and cognitive skills.
Example: typing improves from 10 → 40 WPM in the first month, but only 40 → 42 WPM after a year of additional daily practice.
3c. Massed vs. Spaced Practice (Baddeley & Longman, 1978 — Post Office Keyboard Study)
Workers practicing 1 hour/day (60 days) needed fewer total hours to reach proficiency than 4-hour/day workers (20 days). Spaced practice is more efficient per hour, likely due to memory consolidation during sleep.
Counterintuitive Finding
The most efficient (spaced) group was least satisfied with their training — progress felt imperceptibly slow. Subjective satisfaction is a poor measure of actual learning. This matters for designing training programs.
3d. Constant vs. Variable Practice
Constant: same setting — strong performance in that specific context. Variable: varied contexts — better transfer and generalization. The relative advantage is context-dependent and an active research area.
Section 4
Implicit Learning & the SRT Task
Implicit learning = learning without awareness of what is being learned.
H.M. (amnesic patient): improved at mirror-tracing session by session, yet had NO MEMORY of ever having done the task. This classic dissociation separates declarative memory (hippocampus, damaged) from skill memory (basal ganglia + cerebellum, intact).
Serial Reaction Time (SRT) Task Logic
Sequential block (repeating 12-item pattern) → faster RTs than random blockBUT: participants CANNOT verbally report the repeating pattern
= Learning without awareness = Implicit learning
Why SRT Is "Golden" — Know This for the Exam
The SRT task's simplicity allows it to be used across humans, monkeys, and rodents. This cross-species compatibility enables combining behavioral results with invasive recordings and genetic tools in non-human animals — giving researchers mechanistic access unavailable in purely human studies.
Section 5
Fitts's Three-Stage Model
| Stage | Characteristics | Example |
|---|---|---|
| 1 — Cognitive | Verbalizable rules, active thinking, frequent errors, declarative memory dominant | Learning to drive: consciously remembering each gear change, mirror check, speed limit |
| 2 — Associative | Actions becoming stereotyped, less reliance on explicit rules, errors decrease | Driving without thinking about every gear change |
| 3 — Autonomous | Movements become motor programs, can multitask, cannot easily verbalize | Typing while talking; surgeon doing routine procedure while listening to music |
RL Connection: Model-Based → Model-Free
Cognitive stage = model-based control (deliberate planning, prefrontal + hippocampus). Autonomous stage = model-free control (cached S-R, automatic, basal ganglia direct pathway). Associative stage = the transition. Maps to Week 11 (Habits vs. Goals) and the actor-critic architecture.
Choking Under Pressure
An autonomous-stage skill can revert to cognitive control when the performer focuses consciously on their technique (e.g., a golfer suddenly thinking about each step of their swing). Conscious attention interferes with the automatic motor program. The stages can flow in both directions.
Section 6
Brain Substrates of Skill Memory
| Region | Role | Key Evidence |
|---|---|---|
| Basal Ganglia (dorsal striatum) | Links sensory cues to motor responses (S-R); controls movement parameters (velocity, direction, amplitude); chunking | Radial arm maze double dissociation: BG lesions impair cue-guided (lit arm) task but NOT memory-based task. T-maze recordings shift from decision-point → start/END firing with training = chunking |
| Cerebral Cortex (motor + somatosensory) | Controls complex action sequences; cortical maps expand with practice (structural plasticity) | Violinists' enlarged somatosensory representation for left hand. Juggler study: 3% visual cortex expansion after 3 months. London taxi drivers' hippocampal expansion |
| Cerebellum | Timing of movement sequences; especially critical for precisely timed skills (dance, music) | Cerebellar patients take twice as long on mirror-tracing and transfer less. Early cerebellar activity spike when humans begin learning sequential finger movements |
Two Types of Plasticity
- Structural plasticity: cortical area literally expands — more synaptic connections are recruited (not new neurons).
- Functional plasticity: increased blood flow / fMRI signal in skill-related areas during performance.
Where Is Skill Memory Stored?
Current predominant view: in synaptic connections between cortex and striatum/basal ganglia — not inside neurons themselves. Consistent with LTP/LTD mechanisms from earlier weeks.
Section 7
Brain-Machine Interfaces (BMI) — Lecture Only!
Exam Note: Lecture-Only Content
This entire section comes from lecture — it is NOT in the textbook. The instructor identified BMI as a high-priority exam topic. Know Fetz (1969), cosine tuning, population coding, all four BMI types, neural drift, BMI as skill learning, and the ethics issues.
7a. Fetz (1969) — Origin of BMI
Working alone with one monkey at U Washington, Eberhard Fetz rewarded the animal with food whenever a recorded cortical neuron elevated its firing rate. The monkey learned to volitionally control that single cell's activity via operant conditioning — the first demonstration of intentional modulation of individual neuron activity.
RL connection: auditory click = conditioned reinforcer; food pellet = primary reward; neural firing rate = the conditioned behavior. Extension: connect that neuron's output to an external device → animal controls device with thought alone.
7b. Georgopoulos & Population Coding
Center-out reach task: monkeys reached to 8 targets while motor cortex neurons were recorded. Each neuron is cosine-tuned — it has a preferred direction, and its firing rate follows the cosine of the angle between the actual movement and the preferred direction.
Population coding: summing activity across many neurons with different preferred directions gives a precise vector estimate of intended movement direction. Recording from more neurons yields more accurate decoding.
7c. Types of BMI
| Type | Direction | Example |
|---|---|---|
| Read-only | Records neural signals → translates to machine command | Motor prosthetics (Schwartz monkey/robotic arm; BrainGate clinical trials) |
| Write-only (stimulation) | Records external signals → writes electrical patterns to nervous system | Cochlear implant — most successful clinical BMI; records sound, stimulates auditory nerve |
| Closed-loop (read + write) | Reads from one site, stimulates another | DBS for Parkinson's; future sensorimotor prosthetics |
| Computational bypass | Reads from one region, stimulates downstream to bypass damaged pathway | Berger's MIMO hippocampal memory chip (CA3 → CA1) |
7d. Key Engineering Challenges
- Invasiveness vs. signal quality: EEG (scalp) = non-invasive, low resolution. Intracortical arrays (Utah array, Neuralink) = invasive, single-neuron resolution.
- Neural drift: neurons change preferred directions over days/weeks, requiring daily recalibration. Early patients found this burdensome; one dropped out.
- Recording too few neurons: need more simultaneously recorded cells for richer, stable decoding.
- Longevity: chronic implants must survive years in a dynamic fluid environment without causing inflammation.
- Neural plasticity (opportunity): Ganguly's lab (UCSF) — when one neuron is linked to BMI output, surrounding neurons remap to support it. The brain treats BMI as a new effector.
7e. BMI as Skill Learning
Early patients described training as "like being a football player — 6–8 hours/day." BMI performance follows the power law (early rapid gains, then plateau). Best strategy: early phase — machine adapts to brain (supervised adaptive decoder); later — freeze machine, let brain adapt to it. Long-term goal: stable representation requiring only annual (vs. daily) recalibration.
7f. Ethics
- Agency / legal responsibility: if a BMI-controlled arm injures someone, who is legally responsible?
- Identity: does DBS-induced personality change mean the person is now "different"?
- Data ownership: who owns decoded neural data flowing to a corporate server?
- Access / justice: if cognitive enhancements are available, who gets them and at what cost?
- Augmentation vs. restoration: therapeutic BMIs vs. enhancement raises new ethical frameworks.
Section 8
Clinical Perspectives
| Condition | Mechanism | Key Finding / Relevance |
|---|---|---|
| Parkinson's Disease | Loss of dopaminergic neurons in SNc → reduced dopamine in basal ganglia → disrupted BG-cortical loop. Motor rigidity, tremors, impaired skill initiation. | Persian sitar maestro (~93, Parkinson's): needs help walking onto stage but plays with remarkable skill. Deeply ingrained motor programs robustly stored even when general initiation is severely impaired. DBS partially restores function. |
| Huntington's Disease | Gradual neuronal loss in basal ganglia AND cortex. | Motor symptoms + severe deficits in procedural skill learning: weather prediction task, SRT, rotary pursuit. Abnormal LTP/LTD in mice. |
| Apraxia | Damage to parietal cortex (especially left hemisphere). Can perform individual steps but cannot sequence them. | Disrupts access to stored motor action memories. Difficulty coordinating purposeful skilled movements despite intact individual motor components. |
| Twin Study (rotary pursuit) | Genetic influence on skill learning amplified by extended practice. | With extended training, identical twins became more similar; fraternal twins became more dissimilar. Practice amplifies genetic influences, moving performance toward the genetic ceiling or floor. |
Key Terms — 18 Flashcards
Click any card to reveal its definition. Use the filters to focus on a category.
Behavioral
Skill Memory / Procedural Memory
Click to reveal →
Behavioral
Implicit memory for how to do something; hard to verbalize; improves with practice; relies on BG, cerebellum, and motor cortex rather than hippocampus.
Behavioral
Closed Skill
Click to reveal →
Behavioral
A skill involving a predefined sequence executed in a stable, predictable environment. Example: choreographed ballet, dart throw at fixed target.
Behavioral
Open Skill
Click to reveal →
Behavioral
A skill that must be adapted in real time to an unpredictable, changing environment. Example: basketball defense, catching a frisbee.
Behavioral
Power Law of Learning
Click to reveal →
Behavioral
Early practice yields large improvements; additional practice yields progressively smaller gains. Rate of improvement decreases as a power function. Universal across species and skill types.
Behavioral
Massed vs. Spaced Practice
Click to reveal →
Behavioral
Massed: concentrated sessions — faster but less durable. Spaced: distributed sessions — more efficient per total hour, likely due to sleep consolidation. Baddeley & Longman post office study.
Behavioral
Implicit Learning
Click to reveal →
Behavioral
Learning that occurs without awareness of what is being learned. Preserved in amnesia (H.M.). Demonstrated by SRT task: faster RTs for sequential blocks despite inability to verbalize the pattern.
Behavioral
Serial Reaction Time (SRT) Task
Click to reveal →
Behavioral
Participants press keys responding to cues. Sequential blocks (repeating 12-item pattern) yield faster RTs than random blocks, yet participants cannot verbally report the pattern. Gold standard measure of implicit sequence learning.
Behavioral
Fitts's Three Stages
Click to reveal →
Behavioral
Cognitive (rules, errors, effortful) → Associative (stereotyped, fewer errors) → Autonomous (motor program, multitask, can't verbalize). Parallels model-based → model-free control. Stages flow in both directions (choking).
Neural
Basal Ganglia / Dorsal Striatum
Click to reveal →
Neural
Links sensory cues to motor responses (S-R learning). Evidence: radial arm maze double dissociation. T-maze: firing shifts from decision-point → start/END with training = chunking — whole action sequence becomes a motor program.
Neural
Cerebellum
Click to reveal →
Neural
Critical for timing of movement sequences. Especially important for precisely timed skills (dance, music). Cerebellar patients take twice as long at mirror-tracing and transfer less.
Neural
Structural Plasticity
Click to reveal →
Neural
The cortical area dedicated to a trained skill literally expands — more synaptic connections are recruited (not new neurons). Violinists' enlarged left-hand somatosensory area; juggler visual cortex expansion.
Neural
Functional Plasticity
Click to reveal →
Neural
Increased blood flow / fMRI signal in skill-related areas during performance. Distinct from structural plasticity (which involves physical synaptic growth). Both types occur with practice.
Neural
Cosine Tuning
Click to reveal →
Neural
Motor cortex neurons fire maximally for their preferred direction; firing rate varies as the cosine of the angle between actual movement and the preferred direction. Georgopoulos center-out reach task. Foundation of population decoding.
Neural
Population Coding
Click to reveal →
Neural
Decoding movement direction (or other variables) from the summed activity of many neurons with different preferred directions. More neurons recorded = more accurate decoding. The basis of BMI motor prosthetics.
Clinical/BMI
BMI / Brain-Machine Interface
Click to reveal →
Clinical/BMI
A system that reads from the brain, writes to the brain, or does both. Types: read-only (motor prosthetics), write-only (cochlear implant), closed-loop (DBS), computational bypass (hippocampal chip). Restores or augments function.
Clinical/BMI
Cochlear Implant
Click to reveal →
Clinical/BMI
Write-only BMI — the most clinically successful. Records sound via microphone, converts to electrical signals, stimulates the auditory nerve. Does NOT read from the brain; it writes to it, bypassing damaged hair cells.
Clinical/BMI
Neural Drift
Click to reveal →
Clinical/BMI
Neurons' preferred response properties change over days/weeks, making the BMI decoder (trained on one day's patterns) inaccurate the next day. Required daily recalibration in early clinical trials — a major engineering and clinical challenge.
Clinical/BMI
Deep Brain Stimulation (DBS)
Click to reveal →
Clinical/BMI
Disrupts pathological oscillations in basal ganglia circuits. Treats Parkinson's disease by partially restoring the BG-cortical loop. A closed-loop BMI: reads pathological activity and stimulates to correct it. Raises identity ethics questions.
Practice Multiple Choice — 20 Questions
Click an option to check your answer. Explanations appear automatically.
30
Questions
0
Correct
0
Incorrect
—
Score
Big Picture Synthesis
How the week's concepts connect across levels of analysis and to the course arc.
The Unifying Principle
Skill learning is the brain gradually transferring control from deliberate (model-based, cognitive) to automatic (model-free, motor program) processing. The same basal ganglia system that computes prediction errors (Weeks 5–8) now builds cached S-R action sequences. BMI is the ultimate extension: the brain applies its skill-learning machinery to operate entirely novel external tools.
Levels of Analysis
Behavior
Power law: large early gains, plateau
3 stages: cognitive → associative → autonomous
Implicit vs. explicit learning (SRT task)
Circuit
Dorsal striatum: S-R, chunking
Cerebellum: timing
Motor cortex: cosine tuning, population coding
Synapse
LTP/LTD at corticostriatal synapses
Skill memory in connections, not neurons
Fetz: operant conditioning of single-neuron firing
Structure
Cortical map expansion (violinists, jugglers)
Motor cortex population vector decoding
Neural drift over time → recalibration
Likely Exam Themes
- Fitts's three stages and their RL analogs (model-based → model-free)
- BG vs. hippocampus dissociation — cue-guided vs. memory task (radial arm maze)
- SRT task as proof of implicit learning: faster RTs, no verbal report
- Fetz (1969) — operant conditioning of single cortical neuron = origin of BMI
- Cosine tuning and population coding — Georgopoulos center-out task
- Cochlear implant = write-only BMI — most successful clinical BMI
- Neural drift → daily recalibration problem in early BMI patients
- Ethics: agency, identity, data ownership
- Twin study — practice amplifies genetic influences
Cross-Course Connections
- Autonomous stage = model-free policy (BG direct pathway from Week 8)
- Cognitive stage = model-based planning (prefrontal + hippocampus, earlier weeks)
- Fetz experiment = operant conditioning (Week 8) applied at single-neuron level
- BMI power law matches spacing effect and power law from Week 6
- Thorndike feedback = knowledge of results = TD error signal needed for policy update
- Chunking in BG = caching whole action sequence as one S-R unit = actor in actor-critic
- Neural drift — same plasticity that enables learning can undermine stability of existing representations