A General Theory of the Brain Based on the Biophysics of Prediction Christopher D. Fiorillo KAIST Daejeon, Korea April 7, 2016
Why have we not understood the brain? Scientists categorize the world into people and things, and they analyze them very differently Fiorillo, 2012 Thing (Matter)? Person (Mind)
Dualism in Science Physical Sciences Psychology, Computation Matter, Energy Causation Mechanism Object Mind, Perception Belief, Prediction, Inference Information, Knowledge, Intelligence Observer
The Standard Approach of Science: The Inputs and Outputs of Things
The Conventional Application of Information Theory to Neurons What do spikes tell US about sensory input? The answer varies depending on the prior knowledge of the observer. There could be countless external observers. There is only one neuron generating the spikes. What does a spike mean to that neuron?
How do we understand a person? Theory of Mind Different people have different information We try to see the world from another person s perspective Can this be objective and scientific? Yes.
The Probability Theory of Jaynes A unified mathematics of information and logic (knowledge and reason) Information is Subjective. Logic is Objective We can objectively describe the subjective knowledge of a physical observer with probabilities.
Physics: What are its inputs and outputs? Psychology: What does he know about the world?
General Theory 1. Goal: Minimize Uncertainty (of Neurons about Future Value) 2. Input-Output Relations A. Sensory neurons: prediction error (predictive homeostasis) B. Motor neurons: predictive control 3. Learning A. Hebbian rules to maximize prediction error, and predictive control of value B. anti-hebbian rules to minimize prediction error and promote predictive homeostasis C. Associative learning regulates voltage-sensitive ion channels as well as synapses Fiorillo, 2008; 2012; Fiorillo et al., 2014
Membrane Voltage in a Visual Neuron During a Movie Wang et al., 2007 What is the meaning of membrane voltage and spikes from the Neuron s Point of View? What is the neural code?
Membrane Voltage Spike (Action Potential) 20 mv 5 ms Excitatory Postsynaptic Potential (EPSP)
Known Reference (Old, Prior, Prediction, Expectation) Unknown to Measure (New sensory evidence)
The Physics of Measurement Known Reference (Old, Prior, Prediction, Expectation) Prediction Error Unknown to Measure (New sensory evidence)
Measuring Synaptic Excitation: Prediction Error Inhibitory Postsynaptic Conductance (IPSG) Excitatory Postsynaptic Conductance (EPSG) Excitatory Postsynaptic Potential (EPSP)
Perfect Balance (Homeostasis) Perfect Prediction Prediction Error = Zero An Ideal that Never Happens Spike Threshold Theory: Spike Threshold is The Homeostatic Ideal of Perfect Prediction
Can the Theory Predict Neural Properties? 1. Spike Statistics 2. Synaptic Inhibition 3. T-type calcium channels in sensory neurons (homeostasis) 4. T-type calcium channels in motor neurons (decisionmaking)
Spikes Appear Random and Noisy Ruyter van Steveninck et al., Science 1997
Testing the Theory: Spike Statistics Inaccurate Prediction Accurate Prediction Inaccurate Prediction No Spike Spike or No Spike Spike
Testing the Theory: Spike Statistics Input Neuron s Internal Point of View Scientists External Point of View No Spike Spike or No Spike Spike Output
Spikes are Variable When Predictions are Accurate Ruyter van Steveninck et al., Science 1997
Can the Theory Predict Neural Properties? 1. Spike Statistics 2. Synaptic Inhibition
Using Theory to Predict the Properties of Synaptic Inhibition IPSG EPSG 1 ms 1. How strong should synaptic inhibition be? 2. How long should it last?
EPSG EPSP 10 ns 5 mv 2 ms Measuring Distance from Perfect Balance in Computer Simulations EPSG 1 EPSG 2 Residual 1 Residual 2 Mean Squared Residual Minimize to Find Optimal Synaptic Inhibition
50 Hz 200 Hz 30 ns 30 ms 60 ns 10 mv
500 100 Hz Optimal IPSG for EPSG at 100 Hz 300 IPSG Decay Time Constant (ms) 30 20 10 0 0 1 2 3 Amplitude I/E 100 1 2 3 I/E 30 25 20 15 10 5 Mean Squared Residual (100 ns^2)
anti-hebbian Learning of Optimal Synaptic Inhibition 10 ns 10 ms
Kim and Fiorillo, in preparation Optimal Synaptic Inhibition 800 Hz 30 ns 5 ms 5 Hz 50 Hz
4 3 Theory Minimize MSR Learning Experiment Amplitude of Inhibition (I/E) 2 1 Kim and Fiorillo, in preparation 0 10 0 10 1 10 2 10 Firing Rate (Hz)
40 Theory Experiment 30 Decay Time Constant (ms) 20 10 Kim and Fiorillo, in preparation 0 10 0 10 1 10 2 10 3 Firing Rate (Hz)
Can the Theory Predict Neural Properties? 1. Spike Statistics 2. Synaptic Inhibition 3. T-type calcium channels in sensory neurons (homeostasis) 4. T-type calcium channels in motor neurons (decisionmaking)
4 Types of Input (Ion Channel) Old Evidence (Homeostatic) New Evidence
Hong et al., 2014 Mimicking Natural Conditions In the brain (LGN) while watching a movie Mimicry in a brain slice
T-type Channels Promote Homeostasis in Sensory Neurons T-type Channels Homeostatic Amplification (not too weak or too strong) Hong et al., 2014
T-type Channels Restore Homeostasis 1, 2, and 4 Artificial EPSPs mv 30 40 50 60 50 ms 800 ms 70 80 30 ms Spike Count 2 1-65 -80 mv 2 1-65 -80 mv Nickel (n = 12) control (n=26) Nickel (n = 12) 0 n = 26 0 400 800 Time (ms) 0 0 400 800 Time (ms) 1 2 3 Input (EPSGs) Hong et al., 2014 4
Can the Theory Predict Neural Properties? 1. Spike Statistics 2. Synaptic Inhibition 3. T-type calcium channels in sensory neurons (homeostasis) 4. T-type calcium channels in motor neurons (decisionmaking)
Causation in Decisions Sensory Evidence Perception Prior Evidence Decision Action
Sensory Neurons Input Output Spike Output (Perception) Graded Motor Neurons Input Output Spike Output (Action) All-or-None Sensory Input (Evidence) Sensory Input (Evidence) T-type Calcium Channels
Kim et al., 2015 Hpc A sm MD AV AM mt LD VA 1 mm 200-2.04 mm -3.48-4.68-5.28 VL Rt ic fi eml fr VM Hpc 150 PoM -74-70 -66 Voltage (mv) LP LGN MD PoM VB eml ic Rt AP PoM B C 250 40 D LP LGN 40 Resistance (MΩ) Comparison of 8 Thalamic Nuclei MGd MGv MD VL VB Neurons (%) 30 20 10 LP VB LGN str VG 0-85 -75-65 -55 Voltage (mv) Hpc Neurons (%) opt 30 20 10 MGd AP LP MGv LGN bsc 0 100 300 500 Resistance (MΩ)
Kim et al., 2015 TtCC Cause Cause All-or-None Bursts in Motor but not Sensory Thalamus Motor Sensory
TtCC Cause Bursts in Motor but not Sensory Thalamus A Threshold (mv) C -65-70 4 r = 0.38 p = 10-6 -65-70 3 B CCSW DCSW aepsg r = 0.39 r = 0.22 p = 10-5 2.0 p = 0.04 Threshold (aepsgs) D 1.5 1.0 3 Protocols 8 Thalamic Nuclei 450 Neurons flts spikes E 2 0 100 r = -0.59 p = 10-16 2 1 100 0.5 r = -0.43 r = -0.36 p = 10-6 p = 0.001 0 100 Bursts (%) 50 0 r = -0.83 p = 0.01 50 0 r = -0.91 p = 0.03 Motor Sensory 50 0 r = -0.99 p = 0.07 Kim et al., 2015
The Theory Correctly Predicted Properties of Neurons 1. Spike Statistics 2. Synaptic Inhibition 3. T-type calcium channels in sensory neurons (homeostasis) 4. T-type calcium channels in motor neurons (decision-making)
Acknowledgments Jaekyung Kim Synaptic Inhibition Su Z. Hong T-type Calcium Channels in LGN Haram Kim T-type Calcium Channels in Sensory vs. Motor Thalamus