Notebook 3 — Modeling the licks¶
SWC ENC 2026 · behaviour module
We can now build the design matrix $X$ (Notebook 1) and we've seen strategies leave fingerprints in behavior (Notebook 2). But so far we peeked at ground truth to know each mouse's strategy. This notebook is where we stop peeking and infer the strategy from behavior, by fitting the model the paper is built on: logistic regression.
In this notebook you will:
- Write down the model and its likelihood.
- Fit it by maximizing that likelihood.
- Recover the true strategy weights of single-strategy mice.
- See what a single static fit does to the dynamic mouse — the crack that motivates the rest of the day.
import numpy as np
import matplotlib.pyplot as plt
from scipy.optimize import minimize
import swcbehav as sb
def get_Xy(sess):
# Observable design matrix X and per-image outcome y for a session.
X, cols = sb.build_design_matrix(sess.table)
y = sess.table["bout_start"].to_numpy().astype(float)
return X, y, cols
1. One image at a time: the Bernoulli model¶
Start with a single image. The model turns its design row $x_t$ and the weights $w$ into a probability that a licking bout starts on it:
$$p_t = \sigma(w \cdot x_t), \qquad \sigma(z) = \frac{1}{1 + e^{-z}} .$$
The outcome $y_t$ is binary — a bout either starts ($y_t = 1$) or it doesn't ($y_t = 0$) — so one image is a biased coin flip. The distribution of a single yes/no outcome with success probability $p_t$ is the Bernoulli distribution:
$$P(y_t \mid p_t) = \begin{cases} p_t & \text{if } y_t = 1 \\ 1 - p_t & \text{if } y_t = 0 \end{cases} \;=\; p_t^{\,y_t}\,(1 - p_t)^{\,1 - y_t}.$$
The compact form on the right packs both cases into one expression, using the fact that anything to the power $0$ is $1$: when $y_t = 1$ it reduces to $p_t$; when $y_t = 0$ it reduces to $1 - p_t$. Either way it is the probability the model assigns to what the mouse actually did on image $t$.
2. A whole session: independence makes a product¶
We don't have one image, we have thousands. The likelihood is the probability the model assigns to the entire observed sequence of outcomes, read as a function of the weights $w$:
$$\mathcal{L}(w) = P(y_1, y_2, \dots, y_T \mid w).$$
To get from one image to the whole sequence we make a modeling assumption: the outcomes are conditionally independent given the weights and the stimulus. "Independent" here means that once $w$ and the design rows are fixed, knowing a bout started on image 5 tells us nothing extra about image 6 — whatever one image implies about the next is already carried in the design matrix (that is exactly the job of the timing regressor). Independence is precisely the condition that lets us multiply the per-image probabilities together:
$$\mathcal{L}(w) = \prod_{t=1}^{T} p_t^{\,y_t}\,(1 - p_t)^{\,1 - y_t}.$$
(It's an idealization — real licking surely has leftover dependencies — but it is the assumption the model is built on, and what makes the problem tractable.)
3. Why we work with the log¶
That product is a clean mathematical object but a terrible one to compute. With $T \approx 4800$ images, every factor is a probability below $1$, and multiplying thousands of them underflows to zero in floating point — the true value is smaller than the computer can represent. Taking the logarithm cures this and brings two more benefits:
- $\log$ turns the product into a sum, which can't underflow and is far easier to differentiate (we'll need that to fit);
- $\log$ is monotonically increasing, so the $w$ that maximizes $\mathcal{L}(w)$ is the same $w$ that maximizes $\log \mathcal{L}(w)$ — switching to the log costs us nothing.
The log of the product becomes a sum of logs — the log-likelihood:
$$\log \mathcal{L}(w) = \sum_{t=1}^{T} \Big[\, y_t \log p_t + (1 - y_t)\log(1 - p_t)\,\Big].$$
Substituting $p_t = \sigma(w\cdot x_t)$ and simplifying (worth grinding through once by hand) removes the $\sigma$ entirely and gives the form we actually code:
$$\log \mathcal{L}(w) = \sum_{t=1}^{T} \Big[\, y_t\,(w\cdot x_t) - \log\!\big(1 + e^{\,w\cdot x_t}\big) \Big].$$
This version is also numerically safer: it never forms $p_t$ and then logs it
(which blows up for confident predictions). The awkward $\log(1 + e^z)$ term has
its own stable routine, np.logaddexp(0, z).
Exercise 1 (~8 min · meaty). Implement this log-likelihood inside neg_log_likelihood. We
return the negative log-likelihood because the optimizer in the next step
minimizes rather than maximizes. The extra l2 term adds a small penalty on large
weights — a weak Gaussian prior that keeps the fit well-behaved and turns pure
maximum-likelihood into the maximum a posteriori (MAP) estimate the paper uses;
leave it as given.
Check / unstuck. The next cell should print NLL at w=0 ≈ 3327, matching
n·log(0.5). Stuck? Usesb.neg_log_likelihood.
def neg_log_likelihood(w, X, y, l2=1e-4):
z = X @ w # w . x for every image
log_lik = np.sum(y * z - np.logaddexp(0.0, z))
return -log_lik + l2 * np.sum(w ** 2) # ridge = weak Gaussian prior (MAP)
# Sanity check: at w = 0 every p_t = 0.5, so the log-likelihood is n*log(0.5).
Xv, yv, cols = get_Xy(sb.make_mouse("visual", seed=0))
w0 = np.zeros(Xv.shape[1])
print("NLL at w=0:", round(neg_log_likelihood(w0, Xv, yv), 1))
print("expected :", round(-len(yv) * np.log(0.5), 1))
NLL at w=0: 3327.1 expected : 3327.1
4. Maximizing the likelihood¶
Fitting the model means choosing the weights under which the mouse's actual licking was most probable — the $w$ that maximizes $\log \mathcal{L}(w)$, or equivalently minimizes the negative log-likelihood we just built.
How do we find that $w$? The negative log-likelihood is a smooth function of the weights, so at any point we can compute its gradient — the direction in weight-space in which it rises fastest — and step the opposite way to go downhill. Repeating that is gradient descent: start from a guess, measure the slope, take a step down, and iterate until the slope flattens out. For logistic regression the gradient takes a tidy, interpretable form,
$$\nabla\big(\!-\log\mathcal{L}\big) = \sum_t (p_t - y_t)\,x_t,$$
so each image pushes the weights in proportion to its prediction error $(p_t - y_t)$ — images the model got wrong pull hardest. And this objective is convex (bowl-shaped, no local minima to trap us), so gradient descent is guaranteed to reach the single global optimum.
Plain gradient descent only uses the slope and can zig-zag slowly toward the
bottom. We can converge in far fewer steps by also using curvature — how the
slope is itself changing — to pick smarter step directions and sizes. That is what
the optimizer we'll call, L-BFGS-B, does; treat it as a black box whose one-line
summary is "gradient descent that also exploits curvature." scipy.optimize.minimize
even estimates the gradient for us here — in Notebook 7 we'll supply one by hand.
Exercise 2 (~4 min · easy). Complete fit_static using scipy.optimize.minimize with
method="L-BFGS-B", starting from an all-zeros weight vector.
Check / unstuck. The visual mouse should recover bias ≈ −3, visual ≈ 5, the rest ≈ 0. Stuck? Use
sb.fit_static.
def fit_static(X, y, l2=1e-4):
result = minimize(neg_log_likelihood, np.zeros(X.shape[1]),
args=(X, y, l2), method="L-BFGS-B")
return result.x
w_fit = fit_static(Xv, yv)
w_true = sb.make_mouse("visual", seed=0).true_weights[0]
print(f"{'strategy':14s}{'true':>7s}{'fit':>8s}")
for c, t, f in zip(cols, w_true, w_fit):
print(f"{c:14s}{t:7.1f}{f:8.2f}")
strategy true fit bias -3.0 -2.96 visual 5.0 5.03 omission 0.0 -0.12 post_omission 0.0 -0.46 timing 0.0 -0.17
The fit puts nearly all the weight on visual, with bias near $-3$ and the other strategies near zero — it has recovered the strategy from behavior alone, without ever seeing the true weights. This "can we recover the truth?" check is the backbone of the whole course.
5. Recovery across strategies¶
Does it work for every single-strategy mouse? Fit all three and compare fitted to true weights.
archetypes = ["visual", "timing", "mixed"]
fig, axes = plt.subplots(1, 3, figsize=(12, 3.2), sharey=True)
xpos = np.arange(len(cols))
for ax, name in zip(axes, archetypes):
sess = sb.make_mouse(name, seed=0)
X, y, _ = get_Xy(sess)
w_fit = fit_static(X, y)
w_true = sess.true_weights[0]
ax.bar(xpos - 0.2, w_true, 0.4, label="true", color="0.6")
ax.bar(xpos + 0.2, w_fit, 0.4, label="fit", color="tab:red")
ax.set_xticks(xpos); ax.set_xticklabels(cols, rotation=45, ha="right")
ax.axhline(0, color="k", lw=0.8); ax.set_title(f"{name} mouse")
axes[0].set_ylabel("weight"); axes[0].legend()
plt.tight_layout(); plt.show()
Fitted bars track the true bars across all three mice. A few things worth saying out loud:
- The weights are log-odds. A bias of $-3$ means a baseline bout probability of $\sigma(-3) \approx 0.05$; adding a visual weight of $+5$ on a change image gives $\sigma(-3 + 5) = \sigma(2) \approx 0.88$ — the hit rate we measured in Notebook 2.
- The omission and post-omission weights wobble more around their true value of zero. Those events are rare, so there's little data to pin them down — the same reason the paper found the omission strategy weakly determined.
6. A single fit meets the dynamic mouse¶
Everything above assumed the strategy is fixed. What happens if we fit one static $w$ to the dynamic mouse, whose strategy drifts? The optimizer returns the single weight vector that best explains the whole hour at once — a blurry average that lands between visual and timing and matches the animal at no point in particular.
dyn = sb.make_mouse("dynamic", seed=1)
Xd, yd, _ = get_Xy(dyn)
w_static = fit_static(Xd, yd)
sb.plotting.plot_weights(true_weights=dyn.true_weights, col_names=sb.WEIGHT_NAMES)
for k, name in enumerate(sb.WEIGHT_NAMES):
plt.axhline(w_static[k], color=sb.WEIGHT_COLORS[name], ls=":", lw=1.5)
plt.title("dynamic mouse: drifting truth (solid) vs one static fit (dotted)")
plt.show()
print("static fit visual =", round(w_static[1], 2),
" timing =", round(w_static[4], 2))
static fit visual = 1.27 timing = 1.22
The dotted static estimates sit flat while the true visual and timing weights cross right past them. The model isn't wrong — it's answering the wrong question, forcing one number onto a moving target.
Wrap-up¶
You built the strategy model from its likelihood, fit it by maximizing that likelihood, and recovered the true weights of single-strategy mice. You also saw the static model's blind spot: a mouse whose strategy changes.
Next (Notebook 4): before trusting any of these fits, we need to evaluate them — how well does the model actually predict held-out licking? That's the ROC curve, AUC, and cross-validation.