Earther

Researching Learning Systems | Deep Learning × Vision Mamba x Fusion Models

[Earther @EartherAI]

Day 57 of #100DaysOfML

Today was about generative classifiers—specifically Naive Bayes.

Key takeaways:

Modeling P(x∣y)P(x \mid y)P(x∣y) instead of P(y∣x)P(y \mid x)P(y∣x) changes how you think about learning
Class-conditional independence → simple likelihoods, strong baselines

MLE for Bernoulli & Gaussian NB is surprisingly intuitive

With equal covariance, Gaussian NB yields a linear decision boundary

Zero-frequency issues → Laplace smoothing is not optional

Simple assumptions. Strong theory. Still competitive in practice.

[Earther @EartherAI]

@yur_o_0 Great

[Earther @EartherAI]

@soydotrun @Amrita_Bh @liuhuan @SCAI_ASU Wow!

[Earther @EartherAI]

@chastronomic @WhatsApp Soo True

[Earther @EartherAI]

First round of 7-hour training is done. Next up: 6 hours of A100 experiments and analyzing the first phase results.

[Earther @EartherAI]

Training 2 models simultaneously on NVIDIA A100. Parallel experimentation, faster iteration, better research velocity.

[Earther @EartherAI]

@novasarc01 Wow! Congratulations 🎉🎉

[Earther @EartherAI]

Training 2 models simultaneously on NVIDIA A100. Parallel experimentation, faster iteration, better research velocity.

[Earther @EartherAI]

Day 65 of #100DaysofML

Finished Artificial Neural networks; Multiclass classification

-End of Machine Learning Theory

[Earther @EartherAI]

Day 64 of #100DaysofML

Complated theory of Ensemble methods - Bagging and Boosting (Adaboost)

[Earther @EartherAI]

Day 64 of #100DaysofML

Complated theory of Ensemble methods - Bagging and Boosting (Adaboost)

[Earther @EartherAI]

Day 63 of #100DaysOfML - Ensembles & the Bias–Variance Game

Generative vs Discriminative.
Weak → Strong learners.
Bagging → ↓ Variance (parallel, averaging).
Boosting → ↓ Bias (sequential, adaptive).

Overfit = high variance.
Underfit = high bias.
Smart aggregation > single model.

[Earther @EartherAI]

Day 63 of #100DaysOfML - Ensembles & the Bias–Variance Game

Generative vs Discriminative.
Weak → Strong learners.
Bagging → ↓ Variance (parallel, averaging).
Boosting → ↓ Bias (sequential, adaptive).

Overfit = high variance.
Underfit = high bias.
Smart aggregation > single model.

[Earther @EartherAI]

Day 62 of #100DaysOfML

Today cracking open the math behind Soft-Margin SVMs!

The real magic lies in the Dual Problem and Complementary Slackness.

🟢 The Safe Zone:
Correctly classified and safely past the margin.

🟡 The Support Vectors :
Sitting perfectly on the margin edge. These define the boundary!

🔴 The Margin Violators:
Points that are either inside the margin or completely misclassified.

It's how complex optimization math translates into such clear geometric intuition!

[Earther @EartherAI]

Day 62 of #100DaysOfML

Today cracking open the math behind Soft-Margin SVMs!

The real magic lies in the Dual Problem and Complementary Slackness.

🟢 The Safe Zone:
Correctly classified and safely past the margin.

🟡 The Support Vectors :
Sitting perfectly on the margin edge. These define the boundary!

🔴 The Margin Violators:
Points that are either inside the margin or completely misclassified.

It's how complex optimization math translates into such clear geometric intuition!

[Earther @EartherAI]

Day 61 of #100DaysOfML – Logistic Regression

Today was about the math behind Logistic Regression:

• Sigmoid maps scores → probabilities
• Likelihood → log-likelihood for numerical stability
• Objective = maximize log-likelihood (or minimize cross-entropy)
• Gradient: ∑ xᵢ (yᵢ − σ(wᵀxᵢ))
• Update rule: w ← w + η ∇logL
• Regularization adds λ/2 ‖w‖² to prevent overfitting
• Kernelization possible via w* = Σ αᵢ xᵢ
From probability model → optimization → gradients → regularization.

[Earther @EartherAI]

Reminds of this ... by @ilyasut

[Derya Unutmaz, MD @DeryaTR_]

The next several years will see the greatest destruction of human ego in history!

In the age of AI, there will be three big losers:

1) Those who maintain high ego & arrogance about their intellectual abilities;

2) Those whose careers depends on gatekeeping;

3) AI deniers.

[Earther @EartherAI]

Hard take: If India builds its thorium reactor within the next 2-5 years, it can become one of the world's superpowers.

Otherwise, the train has already left. This is the only visible move - move 37 - left for India.

[Earther @EartherAI]

@theJayAlto Metaphor ; Biology Meets Modern Life.

[Earther @EartherAI]

@learner_03038 Thank you!

[Earther @EartherAI]

Day 61 of #100DaysOfML – Logistic Regression

Today was about the math behind Logistic Regression:

• Sigmoid maps scores → probabilities
• Likelihood → log-likelihood for numerical stability
• Objective = maximize log-likelihood (or minimize cross-entropy)
• Gradient: ∑ xᵢ (yᵢ − σ(wᵀxᵢ))
• Update rule: w ← w + η ∇logL
• Regularization adds λ/2 ‖w‖² to prevent overfitting
• Kernelization possible via w* = Σ αᵢ xᵢ
From probability model → optimization → gradients → regularization.

[Earther @EartherAI]

Day 60/100 #100DaysOfML
Diving into Support Vector Machines intuition!

From perceptrons to max margins for better classifiers. Here's a structured breakdown:

Perceptron Foundation: Mistake bound is # mistakes ≤ R²/γ², where γ is the margin. Larger margins = fewer errors & better generalization.

Goal: Formulate optimization to directly maximize the margin, avoiding small-margin pitfalls.

Key Derivations:Normalize weights: Set ||w||=1, then maximize γ s.t. y_i (w·x_i) ≥ γ for all i (avoids scaling issues).

Equivalent form: Fix functional margin to 1, minimize (1/2)||w||² s.t. y_i (w·x_i) ≥ 1.
Geometric margin: Simplifies to 2/||w||.

[Earther @EartherAI]

@thepanshu_logs Hey I have been working on a paper too can you suggest the platform for digram making.

[Earther @EartherAI]

@saurabhtwq I have to try this out!!

[Earther @EartherAI]

From tomorrow back to maths bulid UP!!

[Earther @EartherAI]

Day 60/100 #100DaysOfML
Diving into Support Vector Machines intuition!

From perceptrons to max margins for better classifiers. Here's a structured breakdown:

Perceptron Foundation: Mistake bound is # mistakes ≤ R²/γ², where γ is the margin. Larger margins = fewer errors & better generalization.

Goal: Formulate optimization to directly maximize the margin, avoiding small-margin pitfalls.

Key Derivations:Normalize weights: Set ||w||=1, then maximize γ s.t. y_i (w·x_i) ≥ γ for all i (avoids scaling issues).

Equivalent form: Fix functional margin to 1, minimize (1/2)||w||² s.t. y_i (w·x_i) ≥ 1.
Geometric margin: Simplifies to 2/||w||.