Cholinergic Modulation of the Hippocampus Computational Models of - PowerPoint PPT Presentation

Cholinergic Modulation of the Hippocampus Computational Models of Neural Systems Lecture 2.5 David S. Touretzky September, 2007

A Theory of Hippocampus ● Suppose CA1 is a hetero- associator that learns: – to mimic EC patterns, and – to map CA3 patterns to learned EC patterns ● Imagine a partial/noisy pattern in EC triggering a partial/noisy response in CA3, cleaned up by auto- association in CA3 recurrent collaterals ● CA1 could use the EC What happens if recall response to call up the isn't turned off during complete, correct EC pattern learning? 2 Computational Models of Neural Systems 09/20/07

Acetylcholine Effects (1) Acetylcholine (ACh) has a variety of effects on HC: ● Suppresses synaptic transmission in CA1: – Mostly at Schaffer collaterals in stratum radiatum – Less so for perforant path input in stratum lacunosum-moleculare patch clamp recording 3 Computational Models of Neural Systems 09/20/07

Effect of Carbachol Experiment 1 ● Carbachol is a cholinergic agonist. ● Can use carbachol to test the effects of 46.0% suppression 90.7% suppression ACh. ● It only activates metabotropic ACh Experiment 2 receptors. ● Brain slice recording experiments show that carbachol suppresses synaptic 54.6% suppression 87.6% suppression transmission in CA1. 4 Computational Models of Neural Systems 09/20/07

Effect of Atropine ● Atropine affects muscarinic-type ACh receptors, not nicotinic type. ● Blocks the suppression of synaptic transmission by carbachol. ● Therefore, cholinergic suppression in s. rad. and s. l.-m. is by muscarinic ACh receptors. same same 5 Computational Models of Neural Systems 09/20/07

Summary of Blockade Experiments 6 Computational Models of Neural Systems 09/20/07

Acetylcholine Effects (2) Acetylcholine also: ● Reduces neuronal adaptation in CA1 by suppressing voltage and Ca 2+ dependent potassium currents. – This keeps the cells excitable. ● Enhances synaptic modification in CA1 – possibly by affecting NMDA currents. ● Activates inhibitory interneurons – but decreases inhibitory synaptic transmission. 7 Computational Models of Neural Systems 09/20/07

Hasselmo's Model: Block Diagram EC pp ACh Sch CA1 CA3 fimbria/fornix Medial Septum 8 Computational Models of Neural Systems 09/20/07

Hasselmo's Model 9 Computational Models of Neural Systems 09/20/07

Initial CA1 Activation Function n  L ij − EC H ij  ⋅ g  EC a j  t  = ∑ a i  t  j = 1 n  R ik − CA3 H ik  ⋅ g  CA3 a k  t   ∑ k = 1 n − ∑ CA1 H il ⋅ g  CA1 a l  t  l = 1 a i  t  is activation of unit i at time t g(x) g  x  is a threshold function: max  x − 0.4, 0  L ij is feedforward synaptic strength for s. lacunosum (EC input) R ij is feedforward synaptic strength for s. radiatum (CA3 input) xx H i _ is feedforward or feedback inhibition of CA1 from layer xx 10 Computational Models of Neural Systems 09/20/07

S. Radiatum Learning Rule ● Note: the only learning in this model is in R ik , the weights on the CA3 → CA1 connections. T wo factors: – Linear potentiation when pre- and post-synaptic cells are simultaneously active. – Exponential decay whenever the postsynaptic cell is active. R ik  t  1  = R ik  t   ⋅ [  CA1 a i  t −  ⋅ g  CA3 a k  t  −  CA1 a i  t − ⋅ R ik  t  ]  is the synaptic modification threshold for LTP to occur.  is the overall learning rate.  is the synaptic decay rate. 11 Computational Models of Neural Systems 09/20/07

Learning Rule: Hebbian Facilitation Plus Synaptic Decay Presynaptic 0 1 0 Postsynaptic ↓ ↑ 1 ↓ 12 Computational Models of Neural Systems 09/20/07

Exponential Weight Decay dx = − x dt x  t  1  = x  t  −  x  t  = x  t  ⋅  1 − x  t  ⋅  1 − n x  t  n  = Example:  = 0.04 x  t  1  = 0.96 x  t  13 Computational Models of Neural Systems 09/20/07

Control of Cholinergic Modulation ● Cholinergic modulation Ψ was controlled by the amount of activity in CA1:  = [ 1  exp  ∑ g  CA1 a i − ] − 1 n i = 1  is a gain parameter  is a threshold value ● This is an inverted sigmoid activation function of form 1 - 1/(1+exp(x)): – With no CA1 activity, Ψ is close to 1. – With maximal CA1 activity, Ψ is close to 0. 14 Computational Models of Neural Systems 09/20/07

ACh Modulation of Recall n   1 − C L  L ij − 1 − C H  H ij  ⋅ g  EC a j  t  ∑ a i  t  = j = 1 n   1 − C R  R ij − 1 − C H  H ik  ⋅ g  CA3 a k  t   ∑ k = 1 n − ∑  1 − C H  H il ⋅ g  CA1 a l  t  l = 1 C L ,C R ,C H are coefficients of ACh modulation. 15 Computational Models of Neural Systems 09/20/07

ACh Modulation of Learning R ik  t  1  = ⋅ [ 1 − C   1 −] ⋅ [  CA1 a i  t  −  1 − C   ⋅ g  CA3 a i  t  ⋅ R ik  t  ] −  CA1 a i  t  −  1 − C    R ik  t  C  ,C  are coefficients of ACh modulation. Note: output threshold  in g ⋅ is also reduced by  1 − C  . This simulates ACh suppression of neuronal adapation. 16 Computational Models of Neural Systems 09/20/07

What Do These T erms Look Like?  1 − C   1 − 1 − C   1 − 17 Computational Models of Neural Systems 09/20/07

T rain T est Recovery from weight decay caused by recall of pattern 2. 18 Computational Models of Neural Systems 09/20/07

Weak suppression in s. rad. and none in s. l.-m. Result: unwanted learning causes memory interference. Strong suppression in s. rad. and also in s. l.-m. Result: retrieval fails. 19 Computational Models of Neural Systems 09/20/07

Larger Simulation Learned 5 patterns. After learning, CA3 input is sufficient to recall the patterns. 20 Computational Models of Neural Systems 09/20/07

Memory Performance A is CA1 output, B is corresponding EC pattern (teacher). For perfect memory, A = B . Recall that A ⋅ B = ∥ A ∥ ∥ B ∥ cos  A ⋅ B = cos  = 1 for perfect memory Normalized dot product: ∥ A ∥ ∥ B ∥ C i is some other training pattern A ⋅ C i M A ⋅ B 1 M ∑ Performance P = − ∥ A ∥ ∥ B ∥ ∥ A ∥ ∥ C i ∥ i = 1 Average overlap with all stored patterns. 21 Computational Models of Neural Systems 09/20/07

C L vs. C R Parameter Space ● Performance is plotted on z axis. ● Grey line shows C L = C R . ● White line shows dose- response plot from carbachol experiment. 22 Computational Models of Neural Systems 09/20/07

Comparison with Marr Model ● Distinguishing learning vs. recall: – Marr assumed recall would always use small subpatterns, perhaps one tenth the size of a full memory pattern. Not enough activity to trigger learning. – Hasselmo assumes that unfamiliar patterns only weakly activate CA1, and that leads to elevated ACh which enhances learning. ● Input patterns: – Marr assumes inputs are sparse and random, so nearly orthogonal. – Hasselmo's simulations use small vectors so there is substantial overlap between patterns. Uses ACh modulation to suppress interference. 23 Computational Models of Neural Systems 09/20/07

A Model of Episodic Memory 24 Computational Models of Neural Systems 09/20/07

ACh Prevents Overlap w/Previously Stored Memories from Interfering with Learning 25 Computational Models of Neural Systems 09/20/07

Simulation of ACh Effects 10 input neurons 2 inhibitory neurons 1 ACh signal 26 Computational Models of Neural Systems 09/20/07

Episodic Memory Simulation ● Each layer contains both Context and Item units. ● T rain on list of 5 patterns. ● During recall, supply ony the context. ● An adaptation process causes recalled items to eventually fade so that another item can become active. 27 Computational Models of Neural Systems 09/20/07

“Consolidation” T rain model on set of 6 patterns. During consolidation, use free recall to train slow-learning recurrent connections in EC layer IV. After training, a partial input pattern (not shown) recalls the full pattern in layer cortex. poor good 28 Computational Models of Neural Systems 09/20/07

Summary ● Unwanted recall of old patterns can interfere with storing new ones. ● The hippocampus must have some way of preventing this interference. ● Cholinergic modulation in CA1 (and also CA3) affects both synaptic transmission and L TP . ● Acetylcholine may serve as the “novelty” signal: – Unfamiliar patterns → high ACh → learning – Familiar patterns → low ACh → recall ● CA1 might serve as a comparator of current EC input with reconstructed input from CA3 projection to determine pattern familiarity . 29 Computational Models of Neural Systems 09/20/07