Slice

Created: 23 Dec 2022, 02:27 PM | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a") Tags: knowledge,

Then run slice == mixture model

domino = DominoSlicer(

y_log_likelihood_weight=40,

y_hat_log_likelihood_weight=40,

n_mixture_components=60,

n_slices=10,

max_iter=10,

n_pca_components=128,

init_params=“confusion”,

confusion_noise=3e-3

)

domino.fit(data=dp, embeddings=“clip(image)”, targets=“target”, pred_probs=“prob”)

dp[“domino_slices”] = domino.predict_proba(

data=dp, embeddings=“clip(image)”, targets=“target”, pred_probs=“prob”

)

What does error-aware mean?

Mixture model == GMM? What’s the diff?

How does it get the slices?

Formulate the MM using input embeddings (clip image), class labels (target), model predictions (prob)
Embeddings modelled as Gaussians, labels and predictions modelled as categoricals
Get the log-likelihood, maximise using Expectation-Maximization (EM)
Where EM is an approach for Maximum Likelihood Estimation (MLE) ⇒ EM is better than MLE, where MLE may not be able to find closed form solution, EM can.
- https://machinelearningmastery.com/expectation-maximization-em-algorithm/
- See subpage
EM is done in fit_predict

What does fit() do?

MixtureSlicer.fit() calls DominoMixture.fit()
DominoMixture.fit() calls DominoMixture.fit_predict()
DominoMixture.fit_predict() does the E step and M step in for n_iter in DominoMixture._e_step and DominoMixture._m_step
Then after EM algo, will have gotten… what? The clusters?

DominoSlicer ⇒ MixtureSlicer, MixtureSlicer.mm ⇒ DominoMixture

What is MixtureSlicer.predict_proba doing?

Gives probabilistic slice assignments / clusters
Calls DominoMixture.predict_proba, which calls _estimate_log_prob_resp that
- Estimates log probabilities and responsibilities for each sample.
- Compute the log probabilities, weighted log probabilities per component and responsibilities for each sample in X with respect to the current state of the model.
- Will return the exponent of log responsibilities:

_, log_resp = self._estimate_log_prob_resp(X, y, y_hat)

return np.exp(log_resp)

fit()Learn a set ofslicing functionsthat partition the embedding space into regions with high error rates.
predict()Apply the learned slicing functions to data, assigning each datapoint to zero or more slices.
predict_proba()Apply the learned slicing functions to data, producing “soft” slice assignments.