"What distribution do you think this is?"
Mira showed Yuki a graph. Data points were scattered.
"I don't know. But how do we find out?"
Aoi joined the conversation. "That's the essence of statistical learning. Estimating distributions from data."
"Estimating?"
"Inferring the underlying probability distribution from observations."
Professor S. entered the club room. "Interesting challenge. This is the fundamental problem of inference."
The four gathered around the table.
Aoi began explaining. "First, assume a distribution form. Normal distribution, Poisson distribution, or something else?"
"But how do we choose?" Yuki asked.
"Look at data properties. Continuous or discrete? Bounded or unbounded range?"
Mira listed data characteristics. Continuous values, symmetric, no negative values.
"Normal or gamma distribution are candidates," Aoi judged.
"Next is parameter estimation," Professor S. continued. "Even deciding the distribution form doesn't tell us mean or variance."
"How do we estimate?"
"Maximum likelihood is common. Choose parameters that maximize the probability of obtaining observed data."
Aoi wrote an equation on the whiteboard.
"L(θ|x) = Π p(x_i|θ). Likelihood function. Find θ that maximizes this."
"Taking logarithm makes computation easier," Professor S. supplemented. "log L(θ|x) = Σ log p(x_i|θ)"
Yuki attempted calculation. "If we differentiate this and set it to zero..."
"Correct. In many cases, analytically solvable."
Mira presented another perspective. Bayesian inference equation.
"There's also Bayesian approach," Aoi explained. "Assume prior distribution, update with data."
"Prior distribution?"
"Initial belief about parameters. Pre-observation knowledge."
Professor S. explained in detail. "p(θ|x) ∝ p(x|θ) p(θ). Posterior distribution is proportional to product of likelihood and prior."
"How is this different from maximum likelihood?" Yuki asked.
"Maximum likelihood estimates parameters as fixed values. Bayesian treats parameters as probability distributions."
Aoi supplemented. "Bayesian can explicitly express uncertainty. Like 'I think the mean is 5, but could be 4 to 6.'"
"Which is better?"
"Depends on situation," Professor S. answered. "If you have prior knowledge, Bayesian is powerful. Without it, maximum likelihood is simpler."
Mira applied maximum likelihood to the data. Normal distribution with mean 7.2, standard deviation 2.1.
"But," Yuki said. "There's no guarantee this is correct, right?"
"Sharp. Estimation always involves uncertainty," Aoi acknowledged.
"That's why we calculate confidence intervals," Professor S. continued. "A range around the estimate that contains the true value."
"A 95 percent confidence interval contains the true value 95 times out of 100 estimates."
Yuki understood. "Quantifying uncertainty."
"Yes. That's the beauty of statistical inference."
Aoi presented a new perspective. "What if we don't even know the distribution form?"
"Non-parametric methods," Professor S. answered. "Like kernel density estimation. Don't assume distribution form."
Mira drew a complex diagram. Hills overlapping around data points.
"Place small distributions centered on each data point. Add them together," Aoi explained.
"Flexible, but needs lots of data."
Yuki asked. "Is machine learning the same as this?"
"Fundamentally the same," Professor S. nodded. "Neural networks also learn distributions from data."
"But estimation in ultra-high-dimensional space," Aoi supplemented. "For images, the number of pixels becomes dimensions."
"Staggering," Yuki murmured.
"But with abundant data, estimation is possible. That's the power of modern machine learning."
Mira looked at the original graph. "The true distribution might never be known."
"Perhaps," Professor S. said quietly. "But the more data we collect, the closer we get to truth. That's learning."
Aoi said. "Chasing an unknown distribution. That's the journey of seeking truth itself."
Yuki looked out the window. "Phenomena everywhere in the world are unknown distributions."
"Yes. And as observers, we continue learning from data."
The four continued quietly gazing at the data, like travelers chasing invisible probability distributions.