Notesale: Turn your study into money

Document Preview

Extracts from the notes are below, to see the PDF you'll receive please use the links above

---

27th November

Hippocampus and Associative Memory

The way memory is found is by a search of the content of memory, rather than an arbitrary filing system
...
g
...
Dewey decimal
system for categorising books in the library, where you have to find a numerical code for the book you are
looking for, so the labelling is via a second party
...

Spurious memories are formed because of lack o capacity of the neurons
...
But if there is more input, (e
...
more facts from a certain
episode), the correct memory episode may be recalled (pattern completed) accurately
...
However, this network is very powerful, and with the correct

27th November
biological properties of neurons taken into account, e
...
synaptic connectivity, neuron dynamics, the
behaviour of human and network is similar
...
To
support a pattern of activation, connections should be positive between units in the same state (i
...
1 / 1 or 1/ -1) and negative between units in different states (1/ -1 or -1 / 1) - i
...
si sj wij > 0
The ‘frustration’ or ‘energy’ of the system is how much this is not true –
i
...
E = -∑ij si sj wij
The update rule changes each units activation to reduce the overall frustration, until the network ends up in
a stable state from which frustration cannot be further reduced
...

The minima is a pattern network that is stable, i
...
least
frustration/energy of the system
...
The graph is called energy
landscape
...
It might not be the exact original pattern
stored, but any other changes are going to make the mismatch greater- increasing the frustration
...
This is the difference is memory recall and recognition
...
g
...

In continuous attract networks, the line on the graph is almost flat
...
There are no discrete points for the network to go to
...
Store numbers, 0 or 1, for each connection in the matrix and
gives the connection strength of the whole network mathematically
...
active inputs (N) so that:

27th November
If x and y are both active or non-active, the product will be big, so the connection weight increases unless it’s
already too big
...

This network learns input-output associations- hetero-associative
...
It does pattern completion, but input pattern is associated with something else,
i
...
a different output rather than itself
...
g
...


Willshaw (1969); Marr (1971); McNaughton & Nadel (1990) - references for hetero-associative memory
matrices
...
This network learns to associate a pattern of
activity with itself: ‘auto-association’
...
e
...


The learning rule is the same
...
When
given a partial cue to completion, the network is going to want to store the input partial cue as a new
memory pattern
...
How
does brain differentiate encoding and retrieval when it is the same network used? If the brain doesn’t
spurious memories will be formed
...

Detonator synapses are super strong synapses
...

Summary of memory matrices

27th November






Analogous to the Hopfield auto-associative network but:
o connection weights are 0 / 1 and don’t need to be symmetric
o connection weights only increase (with pre-and post-synaptic activity)
o neuron activation values are 0 / 1 (not -1 / 1)
Performs pattern completion and error correction (suppressing things that aren’t part of memory)
Prone to interference (spurious memories) – need to use non-overlapping (e
...
sparse) codes, i
...

use small no
...

Need to distinguish learning from recall

The hippocampus as an associative memory network
The human brain is made of very many neurons (~100bn)
sending electrical signals to each other
...

Dentate gyrus- neurogenesis
...

The ‘brainbow’ mouse (Livet et al
...
The red lines is the recurrent collaterals of CA3, i
...
Hopfield
network
...
Small number of mossy fibres input onto CA3 also from dentate gyrus
...
They are the detonator synapses for the Hopfield network
...
So input from DG signals

27th November
brain to learn a pattern in the auto-associative network of CA3
...


Another network model by Hasselmo et al
...
Neuromodulators diffuse into large areas
of tissue and effect the neurons it comes into contact with
...
During encoding, you want lots of synaptic
plasticity because you want to change the strength of connections to match a pattern, but you want to
supress recurrent excitations as you don’t want to retrieve
...

During retrieval, low Ach levels, so the recurrent network is active
...
But it does have strong supporting
evidence
...

1
...
Mossy fibres from dentate gyrus to CA3
induce a sparse pattern of activity for autoassociative storage

27th November
3
...
Schaffer collateral from region CA3 to CA1 mediate hetero-associative storage and recall of
associations between activity patterns in CA3 and the activity patterns induced by entorhinal input
to region CA1
...
Perforant path inputs to region CA1 undergo self-organisation, forming new presentations of
entorhinal cortex inputs for comparison with recall from CA3
...

Categories of human memory

Patient HM (Scoville and Milner, 1957; Corkin et al
...
Posterior
hippocampus is sclerotic
Operation: September 1953 (27 years old)
Memory tested: April 1955, 29 years old)
No new memories formed since operation
...

IQ better than before the operation (112) and fewer seizures
Can’t find new home (after 10 months) remember new people, names etc
...
That’s
why space is more focused on rats
...
She will leave both rooms open and the birds take one
type of food into the other room
...

Semantic memory of new vocabulary, speech is frozen in 1950’s (Gabrieli et al
...
1997)










Jon has developmental amnesia
Peri-natal anoxia caused bilateral hippocampal damage, 50% volume reduction
Jon was premature, born at 26 weeks, on artificial ventilation for two months
Memory impairments first noticed at 5½ years of age
Jon has impaired episodic memory but preserved semantic and recognition memory, a normal
vocabulary, and obtained one GCSE (in History)
Highlights a distinction between episodic memory and semantic memory / familiarity-based
recognition?
This distinction can be demonstrated using the ‘doors and people’ difficulty-matched test of verbal /
visual recognition / recollection (Baddeley et al
...
g
...

HM describes things in the same way- no longer episodic, have become semantic, like a fairy tale, a
chapter from a book… so some people argue he has lost all of his episodic memory
...
For episodic memory, fast learning is indeed because events only happen
once
...
The problem with this network is that it gets corrupted quickly as more
memories are formed- catastrophic interference
...
Therefore the
brain would be better to have two memory systems- one that learns fast and one that learns slowly
...
This also explains the retrograde
amnesia, as memories are slowly transferred from hippocampus to cortex
...
People argue that during sleep, the hippocampus replays the event and trains the neocortex

27th November




to remember and associate the different inputs (smell, sight etc
...

o Is this transfer or training? Do memories truly leave the hippocampus?
Timescale? One night or twenty years → graded retrograde amnesia?
Generalisation / abstraction of meaning from multiple events leads to semantic memory from
episodic memory?

Hippocampus is an index (this index represents cross-modal biding) or convergence zone for disparate
neocortical sites which store the memory content
...
This is anatomical evidence
why memory transfer might be true
...

Familiarity-based unimodal recognition of content is therefore possible without the hippocampus
...
(2002)




Place cell firing is robust to the removal of subsets of cues in normal mice
In mutant mice, with removal of NMDA receptor in CA1, the place cells firing degrades with partial
cue
...
(2005)


Remember the energy minima, one input leads to one pattern completion, but then keep distorting
the input cue and the brain will think it’s a different pattern
...


27th November






There were rats shown two different arenas, square box and round box with different coloured
walls
...

Then they record firing of place cells in one environment while this experiment is morphed into
another experiment
...

This is proof of attractor network

---