The campaign
The Theory wing
Twelve levels in four chapters that turn a computational-neuroscience course into puzzles: the phase plane, rate models and mean field, attractor memory, tuning and decoding.
The Theory wing is the part of the campaign that asks "what does the mathematics of a neuron say?" instead of "which body does it drive?". It has four chapters of three levels each: Dynamics, Rate models, Attractor memory and Tuning and decoding. Each chapter starts with two small puzzles (one idea each) and ends with a larger scored task, the same shape as every other chapter.
The wing is open from the start, like the Bestiary. It is not part of the unlock chain of the five body chapters, and nothing in it gates anything else. Inside a chapter the usual rule holds: a level opens when the one before it is passed. You can start with Theory, with the body chapters, or jump between them. See the core loop for how a level is built, run and checked, and stars, par and the hint ladder for scoring.
Every level here has the usual honesty contract. The game checks the mechanism you built by reading the circuit's structure (which cell connects to which, with what sign), not only by looking at the numbers it produces. A shortcut that happens to hit the target numbers fails with a verdict that names the missing part. Verdicts and the first two hint tiers carry no numbers. The reference solutions are never shown on this page, and the "Stuck?" notes below go no further than the first hint tier.
What the wing adds to the labLink to this section
The Theory levels use the same lab as the rest of the game. A few new pieces appear in it.
| New piece | Where you meet it | What it shows |
|---|---|---|
| Phase plane | World panel of the Dynamics levels | A picture of a two-variable cell: voltage on one axis, a "recovery variable" on the other, the two nullclines, the fixed points and the live path of the cell. |
| Rate units and rate strips | World panel of the Rate models levels | Scrolling traces of the readouts of rate units on one shared scale, one trace per output channel. |
| Siegert units and diffusion projections | Parts bin and the projection inspector in r_meanfield | A unit that stands for a whole population, driven by the mean and spread of its input; a connection that carries the weight J and the number of inputs K of the network it stands for. |
| Pattern grid | World panel of the Attractor memory levels | The cells the level reads, laid out on a grid and lit by their own spikes. Cells that a sensor reaches are ringed and named by the sensor. |
| Tuning dial, bars and curve | World panel of the Tuning levels | A dial for the stimulus shown, a bar per cell, a single cell's tuning curve, and a decoded-direction arrow. |
| Population-vector readout | Output kind of t_decode | A motor channel that turns a ring of cells into one angle. |
| STA tab | Analysis screen | Spike-triggered average: what the input looked like just before the cell fired. Reachable from the More menu once the campaign has reached the chapters that open analysis. |
Plain-language explanations of each follow in the chapters. The parts and wiring options are in the parts bin, synapses and modulators and inputs and outputs.
Honest caveatsLink to this section
The wing is built on real simulator models, but several things are simplified for the sake of a puzzle you can finish in minutes. The full account is in what is real. The points that matter here:
- The "Wilson-Cowan" units are not the original equations. They are NEST's threshold-linear rate neurons with a ceiling, not the sigmoid-with-refractoriness of the 1972 paper. Rates are in arbitrary units, and the drive is added after the nonlinearity, so inhibition cannot cancel it.
- The mean-field network is small and well behaved. The level shows the method, not where it breaks down.
- Tuning curves are given, not learned. The afferents have a fixed bell-shaped rate curve, Poisson spikes and evenly spread preferred angles.
- Memory patterns are groups of cells, not the plus-one and minus-one vectors of the 1982 model. Persistent firing is far faster than real delay activity.
- The "one neuron switch" fires much faster than a real persistent-firing cell. The level says so.
- Weights, time constants, windows and pass bands are tuned for the game. They are properties of the game, not measurements.
Levels at a glanceLink to this section
None of the twelve levels has challenge variants. Par is the neuron count of the reference circuit; staying at or under it, with five seeds passing, within the energy budget (the default of 40 spikes per second per neuron for these levels) and without opening the recipe, earns the third star. See stars, par and the hint ladder.
| Chapter | Level | Kind | Par |
|---|---|---|---|
| Dynamics | Where firing starts (d_class) | micro-puzzle | 1 |
| Dynamics | Only the change counts (d_phasic) | micro-puzzle | 1 |
| Dynamics | A neuron that remembers a pulse (d_switch) | task | 1 |
| Rate models | Two populations that make a rhythm (r_oscillate) | micro-puzzle | 2 |
| Rate models | Two populations, one winner (r_choose) | micro-puzzle | 2 |
| Rate models | The network and its mean field (r_meanfield) | task | 255 |
| Attractor memory | Recalling a whole from a part (h_complete) | micro-puzzle | 8 |
| Attractor memory | Two memories, no blends (h_two) | micro-puzzle | 8 |
| Attractor memory | Writing a memory with plasticity (h_learn) | task | 8 |
| Tuning and decoding | A cell tuned by its inputs (t_tuned) | micro-puzzle | 9 |
| Tuning and decoding | Sharpening a tuning curve (t_sharpen) | micro-puzzle | 16 |
| Tuning and decoding | Reading a direction from a crowd (t_decode) | task | 24 |
Chapter 1: DynamicsLink to this section
This chapter is about one neuron seen as a moving point. A cell like the ones in this chapter has two variables: its voltage and a slower recovery variable (a stand-in for the slow currents that pull the voltage back). A picture with voltage on one axis and the recovery variable on the other is a phase plane. Every moment of the cell is a point on it, and the cell's life is a path.
Two curves on the plane matter. A nullcline is the set of points where one of the two variables is not changing at that instant. The voltage nullcline and the recovery nullcline cross at a fixed point: a state where nothing changes. If the cell is left alone at a stable fixed point, it stays there. That is the resting state. A steady input current moves the voltage nullcline, and the fixed point moves with it, or disappears.
In the lab the phase plane is drawn for the first cell the level's motor channel reads, using the cell's own parameters and the constant current it is getting right now. Stable fixed points are drawn filled, unstable ones hollow, saddles as crosses. The reset (where a spike sends the cell) is marked, and the live path of the cell runs over the picture. It is computed for a constant current, so it ignores synaptic input.
Note: The phase plane works for the two-variable cells: the Izhikevich cell and the adaptive exponential cells. A level may still be scored with another model, but then it has no picture and the mechanism check will say the cell needs two variables. The default cell of the Dynamics levels is the Izhikevich cell. See the parts bin.
| Level | What it teaches | Know first | The real science | Citation |
|---|---|---|---|---|
| Where firing starts | Excitability class: how firing begins as current grows | Threshold and rheobase | Hodgkin found two kinds of axon. Class 1 axons fire at any low rate just above threshold; class 2 axons start at a clear rate. In a two-variable model the difference is how the resting state is lost. | Hodgkin 1948, J Physiol 107:165, doi:10.1113/jphysiol.1948.sp004260; Izhikevich 2003, IEEE Trans Neural Netw 14:1569, doi:10.1109/TNN.2003.820440 |
| Only the change counts | A stable fixed point that survives the input, so firing stops | Excitability class | Some auditory brainstem neurons with low-threshold potassium currents fire at the onset of a step and are then silent: they report change, not level. | Izhikevich 2003; Rothman and Manis 2003, J Neurophysiol 89:3097, doi:10.1152/jn.00127.2002 |
| A neuron that remembers a pulse | Bistability: rest and firing both stable at once | Excitability class, stable fixed point | A model cell can have a stable rest and a stable firing state side by side; a brief pulse that crosses the border flips it. Real persistent firing is a different, channel-based mechanism. | Izhikevich 2003; Touboul and Brette 2008, Biol Cybern 99:319, doi:10.1007/s00422-008-0267-4 |
Where firing starts (d_class)Link to this section
The idea. Raise the current into a neuron step by step and watch where firing begins. A class I cell starts slowly, from almost no spikes, and speeds up smoothly as the current grows. A class II cell starts at once, at a clear rate, as if a switch had been thrown.
The phase plane says why. As the current grows, the voltage nullcline rises and the resting state, the crossing of the nullclines, is lost. If the resting state runs into a second crossing and the two annihilate, the cell can only creep through the bottleneck that is left: class I. If the resting state turns into a spiral (a state the cell circles into, like water down a drain) and the current makes that spiral unstable, the cell starts firing at the frequency of that spiral: class II. In the maths this is the difference between a saddle-node and a Hopf onset. The words "saddle-node" are not needed to play; "the resting state vanishes by collision" and "the resting state spirals out" are enough.
The goal. The sensor current steps up in 21 stairs of 200 ms through the cell's own onset current. The staircase is scaled to the cell, so you cannot make a slow starter look abrupt by choosing a big weight. Build one cell, read on spikes, that is silent at the bottom and whose firing starts at a clear rate, not from zero.
What the pass rule asks. The three lowest steps must be silent; some step must fire; the busiest of the first steps that spike must show a clear burst (three spikes in the last 140 ms of a step in Standard, two in Guided, four in Researcher); and the top step must still fire. In addition the structure check computes the class for your particular cell from the slopes of its two equations at rest. The class is not looked up from the cell's name, so a cell with other parameters can land on either side.
What is simplified. The two classes shade into each other in these models. The default regular-spiking cell sits near the point where they meet, so the level uses a game criterion: class II is decided by the resonance of the resting state just before it is lost (a damped oscillation of about 15 Hz or more). The game says this rather than pretend there is a sharp line.
Par: 1 neuron.
Stuck?Link to this section
How firing starts depends on how the resting state is lost: through a collision of two fixed points (a slow start from zero) or as a spiral that turns unstable (an abrupt start).
Only the change counts (d_phasic)Link to this section
The idea. Some neurons answer a step of current with one spike or a short burst and then fall silent, although the current stays on. They report the change, not the level. A step of current only moves the voltage nullcline. A tonic cell has no stable crossing left afterwards, so it circles for as long as the current is on. A phasic cell still has a stable crossing: the step throws the voltage across threshold once, then the path spirals into the new resting point and stays there.
The goal. The sensor current is zero, steps on for 1.2 seconds (400 to 1600 ms), and goes off again. Build a cell that is silent before the step, fires when the step comes, and is silent again well before the step ends. The size of the step is the weight you give the sensor, and it moves the nullcline as much as any parameter of the cell does.
What the pass rule asks. No spike in the quiet window before the step; at least one spike in the 250 ms after the step starts; no spike in the late window (900 to 1600 ms) while the current is still on. The structure check asks for a two-variable cell that has a stable fixed point at the held current, so that it can settle.
What is simplified. The level uses a generic two-variable cell, not an auditory neuron. The real mechanism in the brainstem involves a particular potassium current; the cell here shows only the geometry.
Par: 1 neuron.
Stuck?Link to this section
A step of current only moves the voltage nullcline. Whether the cell keeps firing depends on whether the nullclines still cross at a stable point afterwards.
A neuron that remembers a pulse (d_switch)Link to this section
The idea. A neuron can have two stable states at once: quiet at rest, and firing on and on. In the phase plane the resting point is a stable crossing, and the firing is a loop around it that the path can also settle into. Between the two runs a border. A brief pulse that carries the cell across the border flips the state for good. This is bistability: one cell, one current, two stable states, so the cell stores one bit.
The goal. Three sensors reach the cell. hold is a steady current that puts the cell near the edge of firing without starting it. set is a short pulse that pushes the cell up, and reset is a short pulse that pushes it down. Both pulses take a positive weight; the sensor itself carries the sign. Build a single cell that is quiet at first, fires on and on after set (long after the pulse is gone), and is quiet again and stays so after reset. The same sequence must work a second time. What makes firing survive after a spike is where the spike resets the cell relative to the border between the states.
What the pass rule asks. Quiet before the first pulse; at least six spikes in each 400 ms "on" window (four in Guided, eight in Researcher), long after the pulse; no spike in each "off" window; and the on state must stay at or below 80 spikes in 400 ms (200 Hz; 100 in Guided, 70 in Researcher). The structure check asks for a two-variable cell with a stable fixed point at the holding current and all three sensors reaching it. The behaviour decides the rest: only a real second, firing state makes the "on" windows pass.
What is simplified. Real persistent firing of single cells runs at 5 to 50 spikes per second and relies on channel-based plateau potentials, which are not modelled. The firing state of this kind of cell is a limit cycle born at a finite rate, about 135 spikes per second at the very lowest, so the 200 Hz cap is as low as the model allows. The level shows the mechanism, not the rate.
Par: 1 neuron.
Stuck?Link to this section
The same current can leave a neuron with two stable states: rest and repetitive firing. A pulse that crosses the border between them flips the state, and it stays flipped.
Chapter 2: Rate modelsLink to this section
Here a whole population of neurons is treated as one number: its firing rate. This is a drastic simplification, and an old one. It trades spikes for something you can analyse with a pencil.
A rate unit has a time constant (how fast it follows its input), a gain, a threshold and a ceiling. A connection between rate units is a weight; a negative weight is inhibition. The units in the game are NEST's threshold-linear rate neurons. Their rate follows X' = (-X + drive + clip(sum of weight x rate)) / tau: the drive is added after the nonlinearity. That has a consequence you will meet: a rate can read a little above the ceiling, and inhibition cannot cancel the drive.
In the lab, rate readouts appear as scrolling traces on one shared scale. A readout of a rate unit is in arbitrary units; a readout of a spiking pool is in spikes per second. See inputs and outputs for the rate_drive input and rate_value output kinds and synapses and modulators for rate connections.
| Level | What it teaches | Know first | The real science | Citation |
|---|---|---|---|---|
| Two populations that make a rhythm | The excitatory-inhibitory pair that oscillates | Negative feedback | Wilson and Cowan wrote equations for coupled excitatory and inhibitory populations. A pair can sit steady, jump between rates, or oscillate, and the oscillation speeds up with stronger stimulus. | Wilson and Cowan 1972, Biophys J 12:1, doi:10.1016/S0006-3495(72)86068-5 |
| Two populations, one winner | Competition between populations that holds its winner | Command switch, the Wilson-Cowan pair | Populations that excite themselves and inhibit each other turn a small input difference into a large output difference. The same wiring has been proposed for decisions, attention and working memory. | Wilson and Cowan 1972 |
| The network and its mean field | Predicting a spiking network's rate from the mean and spread of its input | The Wilson-Cowan pair, inhibition | A population of integrate-and-fire cells with many weak inputs has a rate fixed by the mean and variance of its input. Solving the self-consistent equations predicts the network. | Amit and Brunel 1997, Cereb Cortex 7:237, doi:10.1093/cercor/7.3.237; Hahne et al. 2017, Front Neuroinform 11:34, doi:10.3389/fninf.2017.00034 |
Two populations that make a rhythm (r_oscillate)Link to this section
The idea. The Wilson-Cowan pair is an excitatory population that excites itself and an inhibitory population that it excites and that inhibits it back. Excitation builds and drives the inhibition; the inhibition shuts the excitation down; then the inhibition fades and excitation builds again. The result is an oscillation made of two numbers.
The goal. The sensor drive is zero and then switches on. Build two rate units: e reads the excitatory one and i the inhibitory one. The excitatory rate must stay quiet before the drive, then swing up and down again and again, and the inhibitory rate must follow it a little later in each cycle.
Two things have to be right. The excitation must be able to run away by itself (its self-connection must be strong enough compared with how fast the unit follows its input). And the loop through the inhibitory unit must be strong enough to catch it but not so strong that it holds the excitation down for good. The verdict tells you which side of that narrow band you are on: a rate that sticks near its ceiling means too little inhibition; one pinned just above the drive level means too much; a steady value in between means the excitation does not run away strongly enough.
What the pass rule asks. e is quiet before the drive; after it, e swings by a clear amount, for at least four cycles (three in Guided) in the last stretch of the run; one cycle lasts between 40 and 400 ms; and i follows e by between 5 and 70 percent of a cycle, so the inhibition is the cause of the turn-around and not a copy. The structure check wants e and i on different units, the drive reaching e, e exciting itself and i, and i inhibiting e.
What is simplified. The units are threshold-linear, not the original sigmoid, so the band in which oscillation exists is narrow. The inhibition does not have to be slower than the excitation; a slower inhibition only makes oscillation start more easily and lowers its frequency.
Par: 2 units.
Stuck?Link to this section
Excitation that excites itself, plus inhibition that it drives and that inhibits it back, makes a rhythm: it overshoots, is shut down, recovers and overshoots again.
Two populations, one winner (r_choose)Link to this section
The idea. Two populations compete for the same job, say two possible answers to an ambiguous stimulus. Each is driven by its own evidence (a_in, b_in), each excites itself and each inhibits the other. A small advantage in the drive should become a large advantage in the output: one wins and the other is silenced. The winner then holds itself up, so the choice outlasts the evidence.
The goal. Build two rate units, a and b. In the run, A gets a little more drive than B, then the two are equal, then B gets much more, then the two are equal again. A small difference must be amplified; the choice must be held when the inputs are equal; a much stronger input for the other side must overturn it; and the new winner must be held in turn. The memory is in the loop, so it works whichever side won last. A circuit that simply gives one input more gain than the other holds only its favourite and fails the last phase.
What the pass rule asks. At the end of each phase the leading unit must beat the other by a clear margin: A wins phase one, A still leads in phase two with equal inputs, B leads in phase three, and B still leads in phase four. The structure check asks for self-excitation on each unit and inhibition from each onto the other, directly or through an interneuron.
What is simplified. Rates are in arbitrary units and the units are threshold-linear. The "winner-take-all" reading is the textbook one for decision and attention models.
Par: 2 units.
Stuck?Link to this section
Self-excitation plus mutual inhibition between two populations turns a small difference in drive into a large difference in rate, and holds it.
The network and its mean field (r_meanfield)Link to this section
The idea. A large, sparse network of excitatory and inhibitory integrate-and-fire cells, driven by a lot of independent Poisson input, fires asynchronously (the cells are not in step). Every cell sees the sum of many small inputs, so the total input is nearly Gaussian. Its mean grows with J K r and its variance with J^2 K r, where J is the size of one input in millivolts, K is how many inputs each cell has, and r is the rate of the cells that send them. The rate of an integrate-and-fire cell driven by such an input is given by the Siegert formula. A rate unit that applies it, a siegert unit, can stand for a whole population. The network's rates then follow from a few self-consistent equations: no spikes needed. That is mean-field theory.
The goal. Build both: the spiking network (excitatory and inhibitory cells, random sparse connections, the Poisson drive ext) and its prediction (a siegert unit for each population and one for the drive, with projections that carry the same J and K, driven by ext_mf at the same rate). net reads the mean rate of the excitatory cells and mf the prediction. At three drive levels the two must agree. Where they do not, something differs between the two descriptions: J, K, the sign of the inhibition, or the drive.
In the projection inspector, a "mean-field projection" takes J and K as two separate numbers. NEST itself only has the derived drift and diffusion factors; the game computes them for you.
What the pass rule asks. At the highest drive both the network and the prediction fire (at least 5 Hz); the network's rate rises at each drive level; each prediction projection carries a J and a K within a factor of two of the spiking projection it stands for (a K of 1 for a 60-input connection is rejected, even if a hand-tuned J makes the rates agree); and at all three levels the two rates differ by at most 15 percent of the larger plus 1.5 Hz (25 percent in Guided, 10 percent in Researcher).
What is simplified. The network is small and not very sparse (a few hundred cells, not thousands), yet the prediction still agrees within a few percent. So the level shows the method but not where the mean field fails. The recurrent weights are chosen so that the loop matters; drive levels and measurement windows are game values. Par is 255 because the reference counts all of the spiking cells and the units of the prediction.
Par: 255 cells and units.
Stuck?Link to this section
Many weak, independent inputs make the input to a cell Gaussian: its mean and variance, summed over the inputs, are all the cell's rate depends on. A siegert unit computes that rate.
Chapter 3: Attractor memoryLink to this section
An attractor is a state that a system falls into and stays in. Hopfield's idea (1982) is that a memory can be an attractor of a network. If cells excite each other symmetrically, meaning the weight from A to B equals the weight from B to A, the network has stable states, and a pattern stored in the weights is one of them. Show the network part of the pattern, and it falls into the whole. This is a content-addressable memory: you recall by content, not by address.
In these levels a pattern is a group of cells, not a vector of plus and minus ones. The lab draws the cells that the level reads on a grid, each lit by its own spikes. Cells that a sensor reaches are ringed in amber and named by the sensor: the cue is the part that is touched, and the pattern is the whole.
Symmetry is checked on the actual list of connections between cells, one by one. A directed random graph is not called symmetric. An outer-product wiring inside one population is accepted too.
| Level | What it teaches | Know first | The real science | Citation |
|---|---|---|---|---|
| Recalling a whole from a part | An attractor completes a partial cue and holds it | Persistence, excitation | Hopfield showed that symmetric networks have stable states that can be made to be stored patterns. In spiking networks the same states are persistent, selective activity, a model of delay activity in cortex. | Hopfield 1982, Proc Natl Acad Sci USA 79:2554, doi:10.1073/pnas.79.8.2554; Amit and Brunel 1997, doi:10.1093/cercor/7.3.237 |
| Two memories, no blends | Inhibition between patterns avoids spurious mixtures | Attractor memory, command switch | A network storing several patterns can fall into mixtures nobody stored. Competition between the groups makes the pure patterns win. | Hopfield 1982 |
| Writing a memory with plasticity | Spike-timing-dependent plasticity as Hebbian storage | Attractor memory | A presynaptic spike just before a postsynaptic one strengthens the connection; the reverse order weakens it. Showing a pattern writes it into the weights. | Hopfield 1982; Bi and Poo 1998, J Neurosci 18:10464, doi:10.1523/JNEUROSCI.18-24-10464.1998; Song, Miller and Abbott 2000, Nat Neurosci 3:919, doi:10.1038/78829 |
Recalling a whole from a part (h_complete)Link to this section
The idea. If a group of cells excite each other symmetrically, each weight to another equal to the weight back, the group has a stable state in which all of them fire. A few cells nudged into that state pull the rest after them, and the pattern stays after the cue is gone.
The goal. Build a pattern of several cells that is silent by itself. memory reads the whole pattern; cue reaches only part of it, for a moment (a 200 Hz Poisson burst from 300 to 600 ms). After the cue the whole pattern must fire, including the cells the cue never touched, and keep firing. A group of cells can be built from two groups (the cued cells and the rest) that excite each other and themselves.
What the pass rule asks. At most 2 spikes in the pattern before the cue. In the recall window (900 to 1700 ms, long after the cue) at least 80 percent of the pattern's cells fire at 5 Hz or more on average, and at least 80 percent of the cells that the cue never reached fire. The structure check wants several cells in the pattern, excitation among them in a loop, a cue that reaches some but not all, and connections that are roughly symmetric.
What is simplified. The sustained rate is far above real delay activity (there is no balanced inhibition), and the pattern is a group of cells, not a plus-and-minus vector.
Par: 8 cells.
Stuck?Link to this section
A group of cells that excite each other symmetrically has a stable state in which all of them fire. A part of the group, nudged into it, pulls the rest along.
Two memories, no blends (h_two)Link to this section
The idea. Store two patterns in one network. Each pattern is a group of cells that excite each other. If nothing separated the groups, the network could fall into a blend of both: a spurious state, a memory nobody stored. Inhibition between the patterns, as symmetric as the excitation inside them, makes each a state the network can fall into on its own and a blend not one.
The goal. cue_a calls up pattern A and cue_b pattern B; a and b read the cells of each. After cue_a only A must be lit. After cue_b only B, overwriting A. When both cues come at once the network must pick exactly one pattern, not both. A pattern's cells must stay lit after their cue is gone.
What the pass rule asks. Quiet before the first cue; after cue_a at least 80 percent of A's cells fire and B stays silent; after cue_b the same for B; after both cues together exactly one pattern is lit, the other silent. The structure check looks at the edge list: the two read sets must differ, each must excite itself in a loop, the two must inhibit each other (directly or through an interneuron), the connections between the groups must be roughly symmetric, and each cue must reach its own pattern.
What is simplified. The two patterns are disjoint groups, not overlapping vectors, and the inhibition between them is built in, not the result of an outer-product rule.
Par: 8 cells.
Stuck?Link to this section
With symmetric excitation inside each pattern and symmetric inhibition between them, the pure patterns are the states the network falls into; a blend of both is not.
Writing a memory with plasticity (h_learn)Link to this section
The idea. So far the pattern was built in. In a brain it is written by experience. Spike-timing-dependent plasticity (STDP) makes a connection grow when its presynaptic cell fires just before its postsynaptic one, and shrink in the other order. Cells that fire together wire together.
The goal. Make the connections among a group of cells plastic and start them weak. First, cue reaches part of the group and recalls nothing. Then show drives the whole pattern for a while, so its cells fire together and the connections among them strengthen. Then the same cue must recall the whole pattern, and it must stay lit after the cue. The connections must really have grown, and they must start symmetric. Too strong at the start and the cue completes the pattern before it was learned; too weak or too slow and nothing is learned. The synapse has a learning rate, an asymmetry between strengthening and weakening, and a ceiling for the weight. These are in the inspector of the projection; see synapses and modulators.
What the pass rule asks. Before the exposure the cue completes nothing (at most 3 spikes in the cells it never reached); the mean weight of the plastic connections has grown at least 1.5 times; and after the exposure at least 80 percent of the cells fire at 5 Hz or more, including at least 80 percent of the cells the cue never reached. The structure check wants the pattern's cells to excite each other in a loop through plastic synapses, symmetric at the start, with show reaching every cell and cue only some. Whether the exposure is strong or long enough is left to the behaviour.
What is simplified. One second of exposure stands for minutes of learning. The plastic synapse is NEST's stdp_synapse, which is excitatory only, so the inhibition of a bigger network cannot be learned.
Par: 8 cells.
Stuck?Link to this section
Spike-timing-dependent plasticity strengthens connections between cells that fire together, so showing a pattern writes it into the weights.
Chapter 4: Tuning and decodingLink to this section
Many sensory neurons answer best to one value of a stimulus: an orientation, a direction of movement, a pitch. A plot of response against the stimulus value is the cell's tuning curve, and the value at its peak is the cell's preferred value.
The afferent cells of these levels are driven by an input of the kind tuned_poisson. The task sets a stimulus angle, and each cell the input reaches fires with Poisson spikes at a rate that depends on how far its preferred angle is from the stimulus. The curve has a bell shape, a von Mises function: the circular cousin of the Gaussian bell, with a base rate, a peak rate and a width. The preferred angles are spread evenly in the order the cells were made: cell k of N prefers 180 k / N degrees for an orientation (which repeats every 180 degrees) or 360 k / N degrees for a direction.
In the lab a dial shows the stimulus, a bar per cell shows the response, and a single cell's tuning curve fills in as the orientations come round. For t_decode the decoded direction appears as an arrow. See inputs and outputs for the tuned_poisson input and the population-vector output.
| Level | What it teaches | Know first | The real science | Citation |
|---|---|---|---|---|
| A cell tuned by its inputs | Tuning from convergence | Convergence | Hubel and Wiesel found cat cortical cells that fire for a bar at one orientation and proposed that they receive input from cells whose receptive fields lie along a line. Whether real tuning is feed-forward, recurrent or both is debated. | Hubel and Wiesel 1962, J Physiol 160:106, doi:10.1113/jphysiol.1962.sp006837 |
| Sharpening a tuning curve | Narrowing tuning with inhibition among cells | Tuning curve, inhibition | In recurrent models, inhibition among cells of different preferences makes the response tuning much sharper than the input's. Evidence for and against has accumulated for decades. | Ben-Yishai, Bar-Or and Sompolinsky 1995, Proc Natl Acad Sci USA 92:3844, doi:10.1073/pnas.92.9.3844 |
| Reading a direction from a crowd | The population vector | Tuning curve | Georgopoulos and colleagues found monkey motor-cortex neurons broadly tuned to arm-movement direction. One cell predicts the direction badly; the weighted sum of preferred directions points the right way. | Georgopoulos, Schwartz and Kettner 1986, Science 233:1416, doi:10.1126/science.3749885 |
A cell tuned by its inputs (t_tuned)Link to this section
The idea. A cell that listens to the right inputs inherits a preference. This is the Hubel-Wiesel picture of an orientation-selective cell built from aligned inputs.
The goal. The sensor stim is the orientation of a bar, from 0 to 180 degrees, shown for a few hundred milliseconds in turn (eight trials of 300 ms: 0, 45, 90, 135, then the same again). Create a population of afferent cells (parrot cells, which just repeat the spikes of their input) and connect stim to it. Then build one more cell, resp, that listens to some of the afferents and not to others, so that it responds to 90 degrees much more than to 0, and more than to 45 and 135. Which afferents it listens to, and how strongly, is the whole puzzle. A cell that weighs them all alike has no preference. The afferents can be separate populations, taken in the order you create them.
What the pass rule asks. At 90 degrees the cell fires at 12 Hz or more (8 in Guided); its rate at 90 is at least three times its rate at 0; and at least 1.4 times its rate at 45 and at 135, so that it is not just "anything near 90". The structure check wants resp to be a cell that is not one of the afferents and that weighs them unequally. The preference is judged by weights, not by the number of connections: graded excitation (the preferred afferents strongest) and inhibition from the flanks both count.
What is simplified. The tuning curves are given: a game-chosen width, peak and base, Poisson spikes, evenly spaced preferred angles. Real afferents are neither evenly spaced nor Poisson. The level shows that convergence can work, not that the cortex does it this way.
Par: 9 cells.
Stuck?Link to this section
A cell inherits the preference of the inputs it listens to: connect it to afferents that prefer one orientation and leave out the rest.
Sharpening a tuning curve (t_sharpen)Link to this section
The idea. Afferents with broad tuning curves feed a population of cells, one cell to one afferent, and the cells inherit the broad tuning. Inhibition among the cells can narrow it: the cell that prefers the stimulus pulls ahead and pushes its neighbours down.
The goal. stim shows an orientation from 0 to 180 degrees. Make eight afferent cells and connect stim to them. Make eight cells that each listen to their own afferent, using the "neuron k to k" segment wiring (neuron k of one group to neuron k of the next). resp reads the eight cells. The cell that prefers the orientation shown must fire well, and the cells that prefer 45 degrees away from it must fire at only a small fraction of that. Without inhibition among the cells the tuning is as broad as the afferents'. A cell that also excites itself holds its lead against the others; that is optional, and how strong each is decides how much it sharpens.
What the pass rule asks. The cells fire at 10 Hz or more at their preferred orientation (7 in Guided), and the side rate, the cells 45 degrees away, is at most 58 percent of the peak rate (62 in Guided, 52 in Researcher). Without any loop the ratio is about 67 percent. Because a ratio of two small spike counts is noisy, the episode is long: 24 trials of 300 ms, read over 48 trial responses. The structure check wants the eight cells driven by the afferents by excitation, not themselves afferents, and inhibiting one another directly or through a cell they excite.
What is simplified. Whether real cortical tuning is sharpened by inhibition is debated; the level shows the mechanism working. Self-excitation is optional: the reference circuit has it and sharpens more steadily, but inhibition alone can pass.
Par: 16 cells.
Stuck?Link to this section
Inhibition among cells with different preferences narrows the tuning of each: the cell with the strongest input wins and pushes the others down.
Reading a direction from a crowd (t_decode)Link to this section
The idea. No single cell tells the direction of a stimulus: each is broadly tuned and noisy. A crowd does. Weight every cell's preferred direction by how much it fires and add them up. The population vector points where the stimulus is.
The goal. The sensor stim shows a direction from 0 to 360 degrees, a different one in each of six 400 ms trials. Make twelve afferent cells and connect stim to them; cell k of twelve prefers 30 k degrees. Then make a population of cells that listen to the afferents, and let heading read it as a ring. heading takes the cells in the order they are listed and puts cell k of N at 360 k / N degrees, with cell 0 at the top, so the cells must be in the order of their preferred directions. The decoded direction must be near the shown one, trial after trial. If the cells are listed in the wrong order, or each listens to the wrong afferent, the arrow points somewhere else. The ring readout is described in inputs and outputs.
What the pass rule asks. The readout fires (at least 10 spikes per trial on average); the mean absolute error between decoded and shown direction over the six trials is at most 25 degrees (35 in Guided, 18 in Researcher); and no single trial is more than 60 degrees off (80 in Guided). The shown directions are random, a different draw for each seed, so a lucky guess does not repeat. The structure check wants heading to read cells that are not the afferents, that listen to them by excitation, and at least eight listed cells.
What is simplified. The population vector is one linear decoder, not the best one: for Poisson cells with bell-shaped tuning a maximum-likelihood readout is more accurate. The tuning curves are given.
Par: 24 cells.
Stuck?Link to this section
One cell is broadly tuned and noisy; the sum of all the cells' preferred directions, weighted by their firing, points at the stimulus.
The spike-triggered average tabLink to this section
The wing adds a seventh tab to the Analysis screen, STA. It belongs to the same family of ideas as tuning: instead of asking which stimulus a cell prefers, it asks which input pattern makes a cell fire. This is reverse correlation.
You choose a population and press Record. The first cell of the population is driven by a Gaussian white-noise current (random, with no pattern; it is seeded from the circuit's seed, so a run repeats) for the length you choose, at most 20 seconds; 10 seconds is usually plenty. Each time the cell fires, the program takes the stretch of current that came before the spike. The average of all those stretches, minus the average of the whole current, is the spike-triggered average, drawn as a deviation from the mean input in picoamperes.
Reading it: a bump just before time zero means the cell responds to a brief rise in current; a bump followed by a dip means it responds to change, like the phasic cell of Only the change counts. A grey band shows what chance alone gives with that many spikes, so a lobe inside the band is not a filter. If the cell fires fewer than five spikes under the noise, the tab says "too few spikes" instead of drawing noise. The mean and spread of the noise current come from a per-model table chosen near each cell's threshold. For the other analysis tabs see the lesion lab and analysis tools.
Related pagesLink to this section
- The lab: where the new views sit.
- The core loop and stars, par and the hint ladder: build, check, verdicts and the stars.
- The parts bin: the Izhikevich and adaptive cells, the rate units and the siegert unit.
- Inputs and outputs:
rate_drive,rate_value,tuned_poissonand the population-vector readout. - Glossary: every term used above, one line each.
- What is real: the honest account of what is NEST, tuned and illustrative.
- The Bestiary and the body chapters: the other two wings.
Something wrong or missing? Edit this page on GitHub (source) or open an issue.