What is working memory? (a.k.a. short-term memory) Sustained activity, Working Memory, Associative memory. Sustained activity in PFC (1) Readings:

Size: px

Start display at page:

Download "What is working memory? (a.k.a. short-term memory) Sustained activity, Working Memory, Associative memory. Sustained activity in PFC (1) Readings:"

Godwin Chapman
5 years ago
Views:

What is working memory? (a.k.a. short-term memory) The ability to hold information over a time scale of seconds to

.) Sustained activity, Working Memory, Associative memory Readings: C.

sample task: remember!red" Oculo-motor delayed response task: remember location of cue.

1 What is working memory? (a.k.a. short-term memory) The ability to hold information over a time scale of seconds to minutes a critical component of cognitive functions (language, thoughts, planning etc..) Sustained activity, Working Memory, Associative memory Readings: C.Constandinis and XJ Wang,, a neural circuit basis for spatial working memory, Neuroscientist, 2004 elayed match-to sample task: remember!red" Oculo-motor delayed response task: remember location of cue. Sustained activity in PFC (1) Lesion and inactivation studies demonstrate crucial role of Prefrontal Cortex (PFC) in working memory, in particular dorsolateral PFC (PFdl). 3 Fig. 1. Successive frames illustrate the sequence of events in

2 5211-O-3 ll ll ll i ll llllll R R i11 llll ll1 u ll O ~ ll lllll ll ;1 R llnll Sustained activity in PFC (1) Sustained activity in PFC (2) Lesion and inactivation studies demonstrate crucial role of Prefrontal Cortex (PFC) in working memory, in particular dorsolateral PFC (PFdl) c ll U ll ll lllll ll1 Y ll ll1 ll ll1 ll ll ll1 ll ll llld l ll ll1 ll ll ll lull 111 ll ll ll lull 450 C lllll ll ll1 F ll a lllll ll ll ll1 ll 3 llllll 11 ll ll ll ll1 ; ; iiiiimu ll Electrophysiology in macaques show that neurons in (PFC) continue to discharge even after the offset of transient sensory stimuli that animals are required to remember. -- interpreted as cellular basis of working memory C R lllll ll!1 m 11~11~11 111! A A :ll1,l ll ll Yl lllll ll llll ~ 1 : ll 1111 llllll A ll1 allll llllllll 1 ly yml 1 ll mi ll1 ll ll 11 lly 1 R ll ll l l ll ll lllll lllllllll ll cl El 225 O ' 225O c ll ll1 lmllll ll m 1 ll1l ll 111 ll ll ll ll ll 270" C R 11111~11~ lllll lllln~mmm ll l t H u 11 ii1111 n~~~wml~iilllilmiimi~~~~~~i i iii Fl ll :: ~~rzkl~:~l~ ll ammlll i /l 315O c ll : 1 ll ll Y a l!l ilmil ii ll ll ll\ ll/ ll1 ll! W ll ll ll O O-6 FG. 3. irectional delay period activity of a principal sulcus neuron during the oculomotor delayed-response task. This Funahashi et al, 1989 Sustained activity in PFC (3) Working memory vs Long-term memory Lesion and inactivation studies demonstrate crucial role of Prefrontal Cortex (PFC) in working memory, in particular dorsolateral PFC (PFdl). Electrophysiology in macaques show that neurons in (PFC) continue to discharge even after the offset of transient sensory stimuli that animals are required to remember. -- interpreted as cellular basis of working memory. Long-term memory : molecular or structural changes Short-term/ working memory: dynamic process that has not yielded to molecular characterization. Sustained Activity. PET and fmr studies confirm this in humans. E.g. Oculo-motor delayed response task. Neurons in PFC show activity during the delay -- tuned!memory fields". This activity was shown to represent the memory of the cued location, not the preparation of the motor response.

Sustained activity is very widespread Sustained activity in T Sustained activity is a widespread phenomenon LP and PP also have neurons which direction-specific memory fields, similar to PFC.

Memory related activity is also described in V3A, MT, V1, entorhinal cortex, Pre motor cortex, SMA, SC, basal ganglia... The distinct and cooperative roles of these areas remain unresolved.

corresponding to the cell's "place field".

3 Sustained activity is very widespread Sustained activity in T Sustained activity is a widespread phenomenon LP and PP also have neurons which direction-specific memory fields, similar to PFC. Also found in inferotemporal cortex (T), see e.g. Fuster and Jervey Example of a discrete working memory. Memory related activity is also described in V3A, MT, V1, entorhinal cortex, Pre motor cortex, SMA, SC, basal ganglia... The distinct and cooperative roles of these areas remain unresolved. Fuster and Jervey 1982 Place cells in hippocampus Head-direction cells Place cells : principal neurons in hippocampus that fire strongly whenever an animal is in a specific location of environment corresponding to the cell's "place field". often direction-selective suggests that the primary function of rat hippocampus = form a cognitive map of the rat's environment Neurons that are active only when the animal's head points in a specific direction within an environment. These cells are found in many different structure of the limbic system. Also continue to fire in darkness. visual cues seem to be the primary determinant of place cell firing, but firing persists in the dark, suggesting that proprioception or other senses contribute as well.

Neural ntegrator in Oculo-motor System Brain calculus : integration and differentiation in a pre-motor area in goldfish that is responsible for holding the eyes still during fixation, persistent

4 Neural ntegrator in Oculo-motor System Brain calculus : integration and differentiation in a pre-motor area in goldfish that is responsible for holding the eyes still during fixation, persistent neural firing encodes the angular position of the eyes in a characteristic fashion: below a threshold position the neural is silent, and above it, the firing rate is linearly related to position. integration (persistent activity) seems to be mainly due to network interactions, while differentiation (adaptation) seems mainly cellular and synaptic depression [Aksay,Gamkrelidze, Seung, Baker and Tank, Nat Neuro, 2001] 13 Working Memory and Sustained Activity How does a transient stimulus cause a lasting change in neural activity? A theory of working memory should answer: - how it is initiated? - why does it persist? - what makes it specific? - how does it ends? (a) Cortex Glutamate Glutamate Glutamate Striatum Thalamus (b) Parietal Prefrontal - reason for capacity limit? - relationship with attention, long term memory? GABA (c) GPi/SNr GABA (d) nferior temporal (TE) Mechanism : reverberations through connections (which?), or cellular? Lots of experimental and theoretical work to answer these questions, in PFC, H, Oculo-motor system 50 mv V m Ca / Can 3 Hz 20 Hz 3 Hz 2.5 s TRENS in Neurosciences

Attractor paradigm for persistent activity Since the 1970s it has been proposed that delay activity patterns can be theoretically described by!

5 Attractor paradigm for persistent activity Since the 1970s it has been proposed that delay activity patterns can be theoretically described by!dynamical attractors" Hopfield Networks A Hopfield net is a form of recurrent artificial neural network invented by John Hopfield (1982). Hopfield nets typically have binary (1/-1 or 1/0) threshold units: s i where sj state of unit j, and θ i is the threshold The weights have to follow: wii=0, wij=wji Hopfield nets have a scalar value associated with each state of the network referred to as the "energy", E, of the network, where: Hopfield Networks Associative memories Running: at each step, pick a node at random and update (asynchronous update) The energy is guaranteed to go down and the network to settle in local minima of the energy function. The Hopfield network is an associative/content addressable memory. t can be used to recover from a distorted input the trained state that is most similar to that input. E.g., if we train a Hopfield net with 5 units so that the state (1, 0, 1, 0, 1) is an energy minimum, and we give the network the state (1, 0, 0, 0, 1) it will converge to (1, 0, 1, 0, 1). Learning: the weights are learnt, so as to!shape" those local minima. The network will learnt to converge to learnt state even if it is given only part of the state w ij = 1 N k=n k=1 ξ k i ξ k j

Attractor paradigm for persistent activity The Ring Model (1) Since the 1970s it has been proposed that delay activity patterns can be theoretically described by!

τ r dv(θ) dt = v(θ) + [ h(θ) + The Ring Model (2) N neurons, with preferred angle, θ i,evenly distributed between π/2 and π/2 Neurons receive thalamic inputs h.

6 Attractor paradigm for persistent activity The Ring Model (1) Since the 1970s it has been proposed that delay activity patterns can be theoretically described by!dynamical attractors" Recently, a great effort to build biophysically plausible model of sustained activity / attractor dynamics for memory. τ r dv(θ) dt = v(θ) + [ h(θ) + The Ring Model (2) N neurons, with preferred angle, θ i,evenly distributed between π/2 and π/2 Neurons receive thalamic inputs h. + recurrent connections, with excitatory weights between nearby cells and inhibitory weights between cells that are further apart (mexican-hat profile) -('./*!" "!!"!$""!$!"!#""!!" "!" %&'()*+*'%),! π/2 π/2 dθ ( λ0 + λ 1 cos(2(θ θ )) ) ] v(θ ) π + (7.36) h is input, can be tuned (Hubel Wiesel scenario) or very broadly tuned. h(θ) = c[1 ɛ + ɛ cos(2θ)] The Ring Model (3) The steady-state can be solved analytically. Model analyzed like a physical system. 01!2 "#! "#' "#& "#% "#$ " /!!" "!" ()*+,-.-*(,/! Model achieves i) orientation selectivity; ii) contrast invariance of tuning, even if input is very broad. The width of orientation selectivity depends on the shape of the mexican-hat, but is independent of the width of the input. Symmetry breaking /Attractor dynamics. " $ " % /

The Ring Model (4) A B C 80 h (Hz) 30 20 10 0-40 -20 0 20 40 θ (deg) v (Hz) 60 40 20 0-40 -20 0 20 40 θ (deg) firing rate (Hz) 80 60 40 20 80% 40% 20% 10% 0 180 200 220 240 Θ (deg) Attractor Networks

7 The Ring Model (4) A B C 80 h (Hz) θ (deg) v (Hz) θ (deg) firing rate (Hz) % 40% 20% 10% Θ (deg) Attractor Networks Attractor network : a network of neurons, usually recurrently connected, whose time dynamics settle to a stable pattern. That pattern may be stationary (fixed points), time-varying (e.g. cyclic), or even stochastic-looking (e.g., chaotic). The particular pattern a network settles to is called its!attractor". The ring model is called a line (or ring) attractor network. ts stable states are also sometimes referred to as!bump attractors". Figure 7.10: The effect of contrast on orientation tuning. A) The feedforward input as a function of preferred orientation. The four curves, from top to bottom, correspond to contrasts of 80%, 40%, 20%, and 10%. B) The output firing rates in response to different levels of contrast as a function of orientation preference. These are also the response tuning curves of a single neuron with preferred orientation zero. As in A, the four curves, from top to bottom, correspond to contrasts of 80%, 40%, 20%, and 10%. The recurrent model had λ 0 = 7.3, λ 1 = 11, A = 40 Hz, and ɛ = 0.1. C) Tuning curves measure experimentally at four contrast levels as indicated in the legend. (C adapted from Sompolinsky and Shapley, 1997; based on data from Sclar and Freeman, 1982.) Point Attractor Line Attractor The Ring Model (5): Sustained Activity f recurrent connections are strong enough, the pattern of population activity once established can become independent of the structure of the input. t can persists when input is removed. A model of working memory? Anatomical organization of PFC resembles a recurrent network Biophysical realistic computational modeling has shown that such recurrent networks can give rise to location-specific, persistent discharges (Compte et al 2000, Gutkin et al 2000, Tegner et al 2002, Renart et al 2003a, Wang et al 2004) Fig. 4. Schematic diagram illustrating the pattern of connections between prefrontal neurons in the superficial layers. The figure summarizes results of anatomical tracer injection experiments and retrograde labeling. From Kritzer and Goldman- Rakic (1995), with permission.

8 (a) 360 Modeling studies show that stability is an issue in such network. Strong recurrent inhibition is needed to prevent runaway excitation and maintain specificity Models are also challenged by accounting for spontaneous activity in addition to memory state Oscillations can destabilize the memory activity. Working memory is found to be particularly stable when excitatory reverberations are characterized by a fairly slow time course, e.g. when synaptic transmission is mediated by NMA receptors (prediction) Neuron label ( O ) 0 C R 20 Hz 1 s (b) (i) (ii) 20 Hz 20 Hz elay activity Cue location ( O ) Cue location ( O ) [Compte, Brunel, Goldman-Rakic and Wang, 2000] Network of ~2500 integrate and fire neurons, mexican hat connectivity, NMA excitation. Fig. 6. Stability of persistent activity as a function of the AMPA:NMA ratio at the recurrent excitatory synapses. A, Temporal course of the average firing rate across a subpopulation of cells selective to the presentated transient input, for different levels of the AMPA:NMA ratio. As the ratio is increased, oscillations of a progressively larger amplitude develop during the delay period, which eventually destabilize the persistent activity state. E, Snapshot of the activity of the network in (C) between 3 and 3.5 seconds. Top, Average network activity. Bottom, ntracellular voltage trace of a single neuron. nset, Power spectrum of the average activity of the network, showing a peak in the γ (40 Hz) frequency range. Persistent activity is stable even in the presence of synchronous oscillations. Adapted with permission from Renart, Brunel, and others (2003). [Renart, Brunel, Wang, 2003] Box 2. Outstanding questions Recent theoretical models have raised several neurophysiological questions that can be investigated experimentally. Answers to these questions will help to elucidate the mechanisms of neural persistent activity. What is the minimum anatomical substrate of a reverberatory circuit capable of persistent neural activity? s persistent activity primarily sustained by synaptic reverberation, or by bistable dynamics of single neurons? What is the NMA:AMPA ratio at recurrent synapses of association cortices, especially in the prefrontal cortex? How does this ratio depend on the frequency of repetitive stimulation and on neuromodulation? What are the negative feedback mechanisms responsible for the rate control in a working memory network? s delay period activity asynchronous between neurons, or does it display partial network synchrony and coherent oscillations? s delay period activity more sensitive to NMAR antagonists compared with AMPAR antagonists? oes persistent activity disappear in an abrupt fashion, with a graded block of NMAR and AMPAR channels, as predicted by the attractor model? How significant are drifts of persistent activity during working memory? Are drifts random or systematic over trials? What are the biological mechanisms underlying the robustness of a memory network with a continuum of persistent activity patterns?

But cellular mechanisms should not be forgotten.

-dependent plateau potential. This leads to sustained firing at a constant rate > 13 min independent of synaptic transmission.

(s sustained activity helpful for storage of memory?) Ardid et al. Microcircuit Model of Attentional Processing Ardid et al.

Scheme of the loop architecture (red is excitation, and blue is inhibition).

9 But cellular mechanisms should not be forgotten... Lots of interesting questions [Egorov et al, Nature, 2002] Layer 5 of EC in vitro, intracellular depolarization + bath application of the AChreceptor agonist leads to a Ca2+ -dependent plateau potential. This leads to sustained firing at a constant rate > 13 min independent of synaptic transmission. Level of activity can be increased or decreased using repeated inputs. How are these attractors learnt? What is the relation with Attention? What is the relation with Long-term Memory? (s sustained activity helpful for storage of memory?) Ardid et al. Microcircuit Model of Attentional Processing Ardid et al. Microcircuit Model of Attentional Processing Could attractors be suited for remembering learned stimuli while such a system could help maintaining new stimuli? Figure 1. Scheme of the loop architecture (red is excitation, and blue is inhibition). Two kinds of motion stimuli are considered (random-dot patterns; yellow arrows indicate signal motion directions): single (left) and transparent (right) motion. WM, Working memory. Ardid, Wang and Compte 2007 A related problem: spontaneous activity Where does it come from? How is it maintained? How does it!move"? Are these!attractor states"? s it structured? Why is it there? (any functional advantages?) s it noise? s it the brain trying to!predict" the input? Arieli et al 1997; Tsodyks et al, 1999; Fiser et al, Nature, 2004 evoked (horizontal orientation) spontaneous (one frame)

Dopamine modulation in a basal ganglio-cortical network implements saliency-based gating of working memory

Dopamine modulation in a basal ganglio-cortical network implements saliency-based gating of working memory Aaron J. Gruber 1,2, Peter Dayan 3, Boris S. Gutkin 3, and Sara A. Solla 2,4 Biomedical Engineering