Notebook 8 — Strategy is not engagement (optional)¶
SWC ENC 2026 · behaviour module
Every mouse so far licked throughout the session. Real mice disengage — they stop licking for stretches, then re-engage. A natural worry: is "disengaged" just another strategy the model should include? Piet et al. (Figure 3) argue no — engagement is a separate axis from strategy. Because we generate engagement as its own latent, independent of the strategy weights, we can demonstrate that cleanly: the analysis recovers the two axes as separately as we built them.
In this notebook you will:
- Meet a mouse that disengages, and see it in the licking rate.
- Recover the hidden engagement state from rolling rates (Fig 3A–B).
- Show the recovered strategy is unchanged by engagement.
The paper panel this notebook reproduces
Fig. 3B — one session's lick-bout rate and reward rate over time, with the engaged/disengaged threshold. This notebook recovers that engagement state and shows it is a separate axis from strategy.
import numpy as np
import matplotlib.pyplot as plt
import swcbehav as sb
# A visual-strategy mouse that also drifts in and out of task engagement.
eng = sb.make_mouse("visual", seed=2, engaged=True)
tbl = eng.table
print(f"fraction of time truly engaged: {tbl['true_engaged'].mean():.2f}")
fraction of time truly engaged: 0.85
1. Disengagement in the raw behavior¶
When the mouse disengages, its licking nearly stops. Here is the smoothed lick-bout rate across the session, with the true disengaged periods shaded. The rate collapses in the shaded stretches — engagement gates licking on top of whatever strategy the mouse is running.
def rolling_rate(events, window=100):
box = np.ones(window) / window
return np.convolve(np.asarray(events, float), box, mode="same")
bout_rate = rolling_rate(tbl["bout_start"].to_numpy())
plt.figure(figsize=(11, 2.8))
plt.plot(bout_rate, color="k", lw=1)
diseng = ~tbl["true_engaged"].to_numpy()
plt.fill_between(np.arange(len(tbl)), 0, bout_rate.max(), where=diseng,
color="tab:red", alpha=0.15, label="truly disengaged")
plt.xlabel("image"); plt.ylabel("bout rate"); plt.legend()
plt.title("licking collapses during disengaged periods")
plt.show()
2. Recovering engagement from rolling rates¶
We don't get the true engagement state for real mice — we infer it. Following the paper, we compute rolling rates of licking and reward and call the mouse disengaged wherever the lick-bout rate falls below a threshold.
Exercise 1 (~5 min). Complete classify_engaged: smooth the bout-start vector into a
rate and threshold it. Then we'll check the recovered state against the truth.
Check / unstuck. The recovered engagement should match the hidden truth on ~90%+ of images. Stuck? The solutions copy has it.
def classify_engaged(sess, window=100, threshold=0.05):
rate = rolling_rate(sess.table["bout_start"].to_numpy(), window)
return rate > threshold
engaged_hat = classify_engaged(eng)
agreement = (engaged_hat == eng.table["true_engaged"].to_numpy()).mean()
print(f"recovered engagement matches the hidden truth on {agreement:.1%} of images")
recovered engagement matches the hidden truth on 92.6% of images
The rolling-rate classifier recovers the hidden engagement latent well; the mismatches sit at the transitions, where smoothing blurs the exact switch time. The paper draws its engaged/disengaged boundary in the 2D plane of lick-bout rate vs reward rate — the two are tightly linked, since rewards can only come while licking:
reward_rate = rolling_rate(tbl["reward"].to_numpy())
plt.figure(figsize=(5, 4))
plt.scatter(bout_rate, reward_rate, c=tbl["true_engaged"].to_numpy(),
cmap="coolwarm", s=6, alpha=0.4)
plt.xlabel("lick-bout rate"); plt.ylabel("reward rate")
plt.title("the engagement plane (cf. Fig 3A)")
plt.colorbar(label="truly engaged"); plt.tight_layout(); plt.show()
3. Strategy survives engagement¶
Now the key claim. If engagement were just another strategy, adding it would distort the strategy readout. It doesn't. We compute the strategy index on the engaged images of this disengaging mouse and compare it to the same mouse run fully engaged — the strategy should be unchanged, because we built engagement to act only on the overall drive to lick, not on the visual/timing balance.
Exercise 2 (~4 min). Complete strategy_index_on: restrict $X, y$ to a boolean mask of
images, then run the ablation from Notebook 5.
Check / unstuck. The two strategy indices (fully-engaged vs engaged-only) should land close together (~+0.22). Stuck? The solutions copy has it.
def strategy_index_on(sess, mask=None):
X, cols = sb.build_design_matrix(sess.table)
y = sess.table["bout_start"].to_numpy().astype(float)
if mask is not None:
X, y = X[mask], y[mask]
d = sb.ablation_loglik_deltas(X, y, cols)
return d["visual"] - d["timing"]
fully_engaged = sb.make_mouse("visual", seed=2) # same strategy, no disengagement
si_reference = strategy_index_on(fully_engaged)
si_engaged_only = strategy_index_on(eng, mask=engaged_hat)
print(f"strategy index, fully-engaged mouse : {si_reference:+.4f}")
print(f"strategy index, engaged images only : {si_engaged_only:+.4f}")
strategy index, fully-engaged mouse : +0.2379 strategy index, engaged images only : +0.2203
The two strategy indices land in the same place: engagement did not change the strategy we recover. One latent controls whether the mouse is playing; a separate set of weights controls how it plays. We recover them as independently as we built them — which is exactly the paper's conclusion that strategy and engagement are distinct axes.
Wrap-up¶
Engagement gates overall licking; strategy sets the visual/timing balance. They're separate knobs in the generator, and the analysis pulls them apart: rolling rates recover engagement, ablation recovers strategy, and neither contaminates the other.
That completes the behaviour module — Figures 1–3 of Piet et al., built up from a design matrix to a drifting-strategy model, every step checked against a ground truth you controlled.