APRIL 1

Question 1

paper: “a primer on the use of _______ _______ to investigate affective states, affective disorders and animal welfare in _____ _____ ______”

Answer 1

a primer on the use of COMPUTATIONAL MODELLING to investigate affective states, affective disorders and animal welfare in NON-HUMAN ANIMALS”

Question 2

computational psychiatry

Answer 2

uses computational modelling to capture the UNOBSERVABLE UNDERLYING PROCESSES that drive observable changes in behaviour

Question 3

affect is a latent construct

Answer 3

affect = hard to measure accurately

because it’s a LATENT CONSTRUCT

Question 4

latent construct

Answer 4

a theoretical, unobservable trait, condition or concept

can ONLY BE INFERRED or MEASURED INDIRECTLY

using OBSERVABLE INDICATORS

Question 5

fear example as a latent construct

Answer 5

fear is the latent construct

observable indicators:
- increased HR
- sweaty palms
- self-reported fear

Question 6

depression-like state as a latent construct

Answer 6

depression-like state is the latent construct

observable indicators:
- sucrose preference
- social avoidance
- HPA-axis function

Question 7

connection to A&A framework

Answer 7

stimuli (inputs) lead to this CENTRAL EMOTIONAL STATE (CES) which leads to outputs

the CES is unobservable, in between the inputs and outputs

just like the latent construct

Question 8

reasons why relationship between objective, observable measures and affective state is complex

Answer 8

1. potentially MULTIPLE UNDERLYING PROCESSES are involved (but integrated)
2. behaviour can be influenced by MANY FACTORS BEYOND the affective state (ie. experimental setup - if you put rat in field, the texture of the field may elicit certain behaviours)
3. observable measures can have COMPLEX RELATIONSHIPS to underlying states (ie. increased locomotor activity in open-field can be slower habituation - INCREASED ANXIETY - or increased willingness to explore - REDUCED ANXIETY)
4. INDIVIDUAL DIFFERENCES can make observations variable

Question 9

what’s the problem with the output variables that most experiments measure?

Answer 9

most experiments measure ONLY ONE OUTPUT VARIABLE

ie. running speed

this OVERLOOKS the potential contribution of many diff variables

Question 10

what does computational modelling do to the various influences on observable outcomes?

Answer 10

UNMIXES them

provides FORMAL DESCRIPTIONS of how latent constructs intermediate between experience and action

Question 11

3 areas under computational modelling

Answer 11

1. data analysis
2. mathematical modelling
3. information processing modelling

Question 12

data analysis

Answer 12

data-driven, quantitative methods for analyzing data

ie. MACHINE LEARNING ALGORITHMS for video tracking & image classification

applying ML to find patterns/structure within data

Question 13

data analysis example

Answer 13

building a CLASSIFIER to identify MOUSE ‘EMOTION’ from videos of mouse faces as mice experience diff outcomes

ie. sucrose, tailshock

Question 14

data-driven approaches (data analysis) capture _________ relationships in the data, and remain _________

Answer 14

capture STATISTICAL relationships

they remain DESCRIPTIVE

Question 15

data driven approaches don’t inform us about ________

Answer 15

mechanism

Question 16

mathematical modelling

Answer 16

descriptions of MECHANISMS or SYSTEMS using MATHEMATICAL FORMULATIONS

Question 17

example of mathematical modelling

Answer 17

Hodgin-Huxley model that explains how action potentials occur in neurons

set of non-linear differential equations that approximate electrical mechanisms of action potential generation

Question 18

information processing modelling

Answer 18

characterizations of brain in terms of COMPUTATIONAL PROBLEM it solves

Question 19

example of information processing modelling

Answer 19

modelling the brain as MAXIMIZING REWARD using REINFORCEMENT LEARNING to understand decision making

Question 20

key diff between data analysis and mathematical modelling/information processing modelling

Answer 20

data analysis DESCRIBES the data

mathematical modelling and information processing modelling seek to UNDERSTAND the data

Question 21

goal of mathematical modelling and information processing modelling

Answer 21

to understand the RELATIONSHIPS WITHIN THE DATA by applying a formal characterization of the UNDERLYING PROCESSES through which these data might have arisen

Question 22

computational models are generative - how?

Answer 22

1. aim to understand how data was generated by MATHEMATICALLY SPECIFYING HYPOTHESES about latent processes that generated data

Question 23

what should a useful computational model be able to do?

Answer 23

1. REPRODUCE the data
2. PREDICT data from other experiments

Question 24

how can we test the strength of a computational model?

Answer 24

by comparing the SIMULATED DATA FROM THE MODEL to the ACTUAL DATA

Question 25

T/F: models can generate new predictions and hypotheses which can be tested experimentally

Answer 25

true

Question 26

computational models are formal specifications of the relationship between what?

Answer 26

inputs and outputs this makes them POWERFUL in STUDYING LATENT CONSTRUCTS like AFFECT

Question 27

emotion states can operate as 'black boxes' - what can computational modelling do?

Answer 27

it can open black boxes

Question 28

2 types of questions computational models can be used to answer?

Answer 28

1. what's the BEST EXPLANATION for the data? 2. what are the values of the FREE PARAMETERS of the same model that BEST FIT THE DATA of each individual? (parameter estimation)

Question 29

comparing between models versus comparing within models

Answer 29

COMPARING BETWEEN: asking what's the best explanation for the data? COMPARING WITHIN: asks what values of the free parameters of the same model best fit the data of each individual

Question 30

what does computational psychiatry do?

Answer 30

uses computational modelling to CHARACTERIZE and UNDERSTAND psychiatric disorders such as MDD and anxiety in a more MECHANISTIC manner

Question 31

MDD may use different kinds of _______ in decision making

Answer 31

models

Question 32

what models may people with MDD use, in contrast with healthy controls?

Answer 32

MDD: 'model-free' (stimulus-response, habit based, reflexive) HEALTHY CONTROL: 'model-based' (goal directed, flexible, reflective)

Question 33

MDD may result in different weighting of ________

Answer 33

effort (WEIGH EFFORT MORE HIGHLY) people with MDD weigh effort costs more highly than non-depressed control antidepressant drugs reduce effort costs

Question 34

changes in learning in depression

Answer 34

(suggested by computational modelling) 1. reduced ability to learn about NEGATIVE OUTCOMES could make AVERSIVE EVENTS MORE SURPRISING 2. increased surprise INCREASES 'PREDICTION ERROR' - this has been linked to NEGATIVE MOOD 3. prediction errors CORRELATE with MOOD

Question 35

how do prediction errors correlate with mood?

Answer 35

1. outcomes that are BETTER THAN EXPECTED (positive prediction error) associate with more POSITIVE mood 2. outcomes that are WORSE THAN EXPECTED (negative prediction error) associate with more NEGATIVE mood

Question 36

prediction error and mood in naturalistic datasets

Answer 36

SUNSHINE PREDICTION ERROR (as measured by diff in weather forecast and actual weather) MOOD (assessed at city level by scraping Twitter) PER CAPITA LOTTERY GAMBLING (state lottery gambling records)

Question 37

variability in findings reflect ________ in MDD (and provide examples)

Answer 37

heterogeneity 1. some studies report REDUCED REWARD SENSITIVITY in anhedonic depression 2. other studies report FASTER LEARNING ABOUT PUNISHING OUTCOMES and slower learning about rewarding outcome 3. other studies FAIL TO FIND DIFFERENCES in reward learning rates/sensitivity

Question 38

while setup of experiments and specific behaviours across humans and non-humans will substantially vary...

Answer 38

the underlying COMPUTATIONS that humans and non-humans perform may be conserved and computational models can capture these commonalities

Question 39

what do reinforcement learning models of decision making do?

Answer 39

1. estimate the VALUE of a STATE or ACTION to GUIDE CHOICES about which actions should be taken to MAXIMIZE REWARD (concept of prediction error = foundational)

Question 40

Rescorla-Wagner learning rule

Answer 40

describes how animals LEARN THE VALUE OF A STIMULUS or ACTION by experience change in predictive value after an outcome depends on HOW SURPRISING the outcome is (prediction error) prediction error is SCALED by a LEARNING rate

Question 41

what is prediction error scaled by?

Answer 41

learning rate HIGHER learning rate = faster change LOWER learning rate = slower change

Question 42

'optimal' parameters depend on what?

Answer 42

the context ENVIRON 1: systematically, rapidly fluctuating learning rates ^ favours HIGH learning rates - because need ability to adapt to change in reward ENVIRON 2: noisy but stable overall average reward rates ^ favour MODERATE learning rates - because it's so noisy and unclear

Question 43

noisy but stable overall average reward rates call for what type of learning rate?

Answer 43

MODERATE because high learning rates CHASE THE NOISE

Question 44

what does choice depend on?

Answer 44

1. PREDICTED VALUE of an action 2. VALUE of alternatives

Question 45

should individuals always choose the action with highest predicted value?

Answer 45

no - they should explore other options exploration = important for OBTAINING INFO about other options (which could be of higher value)

Question 46

T/F: predicted value directly translates into the probability of actions

Answer 46

false

Question 47

softmax function

Answer 47

calculates action probabilities from action values includes a BETA PARAMETER that CAPTURES RANDOMNESS OF CHOICE

Question 48

higher beta results in...

Answer 48

results in choices that CLOSELY MIRROR ACTION VALUES

Question 49

lower beta results in...

Answer 49

choices that are MORE RANDOM

Question 50

beta parameter can also be thought of as measure of...

Answer 50

exploration versus exploitation (higher beta = less exploration, closer to action values) (lower beta = more exploration, more randomness)

Question 51

positive RPE induces what?

Answer 51

positive valence

Question 52

negative RPE induces what?

Answer 52

negative valence

Question 53

role of beta parameter in modulating extent to which values influence actions can be thought of as parallel to what?

Answer 53

to how ANHEDONIA in MDD impairs ability to use info about REWARD in decision making some evidence that B parameter is LOWER in people with MDD (associated with MORE EXPLORATION and MORE RANDOMNESS)

Question 54

B parameter in people with MDD may be...

Answer 54

lower reflects more exploration/randomness in choice

Question 55

judgment bias as marker of affective state

Answer 55

judgment bias tasks leverage PERCEPTUAL AMBIGUITY to probe OPTIMISM and PESSIMISM experimental manipulations inducing NEG affective state increase PESSIMISTIC BIASES manipulations that induce POS affective state increase OPTIMISTIC BIASES

Question 56

how can computational modelling be applied to study judgment bias?

Answer 56

diff kinds of models can be fit to these tasks to understand more about the NATURE of the UNDERLYING PROCESSES diff models reveal distinct aspects ie. DRIFT-DIFFUSION MODELING

Question 57

drift diffusion models

Answer 57

developed in PERCEPTUAL DECISION MAKING fields

Question 58

DDMs applied to judgment bias tasks - what can it reveal?

Answer 58

use DECISION TIME and CHOICE ACCURACY to estimate a threshold for decision-making for diff choices can REVEAL how DIFF MANIPULATIONS MODULATE the THRESHOLD for certain kinds of decisions ie. can show that a NEGATIVE AFFECTIVE STATE INCREASES the THRESHOLD for an optimistic decision

Question 59

what can formally show that a negative affective state increases the threshold for an optimistic decision?

Answer 59

drift diffusion model

Question 60

Bayesian decision-theoretic modelling in judgment bias

Answer 60

another computational framework estimates EXPECTED VALUE of EACH CHOICE given the ANIMAL'S PERCEPTION of the stimulus reveals AFFECT-INDUCED CHANGES in OUTCOME VALUATION ie. mild stress increased valuation of reward SHOWS THAT JUDGMENT BIASES AREN'T ONLY DRIVEN BY OPTIMISM/PESSIMISM, but also are influenced by CHANGES IN HOW OUTCOMES ARE EVALUATED (models can dissociate impact of related but distinct processes)

Question 61

without computational modelling, Bayesian decision theoretic and DDM show that...

Answer 61

conventional descriptions of judgment bias tasks may be OVERLY SIMPLISTIC multiple sources of data can be integrated into computational models to provide a MORE SENSITIVE and NUANCED UNDERSTANDING ie. effects of optimism pessimism AND outcome valuation

Question 62

computational modelling can disentangle interacting processes - foul smelling room study

Answer 62

foul smelling room used to induce NEGATIVE AFFECTIVE STATES (versus normal room) then tested judgment bias NO EFFECT WAS FOUND IN STANDARD METRICS but Bayesian modelling revealed that the manipulation made participants WEIGH LOSSES MORE HEAVILY which led to MORE OPTIMISTIC DECISION MAKING this way, the effect on VALUATION likely CANCELLED OUT the effect on judgment bias (this is what led to the observed non-effect)