## Introduction

### Computational Mechanisms of Biased Information Processing

### Volatility Modulates Learning Rates

### Augmenting Cognitive Training with tDCS

### Neural Substrates of Learning Rates

### Current Study

## Methods

### Participants

### Study Procedure

### IBLT Training

### tDCS Protocol

### Facial Expression Recognition Task

### Questionnaire Measures

### Computational Modelling

_{(i+1)}is the estimated win outcome probability for the i + 1st trial, rwin

_{(i)}is the estimated outcome for the ith trial, αwin represents the learning rate for win outcomes, and ε

_{win(i)}indicates the prediction error on the ith trial. Prediction error is calculated as the predicted outcome value minus the actual outcome value. At the start of each block, rwin was initialised at 0.5 because participants could not have prior expectations about which shape was most likely to be associated with a win outcome. rloss

_{(i+1)}, rloss

_{(i)}, aloss and ε

_{loss(i)}indicate the same variables for the loss outcome. Next, estimated outcome probabilities were transformed into a single choice probability using a softmax function:

_{(choice=A(i))}is the probability of choosing shape A in trial i, with β representing the inverse decision temperature and t reflecting an added parameter used to estimate a general tendency to select one of the options over the other. The inverse temperature indicates the degree to which the expected values are used to determine choice for a particular shape. Learning rates and β-values were calculated separately for each task block and participant. This was achieved by calculating the full joint posterior probability of the parameters given participants’ choices, deriving the expected value of each parameter from their marginalised probability distributions (Behrens et al. 2007; Browning et al. 2015). As the purpose of this model was to measure change in learning rates between blocks rather than describe the mechanism by which estimated information content is calculated, it was fit separately to each of the task blocks. The first 10 trials of each block were omitted when fitting the model parameters to participants’ choices, as initial learning rates are generally inflated due to estimation uncertainty during new tasks.

### Win- and Loss-Driven Behaviour

### Statistical Analyses

## Results

### Baseline Questionnaire Measures

Study 1 Negative IBLT training | Study 2 Positive IBLT training | |
---|---|---|

BDI-II | 3.70 (1.33) | 4.00 (33.15) |

STAI-Trait | 34.35 (2.08) | 33.15 (1.22) |

PANAS Positive | ||

Sham tDCS session | 33.40 (1.39) | 32.35 (1.70) |

Bifrontal tDCS session | 33.45 (1.26) | 33.60 (1.87) |

PANAS Negative | ||

Sham tDCS session | 11.55 (0.62) | 11.35 (0.43) |

Bifrontal tDCS session | 10.50 (0.18) | 11.80 (0.63) |

STAI-State | ||

Sham tDCS session | 27.40 (1.52) | 26.15 (1.29) |

Bifrontal tDCS session | 26.35 (1.36) | 26.60 (1.15) |

### Study 1: Effects of Negative IBLT Training and Concurrent tDCS on Reward Learning

#### Computational Learning Parameters

^{2}= 0.112). It was also predicted that, through gaining experience on the task, learning rates for loss outcomes would increase over time. Contrary to this hypothesis, however, there was neither a main effect of Block (F(2,36) = 0.30, p = 0.746) nor an interaction of Block and Outcome valence (F(2,36) = 1.62, p = 0.211). As shown in Fig. 3a, a negative learning bias was rapidly induced in ‘Training’ block 1 and persisted in the subsequent two blocks.

^{2}= 0.022), with learning rates being higher (for both wins and losses) before than after IBLT training. Importantly, however, there was also a numerical trend towards an interaction of Time with Outcome valence (F(1,18) = 3.51, p = 0.077, η

^{2}= 0.008), with average learning rates decreasing more for win (t(39) = 2.79, p = 0.008) than loss outcomes (t(39) = 0.64, p = 0.525) over time (see Fig. 3b). This trend is consistent with the aim of the training procedure to encourage faster learning from negative relative to positive outcomes.

#### Win- and Loss-Driven Choice Behaviour

^{2}= 0.090) as shown in Fig. 4b. This increase in loss-driven choices after negative IBLT training is consistent with the trend towards near-transfer for loss learning rates.

#### Emotional Face Recognition

#### Mood and Anxiety Measures

^{2}= 0.056). Potentially, this change in positive mood could be attributed to the negative IBLT training. Alternatively, it is possible that there may be other non-specific factors (e.g. fatigue, loss of motivation) which contribute to reduced positive affect over time.

### Study 2: Effects of Positive IBLT Training and Concurrent tDCS on Reward Learning

^{2}= 0.452). In addition, non-computational analyses indicated that participants chose the win-driven option over the loss-driven option on 62.8% of the trials. These findings demonstrate that the positive IBLT training effectively induced a congruent learning bias in the ‘Training’ blocks (see Fig. 5). However, we found no evidence of near transfer with positive IBLT training/tDCS for learning rates, win-driven behaviour, FERT performance, or mood (see Online Resource 1).