Lecture 20 - Spatial & Cognitive Maps II

What are the three main types of cells that make up spatial representations?

1. head direction cells

• compass like

2. place cells

represent given location in map

3. grid cells

• regular pattern of firing across space

What are place cells?

place cells fire selectively at one (or a few) locations in the environment

representation of a location in allocentric coordinates

What does place cells look like in a linear track?

place cells can also be measured in a linear track

they respond at one location in the track

cells firing at all locations of the track allowing you to represent the space 

What does shortening the linear track reveal about place cells and hippocampal spatial maps?

When track is shortened, place fields compress (place cells fire in proportionally shifted positions)

• shows that the hippocampal spatial map is flexible and scales with environmental deformation

What is shown in experiment where rats control sound frequency with a joystick?

1. Rats press on a joystick to generate a tone and increase its frequency

2. They must release the joystick within a specific frequency range

3. This creates a 1-dimensional “track” in sound-frequency space

What was found from the sound 1D experiment?

hippocampal cells are often tuned to a specific location in the sound sequence

similar to tuning in linear track

How do we know that a “sound place cell” is not just an auditory-system neuron?

Because the neuron fires only during the active behavioural task (when the rat uses sound frequency as a navigational dimension with joystick), but does NOT fire when the exact same sound is played passively

a place cell - like neuron does not respond if the same sound frequency profile is played passively 

If it were just an auditory neuron, it would respond to the sound in both cases

What does tuning correspondence across spaces mean?

some cells in CA1 has a single preferred point in both:

• Physical space

• Sound frequency space

This means the neuron is not just coding physical position.
- It is coding a generalized position along any ordered dimension.

therefore hippocampal cells can encode cognitive spaces beyond real physical locations

Do MEC cells just encode physical space?

they can also encode positions along a sound sequence, meaning they act like “grid cells” in an abstract space

some cells show tuning to multiple locations in sound sequence as well as grid cell tuning in random exploration

• cells are representing positions in ABSTRACT COGNTIVE SPACE

What does this show in terms of MEC cells and response in different tasks?

These panels show the same neurons recorded in two different tasks:

1. Sound sequence task

2. Spatial exploration task

• Cell 5: Active in both sound sequence and spatial exploration — shows tuning to both.

• Cell 6: Not active during the sound sequence task, but shows spatial grid firing — so it’s only a “normal” grid cell.

Does grid cell maps spacing vary?

yes different grid cells tile the space with lattices with different orientations and different scales

• grids more ventral in entorhinal cortex have a larger metric

• grids more dorsal in entorhinal cortex have a smaller metric

What does the “spacing correspondence across spaces” result show about grid cells?

grid cells with wider fields in the spatial environment tended also to have wider fields in the SMT (sound mapping task)

• precise cells in navigation = precise cell in SMT

• suggesting shared neural mechanisms

  • sharing properties in completely abstract spaces

What does the spacing correspondence across spaces task show for shared neural mechanisms?

these results suggest that common circuit mechanisms in the hippocampal entorhinal system are used to represent diverse behavioural tasks

• supporting cognitive processes beyond spatial navigation

Do we have cognitive maps across dimensions?

yes

spatial maps represent cognitive maps that are abstract

neurons encode abstract audio space in the same way as they do for spatial navigation

How do researchers identify grid-like representations in the human brain? What techniques do they use?

entorhinal cortex is not a common area of epilepsy so we often don’t need to put electrodes in this area

using fMRI and symmetry we can leverage the structure of grid fields to identity area that have grid-like representations in the human brain

What regions in human brain have grid-like representations?

entorhinal cortex and mPFC 

• identified when humans navigated a virtual world

How can we navigate abstract spaces? Give example.

we can ‘navigate’ a 2D space of bird shapes in which the leg length and the neck length varies

• changing just neck length means only navigating one direction in 2D space

• changing neck and leg length means travelling at a 45 degree angle in the 2D space

this is similar to navigation in a physical/virtual 2D space

What areas in the brain are activated while navigating virtual spaces and 2D abstract spaces? What does this suggest?

navigation in these two spaces identifies similar areas:

• entorhinal cortex

• mPFC
  

suggests that the navigation is not used simply for navigation in the physical environment but that it is a more general system for navigating spaces that have relational structure 

What is transitive inference?

type of reasoning where you figure out a new relationship by linking together known ones

Example:

• You learn A > B

• You learn B > C
  → Without being told, you infer A > C

Why does transitive inference require a “map”?

the brain uses the structural relationships between items—just like it uses the layout of 2D space

• to make new inferences it was never directly taught

• By learning relations (A > B, B > C, …), the brain builds an internal map of the hidden structure, allowing it to infer A > G even without seeing that comparison before

What is the Transitive Inference Task — Step-by-Step? What is the transitive test vs novel non-transitive test?

1. Train the animal on adjacent pairs
   The animal learns which item is “better” in each pair:

• A > B

• B > C

• C > D

• D > E
  (Choosing the correct item gets a reward.)

2. Animal forms an internal map
   Even though it never sees the whole sequence, the animal can infer:
   A > B > C > D > E.

3. Test with new pairings that were never shown during training

• Transitive test: B vs. D → should choose B.

• Novel non-transitive test: A vs. E → should choose A

Why is A vs. E a non-transitivity test?

E is never rewarded

A is always rewarded

you don’t need a map to know that A gets a reward and E does not get a reward

vs.

B vs. D

it depends on where the other is in the map you’ve built 

What did the experiment for transitivity show?

If the animal chooses correctly on never-seen-before pairs, it shows it understands the implicit order—this is transitive inference (IT HAS BUILT THE MAP)

What brain systems are necessary for transitive inference in preferences?

The hippocampus and navigation system

• entorhinal cortex/perirhinal cortex

• fornix (hippocampus output)

Which probe pair shows a deficit after hippocampal/entorhinal/perirhinal lesions?

The BD relational probe, which requires computing transitivity

since you have to compare (B > C > D)

• depends on hippocampal relational mapping

Do lesion animals show deficits on the end-anchored test (A vs. E)?

no

A is always rewarded and E is never rewarded, so the animal can solve it without relational reasoning

Do lesion animals show deficits in BC & CD tests?

no!

since they are inner premise pairs that don’t require transitivity inference

What is the overall conclusion from the lesion study?

The hippocampal–entorhinal system is required for inference, not for simple pairwise learning.

Recap: What do hippocampal place cells represent in a spatial task?

They fire in specific locations

• create a map of where the animal is in the environment

How does reinforcement learning relate to place cells?

Place cell firing can be viewed as encoding an environmental state (RL) space

• each place cell represents a state, and the pattern of firing shows how states connect

What does the bottom (odor-choice) task show about non-spatial tasks?

Neurons can also represent states that are not spatial

• bottom task is for odor, choice, and outcome

  • mapping the structure of the task just like place cells map space

What is the key idea of the slide?

Both spatial and non-spatial tasks rely on the brain representing an organized “state space” of events or locations

What is Harlow’s task? What does it test?

object A gives reward

object B gives no reward

• The animal is shown two objects at a time (like shapes).

• Only one of the two objects gives a reward.

• The objects change every trial (different shapes, colors, etc.).

They are not just learning each pair → they learn how to learn the structure of the task

What does the performance graph in Harlow’s learning set experiment show?

Once the monkey has learned the high-level rule (abstract structure) of the task — that in every new pair, only one object is rewarded — it no longer has to explore both objects over many trials

ONLY NEEDS ONE TRIAL

a. you pick right one and keep picking

b. you pick wrong one, pick other and keep picking that one

What do OFC lesions lead to for tracking changing reward probabilities?

monkeys with no deficit learn to track changing probabilities of reward

OFC lesions cause deficit so that they can track changes in reward unless the best option switches (they keep picking same option even when reward level changes)

How do healthy (non-lesion) vs. OFC-lesion patients differ in how they use reward history to make decisions?

• Non-lesion patients:
  They track previous rewards using the task’s internal map. They use past information about whether an option was rewarded or not to compute which option is best on the next trial.

• OFC-lesion patients:
  They don’t use the past reward info stored in the map.
  Instead, they rely on off-diagonal terms and choose based on immediate or mis-assigned signals, leading to choices that ignore actual reward history.

Why can’t OFC patients use past information in maps to help in task?

Because their internal “map of the task” (state space) is broken.
If you can’t represent the states correctly, you can’t correctly connect which choice caused which outcome.

• past rewards can’t be assigned to the right action

How does a normal, intact OFC support correct credit assignment in tasks?

1. The brain has a correct state space, representing the sequence:
   odor → left/right choice → reward or no reward.

2. When a reward occurs, the brain correctly links it to the action just taken:
   “This reward belongs to the choice I just made.”

3. Behavior shows diagonal influence in the rewards × choices matrix → recent rewards guide the next choice.

Why can’t OFC-lesioned animals use past rewards to guide their choices? What do we call this structure?

1. The OFC lesion breaks the state space, so the animal can’t distinguish the task’s states

2. cannot link either reward to the actual choice that produced it (right or left)

• Rewards fail to be assigned to their causal choice

• Reward history becomes useless.

• Their behavior shows off-diagonal influence → they rely on mismatched, incorrect associations

BREAKING OF TASK STRUCTURE 

Can knowledge of task structure help with performance? Give example with big vs. small reward environment.

yes!

low reward environment → give large reward → positive RPE

if you are in big reward environment, you know you will get big rewards

• as soon as you get small reward, you know that the block identity has switched

behaviour can adapt in ‘one-shot’ as experience of a small/large reward is enough to know that block identity has switched

How do we see evidence of the abstract cognitive map in the dopamine firing?

firing of dopamine neurons is consistent with a map of task space (model-based)

animal has expectation to be in a high or low reward block and RPE scales accordingly

1. The animal tracks which block it’s in (high-reward vs low-reward block)

2. RPEs scale based on the inferred state

e.g. When mice receive an intermediate reward, dopamine responses shift depending on what block the animal believes it is in (high vs. low reward block)

What is a map-based representation?

metric based

locations are coded in terms of Euclidean coordinates

• Locations have positions on a grid

  • latitude/longitude

Distances matter.

Example: Knowing two buildings are 300m apart

What is a graph-based representation?

• Represents information as nodes connected by links

  • locations = nodes

  • paths between locations = links

• Only the connections matter (not physical distance)

• Example: Subway maps

Can map-based vs. graph-based representations exist together?

yes: both can exist simultaneously.

• We can switch between them depending on what’s needed

• Example: Knowing both the physical layout of a town (map) and the ordered path you must follow (graph)