Skip to content

Task_parameters.md

Latest commit

 

History

History
146 lines (99 loc) · 6.12 KB

Task_parameters.md

File metadata and controls

146 lines (99 loc) · 6.12 KB

Change detection task parameters

Our version of the change detection task (CDT) introduces two modifications:

  • 2 CDT blocks within session separated by a break to allow injection
  • Each CDT block is further divided in two sections (short & long) in which the image changes are drawn from different distributions

overview_task

Some parameters regarding these modifictions have to be optimized

  1. Duration of the pre and post blocks
  2. Duration of the short & long sections within each CDT

1. Duration of the pre and post blocks

Pre block should be 1268 s long. This should corresponds to ≈ 191 total trials. Trials should be equally distributed across the two sections with different taus.

See calculations in "Duration of the pre and post blocks" --> "Option 2". Ideally the post block should not have a predetermined end but should last until animal is engaged in the task.

2. Duration of the short & long sections within each CDT

With:

tau_1 = 0.4
tau_2 = 0.1

and a block of duration X. Time allocation between sections with different taus should be:

T_tau2 = X / 1.776
T_tau1 = 0.776 * (X / 1.776)

If we choose a duration of 1268 s:

T_tau2 = 714 s
T_tau1 = 554 s

See "Duration of the short & long sections within each CDT" below for details.

Duration of the pre and post blocks

The aim is to have similar performance/engagement between the two pre- post- injection blocks. It is paramount that the mouse is not satiated after the first block, therefore not engaging in the task afterward.

2 options are available:

Currently option 2 is choosen.

Option 1

To ensure similar engagement we can look at the the mean reward amount or number of hit trials per session in the visual-behavior-neuropixels dataset and dynamically change the block duration contingent on the number of hit trials/rewards accumulated.

Training is structured in the following stages (12 example mice): Training_stages_example_mice

We compute the average reward per mouse during the last 2 stages of training: total_reward_volume or on recording days only: total_reward_volume

We also compute the number of hit trials during the last 2 stages of training: hit_trial or on recording days only:hit_trial

Option 2

We can hardcode the duration of the blocks by looking at the mean time necessary to reach 50% of the total reward in each session. Analysis run on 153 session of the visual-behavior-neuropixels dataset, all recording sessions available, including sessions with abnormalities in either histology or recorded activity (cache.get_ecephys_session_table(filter_abnormalities=False)).

Half_reward

Let´s subtract 1 std from the mean to find a good empirical limit for the first repetition, either in number of trials:

311-120 = 191 trials

or seconds:

1631-363 = 1268 s

Given the use of different taus in the distributions to select image changes (long and short sections) the situation is more complicated. However, the duration of each sections is weighted to produce similar amounts of hit trials (assuming cnstant performance) the final effect should not deviate greatly (5%) from the data analyzed here with one distribution with tau = 0.3.

Proportion of hit_trials/total_trials during recording sessions: ratio_hit_total_trials.png

This plot does not exclude periods when animal is disengaged. This happens often in the last portion of the session when animal is not thirsty anymore.

Duration of the short & long sections within each CDT

The aim here is to have the same number of hit trials in each section. We can expect the mice to collect more rewards in the section where image changes tend to happen earlier. We should adjust for this by reducing the duration of this section. First step here is to decide precisely the parameters of the distributions. From the proposal: "The coefficient of the geometric distributions (change_time_scale) will be 0.1 and 0.4." Change times are selected from a truncated geometric distribution from 4 to 11 stimulus flashes after the trial start (in the Neuropixel Visual behavior the actual truncation is 4 to 11, not 5 to 11 as stated in the whitesheet).

overview_task

To determine how much faster tau_1 is compared to tau_2, we can use the ratio of the sum of times generated by the two distributions. Given that

tau_1 = 0.4
tau_2 = 0.1
size = 1000  # Number of samples

# Function to simulate truncated geometric distribution
def truncated_geometric(tau, size, min_val, max_val):
    samples = []
    while len(samples) < size:
        sample = np.random.geometric(tau)
        if min_val <= sample <= max_val:
            samples.append(sample)
    return np.array(samples)

# Simulate truncated geometric distributions
truncated_dist_1 = truncated_geometric(tau_1, size, 4, 11)
truncated_dist_2 = truncated_geometric(tau_2, size, 4, 11)

# Convert number of flashes to time in milliseconds
# Each image flash is 250 ms and followed by 500 ms of grey screen, total 750 ms per cycle
times_1 = truncated_dist_1 * 750
times_2 = truncated_dist_2 * 750
sum(times_1) / sum(times_2) = 0.7761322789360172

it indicates that the average time for tau_1 = 0.4 is approximately 77.6% of the average time for tau_2 = 0.1.

Speed Increase = 1 / 0.7761322789360172 ≈ 1.288

This means that tau_1 = 0.4 is about 1.288 times faster tau_2 = 0.1.

If we have a block of duration X, and we assume stable performance we should allocate:

T_tau2 = X / 1.776
T_tau1 = 0.776 * (X / 1.776)

Using this balanced ratio the toatal amount of change times per duration X should be similar to the one used in the Visual Neuropixel dataset (5% difference)