Dual N-Back

What the task asks

In an n-back task, items come one at a time and you react whenever the current item is the same as the one n steps earlier. At 1-back that is the item just before. At 3-back you keep the last three in mind and drop the oldest every time a new one arrives. Dual n-back runs two streams at once: a square that lights up in one of the 8 outer places of the grid, and a letter in the centre. Each stream has its own key, and n is the same for both.

A short example at 2-back. The square appears top left, then right, then top left again: the third step is a position match. The letters are K, P, K: the third step is also a letter match, so you press both keys. On the next step the square is at the bottom and the letter is H, which matches nothing two steps back, so you press nothing.

The lab lets you explore one question for yourself: how far can you push n before the two streams start to blur, and how does that change when you practise on several days?

Things to try

• Run a session of 6 blocks and read the table: which stream costs you more errors, position or letter?
• On the same day, try Position and Letter on their own and compare the levels with your Dual level.
• Look at the two kinds of error in the table. Do you miss matches, or press when there is none?
• Turn on Show right and wrong presses for one session and see whether your level moves differently.
• Train on several days and follow the mean n on the chart. Find the day where the line flattens.

What to watch for

The table under This session splits every error into a missed match and a press without a match. A missed match means the item n steps back had slipped away; a press without a match means you held a wrong item or guessed. In the dual task the level goes up only when both streams were good, so one weak stream holds you back.

The rules this page follows

Susanne Jaeggi, Martin Buschkuehl, John Jonides and Walter Perrig used this task for training in a study published in 2008, with 70 healthy young adults in trained groups and in control groups without training. Their version:

• a square in one of 8 places, shown for 500 ms, a new one every 3 s (2,500 ms between them);
• at the same time one of 8 consonants, spoken through headphones;
• 6 matches per stream in every block, 4 of them in only one stream and 2 in both, at random steps;
• key A for a position match and L for a letter match, no key when nothing matches;
• a block of 20 + n steps; after it, n went up by 1 after fewer than 3 mistakes in each stream, down by 1 after more than 5 mistakes, and otherwise stayed the same;
• a training day of 20 blocks, about 25 minutes.

This page uses the same timing, number of matches, keys and level rule. It differs in four points:

• The letter is shown, not spoken. A browser voice may be an online service that receives the text, and the moment it starts to speak is not exact, so a spoken letter could not be kept in step with the square. Both streams therefore go through your eyes.
• Counting mistakes. The study does not spell out what a mistake is. Here a mistake is a missed match or a press when there was no match, counted for each stream separately.
• Session length. You choose 3, 6, 10 or 20 blocks, and the next session starts at the level where the last one ended (you can change it under Start at).
• Practice. The first session in each task starts with one 1-back block with feedback, which is not counted.

The letters are B, F, H, K, M, P, R and T, the 8 places are the outer cells of the grid, and no match is placed in the first n steps of a block, because there is nothing n steps back yet.

What training does and does not do

Practice raises your level in this task: the 2008 study shows the mean n of every training group climbing over the sessions. The study also reported that the trained groups gained more on a matrix reasoning test than the untrained groups, and more with more days of training. Its control groups did no training at all.

A meta-analysis by Monica Melby-Lervåg, Thomas Redick and Charles Hulme (2016) pooled 87 publications with 145 comparisons of working memory training. Right after training, people did better on other working memory tests. On nonverbal reasoning, in 67 studies whose control group did a different activity of similar length, the mean effect was g = 0.05 (95% interval −0.02 to 0.13), close to zero. For n-back training alone it was g = 0.15, and 0.10 once the most problematic study was removed, which the authors call negligible. Their conclusion: the programmes give short-term, specific training effects that do not generalise to measures of real-world cognitive skills.

So a rising line on the chart shows that you got better at dual n-back. It says nothing about intelligence, memory in daily life or results at school. Training this task is not a treatment, and your level is not an IQ score.

How the result is computed

• Each step is open for answers from the moment it appears until the next step, 3 s. Only the first press of each key in a step counts.
• For each stream of a block: a match with no press is a missed match, a press on a step without a match is a press without a match, and mistakes are the two added together. Every stream has 6 matches.
• The next n follows the rule above, between 1 and 9. The session shows the mean n of its counted blocks and the highest n; a day on the chart shows the mean and the highest n of all blocks on that day.
• The page measures how often your screen redraws during Get ready and shows each stimulus for a whole number of frames: 500 ms is 30 frames at 60 Hz. A step in which a frame was skipped is marked in the saved data.
• The history chart shows by default only sessions with the same kind of input and a similar refresh rate (within 10%, or both below 50 Hz), for example Keyboard · 60 Hz; untick the box to see all of them.

Limits

• Shown letters are not the task of the 2008 study, where the letters were heard. Levels here are not comparable with versions that speak the letters, and two visual streams may be harder or easier for you than one visual and one spoken stream.
• The mistake count is this page's reading of the rule, so a level here is not the same measurement as a level in the study.
• The chart shows how you do in this task with these settings, nothing more.
• If the tab is hidden during a block, that block is discarded and the session waits for you to resume.
• Below 50 frames a second, or with an uneven screen, a stimulus can last a frame longer or shorter; the page says so when the block starts.

Controls

• Keyboard: A = position match, L = letter match; Esc pauses a running block (the block is discarded and starts again on Resume); Tab moves between the buttons, Space or Enter presses them.
• Touch and mouse: the Position and Letter buttons under the grid; a press counts at the moment your finger or the mouse button goes down.
• Buttons: Start session, Pause, Next block, Resume, End session (the finished blocks are kept), Download CSV (every saved block), Clear history (asks first; your level and settings stay).
• Settings: Task (Dual, Position or Letter), Blocks per session, Start at (n from 1 to 9), the practice block, and feedback during the blocks.

FAQ

Why did my level drop after a block that felt fine?

In the dual task, more than 5 mistakes in either stream lowers n, even when the other stream had none. The table shows both counts for every block, so you can see which stream it was.

Sources

• Jaeggi SM, Buschkuehl M, Jonides J, Perrig WJ (2008). Improving fluid intelligence with training on working memory. Proceedings of the National Academy of Sciences 105(19): 6829–6833. The task, its timing, the matches per block and the level rule.
• Melby-Lervåg M, Redick TS, Hulme C (2016). Working memory training does not improve performance on measures of intelligence or other measures of “far transfer”: evidence from a meta-analytic review. Perspectives on Psychological Science 11(4): 512–534. What training does and does not change.