Working models of working memory Omri Barak and Misha Tsodyks 2014, - - PowerPoint PPT Presentation

Working models of working memory

Omri Barak and Misha Tsodyks 2014, Curr. Op. in Neurobiology Referenced by Kristjan-Julius Laak Sept 16th 2015 Tartu

Working models of working memory

Working models of working memory

• Def. Holding and manipulating information for short periods
• f time with no structural changes involved.

… refers more to the whole theoretical framework of structures and processes used for the temporary storage and manipulation of information WM

The Magical Number Seven, Plu lus or r Min inus Two: Some Limits on Our Capacity for Processing Information

George A. Miller, 1956

For a sequence of words in a sentence you get the structural idea in the

• end. – example of a WM task

Levels of computation

Challenges

Data-driven

Analysis of behaviour

Manipulate several items simultaneously

Neurophysiological observations

∑ Irregular firing patterns ∑ Activity is not stationary ∑ Different neurons have different firing profiles

Challenges

Data-driven Computational-driven

Analysis of behaviour

Manipulate several items simultaneously

Neurophysiological observations

∑ Irregular firing patterns ∑ Activity is not stationary ∑ Different neurons have different firing profiles Network activity should be stable to retain memories

Classical idea of WM (still „working memory)

Itskov, Hansel, Tsodyks, Front Comput Neurosci. 2011

Classical idea of WM (still „working memory)

WM – stationary persistent activation of selective neuronal populations

Classical idea of WM (still „working memory)

WM – stationary persistent activation of selective neuronal populations Recent advantages explain WM also by

• 1. Short-term synaptic plasticity (STSP)
• 2. Recurrent excitatory and inhibitory networks (I-E)
• 3. Intrinsic network dynamics

Delay-effect

How can we hold many items simultaneously?

• r How to overcome the mechanistic challenge of retaining several

items in WM?

How can we hold many items simultaneously?

• r How to overcome the mechanistic challenge of the interference

between the activation of different items? Alternatively: Capacity of the network

Overcoming interference: Sparse patterns

Every item is represented by a small fraction of neuronal population

Experiment I-F spiking neuron model Amit et al. 2003, Cerebral Cortex

Overcoming interference: I-E E balance

The balance of inhibition and exhibition determines a) No. of items network can hold b) Mode of failure (fade out, merge)

Amit et al. 2003, Cerebral Cortex

Dorsolateral prefrontal cortex (dlPFC)

Boosting of capacity through dlPFC top- down signals. dlPFC has nonspecific, excitatory connections to IPS. (A) If dlPFC has low activity, only two items are stored. (B) When dlPFC activity is high, all four items are remembered. Edin et al. 2008, PNAS

• 1. Synaptic facilitation

Short-term SF – temporarily modify synaptic efficacy in response to stimuli Phenomenological model of Ca-mediated transmissioon ->

Rolls et al., 2013 ,PLOS One

SF continued

• 1. .. Prolongs memory lifetime by reducing the inherent drift of the

system

• 2. .. Replaces persistent activity!

WM sustained by Ca+ facilitation

Mongillo et al., 2008, Science

1200 species of proteins in post syn end, and only 6 Ca+ ions

SF continued

• 1. .. Prolongs memory lifetime by reducing the inherent drift of the

system

• 2. .. Replaces persistent activity!
• 3. .. Enables non-linear relationship between pre and post syn neuron

Excitatory recurrent currents (~NMDA) make persistent activity models more realistic

Wang et al. 2012 Neuron

• 2. I-E Balance

We know that there is a balance and the activity is irregular … This has been a puzzle for neuroscientists.

Memories are stabilized by fast inhibition and slow excitation Negative feedback loop idea from engineering

• 3. Intrinsic dynamic mechanism

There’s no persistent activity

• > Idea of stable states

during WM tasks

See also: Maass, et al. 2002, Neural Comput Rainer, Miller, 2002, EJN

Reservoir computing w w Liquid State Machines

But maybe some states?

Training initally random network gives better results

Barak, et al., 2013, Science Direct

Dynamic attractors – chaotic network + learning

Laje, Buonomano, 2013, Nature Neuroscience

Conclusions

Biological systems don't choose one mechanism. It is highly possible that many mechanisms mentioned are utilized by the brain.

Thank you!

Kristjan-Julius Laak julius.laak@gmail.com Computational Neuroscience lab (neuro.cs.ut.ee) University of Tartu