explicit memory theory Flashcards
hippocampus important for what tasks (2)
- spatial tasks
- relational memory
what is relational memory
memory of relationships bw objects and events
fundamental functions of hippocampal network (2)
- spatial and memory problems can both benefit from structural abstractions (give a structure to sensory info)
- figure out how best to organize sensory events into abstract structure (graph or map) for the purpose of sensory predictions
what does neural activity in entorhinal cortex correlate with (animal behavior) (2)
- speed of animal as it walks around (speed cells)
- direction the animal is facing (head direction cells)
input to (a) speed cells (b) head direction cells
(a) proprioception, visual
(b) angular velocity of head shifts pattern of activity (change in direction)
how does spatial info create a map
path integration (by creating vector that represents position and direction in comparison to earlier reference point)
what are abstract structures, such as maps, used for
organizing sequences of sensory information for the purpose of prediction
how is map formation similar to semantic memory formation and what does it suggests
both are a two-step process (remembering life episodes/spatial info + inferring info from what you have already?)
* suggests that maybe neuronal algorithms of map formation are the same as the ones underlying explicit memory formation (neuronal algorithms of real and mental space are the same)
what are cognitive maps and what are they required for (3)
systematic organization of knowledge required for (a) speed of learning (b) ability to generalize (c) ability to make complex inferences
what do cognitive maps directly aid in
construction and accumulation of spatial knowledge, allowing the ‘mind’s eye’ to visualize scene or event
how can spatial thinking also be used for non-spatial tasks
non-spatial info can often be organized in abstract structures (word clouds)
deep neural net machine learning algorithms that are used for pattern recognition have (3)
- structure -> # of layers and nodes, how each node gets activated (inputs, outputs)
- objective function -> goal (like label things in input or predict what next input will be or identify best action); has clear right and wrong answer
- learning function -> method of adjusting strength of each connection to better achieve objective function (backpropagation technique) (what gets me closer to the right answer?)
how recommender systems worked before and how they work now
before -> asked questions to user (do you like comedy? romance?); give a score on how much would like movie
now -> have a matrix of scattered info of likes; program creates matrixes (of likes of people and movie recs) that if multiplied, gives input matrix; fills in matrix (infers) and adjusts connections
what do all spatial environments share
common regularities of euclidean space that define which inferences can be made, and which shortcuts might exist (like moving S->E->N->W brings you back to where you started)
ex of how structural regularities can apply to non-spatial relational problems (2)
- transitive inference problems (depend on hippocampus) require stimuli to be represented on abstract ordered line (A > B and B > C implies that A > C)
- abstraction of hierarchical structure permits rapid inferences when encountering new social situations (learn structure, fill in labels)
what structural relationships can be formalized on a connected graph (3)
- social hierarchies
- transitive inferences
- spatial reasoning
main input pathway to hippocampus
entorhinal cortex in temporal lobe
what does (a) MEC; (b) LEC provide to hippocampus
- learned structural info
- perceived sensory info (what experienced at the time)
role of hippocampus in relation to MEC and LEC
hippocampus combines structural info from MEC and sensory info from LEC (associates sensory info with place -> position in coordinate map or place in abstract theoretical structure)
what does our brain do to account for speed of learning, ability to generalize and ability to make complex inferences
identifies how sensory info is structured, whether on spatial map or in abstract space
what does factorized representation of sensory experiences enable (2)
- flexible recombination of info to represent novel experiences
- generalizing structural knowledge to new situations
structure of task to create a machine learning algorithm to test factorization idea (3)
- function of machine is to predict what it will see after taking a step in the indicated direction
- info provided at each step is (a) directional cue and (b) sensory stimulus
- job of machine is to predict sensory stimulus from directional cue
consequence of (a) not knowing structure of space and of (b) knowing structure of space
(a) have to memorize each transition -> have to enter each spot from each direction; if object appears in different places, multi-step memories are needed, exponentially increases memory load
(b) only need to visit each spot once, from any direction -> may be initially hard to learn underlying structure, but once learned, can be applied to new situations
aspects of supervised learning (4) + ex
- identify how a pattern of inputs related to a pattern of outputs using real-life examples
- training data consists of input-output pairings -> figure out how to adjust synaptic weights to get right answer
- training data is labeled
- used for recognition tasks
* ex = train machine to read hand-written cheques; give it a million cheques and tell it exactly what is written on them (real person has to give answers, time consuming); if good input-output function learned, able to read new cheques wasn’t trained on)
aspects of unsupervised learning (self-supervised learning) (6)
- no humans, give it the internet
- identifies patterns and underlying components in unlabeled data
- ca be thought of as dimensionality reduction (factorization)
- exhibits self-organization as it recognizes hidden patterns in large data sets
- used for generative tasks
- tries to mimic data that it is given, using error in its mimicked output to correct itself (with each word, predict what next word is -> right answer in framework)
constraints in the inferred map that the machine has to learn
- 2 sensory stimuli cannot be associated with the same spatial position
- each unique sensory cue can only be associated with one spatial position (spatial representation of given position must be identical when approached from any direction)
explain machine learning algorithm (7)
- start at position
- given where i think i am, what do i expect to see?
- memory retrieval from structural data
- real sensory data (if error, take note and store it -> will be used in backpropagation to adjust synapse weights)
- given that i see this, where am i likely to be?
- memory retrieval from sensory data
- infer memory and path integration -> adjusting of synaptic weights
data of machine learning (a) after trained in a few environments (b) after visiting more environments
(a) makes many visits to each node to remember it because doesn’t understand structure of graph and hasn’t learned how to use memories
(b) learns common structure, machine correctly predicts node on 2nd visit regardless of edge taken -> understands both rules of graph and how to store and retrieve memories
location, input and info encoded of (a) place cells (b) grid cells
(a) hippocampus; input from sensory info and grid cells; single spot
(b) MEC; spatial position (abstract)
activity of nodes in deep neural net that encode abstract structure looks like
grid cell activity
grid-like representations across environments in TEM (3)
- maintained grid-like structure
- structure rotated
- correlation structure preserved -> grid cells that fire next to eo in env 1 fire next to eo in all env
how are ‘neurons’ of the model similar to hippocampus (5)
- ‘neurons’ only become active when both structural and sensory inputs to them are both active
- place cell appears to randomly relocate across different environments
- place-like fields span multiple sizes
- if allowed to move freely, but biased to prefer hugging the walls -> activity looks like that of border cells
- if allowed to move freely, but biased to prefer approaching objects -> activity looks like that of object vector cells
how are place cell representations not randomly determined
they are constrained by the learned structural representations provided by grid cells
in real animals, is place cell remapping random (relative to grid cell activity) or are there PC-GC pairs
graph correlating PC activity with GC activity shows low, but decent correlation (significant) bw activities
what do environment transitions cause
grid cell realignment, which triggers PC remapping because PC are constrained to GC structure
why isnt correlation bw PC and GC activity perfect
because PC are influenced by more than one GC or if do get input from only one GC, don’t know which one
firing of PCs to arbitrary spatial abstractions (circular track, but only get reward on 4th lap) (3)
- some fired for location on track (classical PC)
- some fired only on one of 4 units (consider whole thing (4 laps) as one unit)
- some variously fired as function of lap number (counting laps)
what do (a) spatial place cells (b) other cells predict in arbitrary spatial abstractions
(a) sensory events that are unchanged between laps
(b) represent position within 4-lap circle
when are hippocampal cells active
only when receive simultaneous input from sensory and structural representations (MEC and LEC)
where can place cells remap to
other peaks within MEC cells field if it also receives sensory input there
lap specific PC activity is driven by what and why
driven by grid cell activity because sensory observations at every lap, but grid cell activity specific for reward lap -> hippocampal cells only fire with both inputs so fire specific for grid cell activity
lap specific PC in dif env
cells that show lap specific activity in env 1 are also lap specific in other env
what could semantic info be a representation of
structural info about how the world work that we extract from repeatedly seeing same basic configurations of sensory stimuli across time
