# Lever Position from Mouse Brain Activity: Official commentary (slides)

*English translation by SOTA – AI Community of the Japanese original, shared under CC BY-NC-SA 3.0 IGO like the original. Organisers who would like this translation removed can email sota.ai.community@gmail.com.*

*Translation of the slide deck "JOAI2026 講評" (JOAI2026 Review) by the JOAI committee, 17 slides. Every slide carries the header "JOAI2026 Review" and the JOAI logo. Figures are described in words; see the [original slides](https://speakerdeck.com/go5paopao/joai2026-jiang-ping) for the images.*

## Slide 1

**The 2nd Japan Olympiad in Artificial Intelligence**
**JOAI2026 Review**

## Slide 2: About me

**Director, JOAI Committee**
**Hibiki Hattori (服部 響)**

*[Photo]*

Got into machine learning through developing a mahjong AI. In a previous job, worked on image recognition apps and user profiling. After joining ABEJA in 2020, worked as a data scientist on a variety of projects and managed a data science organisation, then in 2024 became a Principal Data Scientist as a specialist. Kaggle Competitions Grandmaster. Has won several competitions, including the Kaggle Days World Championship and Japanese competitions such as atmaCup.

## Slide 3: Competition recap

**Task: predict, as a time series, the position of the lever moved by a mouse from the mouse's brain activity**

*[Figure: a table of brain activity (time series) → AI model → predicted values, compared with the ground truth.]*

Participants compete on accuracy, measured by the mean squared error (MSE) between the predicted values and the ground truth.

The prediction target is the position of the lever, not the movement of the mouse itself.

## Slide 4: Competition recap

**The prediction targets are time-series samples cut out of experiments on several mice over several days**

*[Figure: data structure. Mouse A, B and C (mouse_id = 1, 2, 3) each have a 14-day experiment (Day 1 … Day 14). One day's experimental record (lever & brain activity) is cut into short time-series samples (20–40 frames), e.g. sample_id 101, 102, 103. The frames (records) belonging to sample_id 102 form a table with the columns id, time, lever (target variable) and brain (brain activity), for example ids 5001–5030 at times t1–t30 with lever values v1–v30; they are linked to mouse_id 1, day_n 7.]*

Besides the brain activity data, the following are given:
- mouse ID
- which day of the experiment it is
- timestamp

The cut-out samples have a random length of 20–40 frames.

## Slide 5: Competition recap

**Brain activity: wide-field calcium imaging**

*[Figure: map of the 22 cortical regions of one hemisphere (Anterior/Posterior, Left/Right), with the list of region names in English: 1. Main olfactory bulb (OB); 2. Secondary motor area (MOs); 3. Primary motor area (MOp); 4. Primary somatosensory area, mouth (SSp_m); 5. Primary somatosensory area, upper limb (SSp_ul); 6. Primary somatosensory area, lower limb (SSp_ll); 7. Primary somatosensory area, nose (SSp_n); 8. Primary somatosensory area, unassigned (SSp_un); 9. Primary somatosensory area, trunk (SSp_tr); 10. Primary somatosensory area, barrel field (SSp_bfd); 11. Supplemental somatosensory area (SSs); 12. Posterior parietal association areas, anterior area (VISa); 13. Posterior parietal association areas, rostrolateral visual area (VISrl); 14. Anterolateral visual area (VISal); 15. Secondary auditory areas (AUDs); 16. Primary auditory area (AUDp); 17. Lateral visual area (VISl); 18. Primary visual area (VISp); 19. Posteromedial visual area (VISpm); 20. Anteromedial visual area (VISam); 21. Retrosplenial area, lateral agranular part (RSPagl); 22. Retrosplenial area, dorsal part (RSPd).]*

The features are the activity values of 22 symmetric regions on each of the left and right sides of the brain (44 in total).

Later in the session there is also a review from Nakae-sensei (中江先生), who published the paper on building this dataset!!

## Slide 6: Competition recap

- Time-series data based on brain activity
- The prediction target is not the movement itself but the position of the object being moved (the lever)
- The brain activity is measured with a technique called wide-field calcium imaging
- A regression problem
- Information on the mouse and the date is available
- The length of the cut-out sequences is random, between 20 and 40

## Slide 7: Why time-series data matter

**Time-series data (and sequence data in general) come up frequently in machine learning and AI**

- Demand forecasting
- Video analysis
- Recommendation (purchase histories, etc.)
- Physical AI / robotics
- Chemistry (amino-acid sequences, etc.)
- Natural language processing (LLMs, etc.)

and so on.

The skills carry over to future applications, so we wanted to cover this early on in JOAI.

## Slide 8: Why time-series data matter

**Time-series data (and sequence data in general) come up frequently in machine learning and AI**

| Sequence | Input |
|---|---|
| 1 | $X_1$ |
| 2 | $X_2$ |
| 3 | $X_3$ |
| … | … |

→ Prediction $y$

- Is the input
  - numerical data? label data?
  - evenly spaced in time? or not?
  - linked to other data?
- Is the prediction target
  - predicted for each sequence element? or one prediction for the whole?
  - numerical? a label?
  - the present of the sequence? or its continuation?

**The differences are mostly just in these patterns. Many of the model architectures that handle time series are similar.**

## Slide 9: Why time-series data matter

**Time-series data (and sequence data in general) come up frequently in machine learning and AI**

- Is the input
  - numerical data? label data? → numerical data
  - evenly spaced in time? or not? → evenly spaced
  - linked to other data? → linked through the mouse and the date
- Is the prediction target
  - predicted for each sequence element? or one prediction for the whole? → predicted for each sequence element
  - numerical? a label? → numerical prediction
  - the present of the sequence? or its continuation? → predict the present

**What would change if each of these followed a different pattern? Just thinking about that helps you apply what you have learned.**

## Slide 10: Aims of the competition design

- On the amount of data
  - Tuned so that there are enough records to train even at the level of individual sample IDs, while keeping the overall computing resources (GPU) needed low. (Possible shifts between the Public and Private rankings were also taken into account.)
  - A fixed sample length would also allow approaches other than time-series ones, so the length was made random.
- Aims regarding models
  - Given the number of records, we wanted gradient boosting and various deep learning models (GRU, LSTM, CNN, Transformer, etc.) all to be worth trying.
  - We also wanted ensembles and combinations of several architectures to be meaningful.

## Slide 11: Aims of the competition design

- On the train/test split
  - Given the length of the contest, it was deliberately kept simple: a random split, including over mouse_id and date.
  - Using the later dates as the test set, or splitting train/test by mouse_id, would make it harder.
- Choice of the target variable
  - The data contained many kinds of data from the mouse experiments, but the target was simply set to the lever position.
  - Making this more complex would raise the bar, for example by making it necessary to read the paper.
  - Brain activity is linked to the mouse's movements, whereas the lever is not the movement itself but the state of the external environment. This left room for ingenuity.

## Slide 12: Aims of the competition design

- Providing mouse_id / day_n
  - One can expect brain-activity characteristics specific to each mouse, changes between days, and so on.
  - How can these be incorporated into the prediction?
  - This called for measures tailored to the trends in the data rather than simple modelling.
- Brain activity data (this part is in hindsight)
  - With the major methods of measuring brain waves, dedicated pre-trained models and analysis methods are established, so the contest would become a test of research skills.
  - With this data, participants gain experience of analysing unfamiliar data.

## Slide 13: Trends in the top solutions (overall)

- Neural-network-centred models suited to time series
  - Bidirectional GRU/LSTM, TCN and similar models were common
  - Transformers were less common
- Temporal difference features
- Features based on left–right symmetric brain regions
- Absorbing differences between individual mice (embeddings, normalisation)
- Ensembling / stacking of diverse models
  - Diversity was ensured by partially changing the model settings
  - Ensembles also included decision-tree models such as LightGBM

## Slide 14: Key point: models

- This time, BiGRU/BiLSTM, TCN and similar models were strong.
- The models themselves are not new; many have not changed for several years.
- Performance depends heavily on the characteristics and size of each dataset.
  - Even in similar time-series competitions, the strongest model changes every time.
- The top participants reached strong models by trying a wide variety of models.
- Past Kaggle competitions are also a useful reference.

**It is important to work out which models and architectures are plausible for the data, and to try many of them.**

## Slide 15: Key point: differences over time

- The prediction target is the lever position → it changes when the mouse acts to move the lever.
- Aren't relative values (changes) more important than the absolute values of the brain-activity features themselves?
  - Activity increases and a movement is made / a movement is made and, as a result, activity increases.
- Moreover, isn't what is directly linked to the mouse's movement not the lever position itself but the change in the lever position?
  - ✕: brain activity is high → the lever position is small
  - ◯: brain activity is high → the lever position is likely to be changing

**This leads to the hypothesis that using the amount of change (differences along the sequence) as features, or predicting the change in the lever position itself, may be effective.**

## Slide 16: Summary and impressions

- JOAI2026 was a competition that is, in some respects, a classic time-series problem, while leaving some room for ingenuity.
- Several of the aims of the competition design were made public.
- The overall trends of the top solutions were summarised.
  - Look forward to the presentations of the individual solutions for details.

**Impressions**
- Thanks to advances in AI coding, implementation speed felt high across the board.
- Even so, I was surprised by how much the scores improved in such a short period of just over a week.
- I would be glad if this serves as a starting point or a reference for learning AI and machine learning in the future.

## Slide 17

**The 2nd Japan Olympiad in Artificial Intelligence**
**JOAI2026 Review**