AI 101
HMIA 2025
AI 101
HMIA 2025
Goals
Demystify some terms. Fill in some gaps. Not exactly "big picture" but "starting picture." Meet 30 people at 30 different places. Apologies to those who know way more than I do.
THERE ARE NO SILLY QUESTIONS.
I don't have all the answers.
AI 101
HMIA 2025
"Readings"
PRE-CLASS
CLASS
Outline
- Programing vs Machine Learning
- AI landscape
- How machines learn
- Training/Deployment; pretraining/finetuning; supervised/unsupervised; RL/RLHF; loss & optimization
- Perceptron, neuron, layers, forward, backward
- Accuracy, precision, reliability, validity
- Bias/fairness
- Alignment
HMIA 2025 AI Bootcamp
HMIA 2025 AI Bootcamp
HMIA 2025 AI Bootcamp
PRE-CLASS
HMIA 2025 AI Bootcamp
PRE-CLASS
HMIA 2025 AI Bootcamp
PRE-CLASS
HMIA 2025 AI Bootcamp
PRE-CLASS
Intro to Machine Learning (ML Zero to Hero - Part 1) [7m17s]
HMIA 2025 AI Bootcamp
PRE-CLASS
HMIA 2025 AI Bootcamp
PRE-CLASS
HMIA 2025 AI Bootcamp
Text
3Blue1Brown: Large Language Models Explained Briefly [7m57s]
PRE-CLASS
HMIA 2025 AI Bootcamp
PRE-CLASS
HMIA 2025 AI Bootcamp
PRE-CLASS
History
HMIA 2025 AI Bootcamp
CLASS
Models
HMIA 2025 AI Bootcamp
CLASS
Models
HMIA 2025 AI Bootcamp
CLASS
\hat{Y} = m X + b
m
b
X→
→Y
a black box with two dials
prediction
data
parameters
error/loss
Brains Can Be Built of Biological Neurons
HMIA 2025 AI Bootcamp
CLASS
Inputs
Output
f(\text{dendrites})
dendrites
axon
nucleus
Models Can Be Built of Artificial Neurons
HMIA 2025 AI Bootcamp
CLASS
Inputs
Output
1
0
f(1,0)
Models as Functions
HMIA 2025 AI Bootcamp
CLASS
Input
Hidden Layers
Output
X_1
X_2
X_3
X_4
X_5
X_6
\hat{Y} = f(x_1,x_2,x_3,x_4,x_5,x_6)
weights (that's what's "learned")
Deep (multiple hidden layers)
Neural (each node an artificial neuron)
Network
Why Linear Algebra?
HMIA 2025 AI Bootcamp
CLASS
Input
Hidden Layers
X_1
X_2
X_3
X_4
X_5
H_1 = \begin{bmatrix}
x_1 \\
x_2 \\
x_3 \\
x_4 \\
x_5
\end{bmatrix}
h_{11} = x_1 w_{11} + x_2 w_{12} + x_3 w_{13} + x_4 w_{14} + x_5 w_{15}
w_{11}
w_{12}
w_{13}
w_{14}
w_{15}
\begin{bmatrix}
w_{11} & w_{12} & w_{13} & w_{14} & w_{15} \\
w_{21} & w_{22} & w_{23} & w_{24} & w_{25} \\
w_{31} & w_{32} & w_{33} & w_{34} & w_{35} \\
w_{41} & w_{42} & w_{43} & w_{44} & w_{45} \\
w_{51} & w_{52} & w_{53} & w_{54} & w_{55}
\end{bmatrix}
Cat Detector as Function
HMIA 2025 AI Bootcamp
CLASS
=(12,98,231,...)
(0,1)
point in super high dimensional space
f
Reinforcement Learning
HMIA 2025 AI Bootcamp
CLASS
STOP+THINK
What else about learning to ride a bike strikes you as interesting? Do you think it's important that it's fun? Is it important that mom is there?
Living is One Fork in the Road After Another
https://devblogs.nvidia.com/deep-learning-nutshell-reinforcement-learning/ (Video courtesy of Mark Harris, who says he is “learning reinforcement” as a parent.)
Agent Models
Reinforcement Learning with Human Feedback
(RLHF)
Remember the other day when we talked about having "other people in our head"?
Train your model to produce outputs from inputs.
Let it run on a lot of different inputs and rate the outputs.
Build a model that predicts the rating of each output, given a particular input.
IMAGES
|
Try out everything over and over and over again and keep track of how everything works out each time.
The Plan
What can models do?
HMIA 2025 AI Bootcamp
CLASS
Classify: "look" at an item and put it in a category
Predict: "look" at some information and say what comes/happens next
Optimize: "look" at situation, objective, and resources and choose best solution
Generate: "look" at a description, create a text or image or sound
Translate: "look" at something in one form and transform it to another form
Networks of Neurons
HMIA 2025 AI Bootcamp
CLASS
Imagine you are a worm learning how to move around in the world. You are constantly bumping into things. Some things are sharp and threatening. You can defend yourself against sharp, threatening things by curling into a ball. But if you curl into a ball everything something touches you, all the other worms will laugh and you'll never get anywhere.
So you have this neuron that responds to touch
and drives the curl-up action in response. How does the worm train the neuron to respond to the proper levels of touch (only curl up when you encounter a sharp and threatening touch)?
Let's suppose the neuron works like this: the input is a measure of the touch; this is compared to a threshold and if it is higher, the neuron triggers a curl-up action.
Initially, the threshold is zero: the worm curls up at the slightest touch. This happens a few times, touches of level 1 and level 2 and level 9, and the worm notices that nothing happened, the touch did not cause any harm. So, it raises the threshold to 1. Now when it gets a level 1 touch it just keeps moving. But level 2 touches cause it to curl up. And then some 5s and 6s and it curls up. Again, it notices that nothing happened so it raises the threshold to level 2.
This continues for a bit and it has raised the threshold to level 6. Then
one day, it gets a level 5 poke that really hurts — a sharp thorn or a
hidden barb. So the worm lowers its threshold by 1 to level 5. Now it curls up when it feels a level 5 poke. For a while, it seems to be avoiding injury quite well so it lowers the threshold to 4. Then it notices that sometimes a level 5 poke is painful but sometimes not and so it leaves the threshold at 4. The worm has learned an approximate rule: curl up when things feel like level 5 or higher — but not always, and not perfectly.
HMIA 2025 AI Bootcamp
CLASS
We all know about the game of guessing the number of jelly beans in a jar. And if you have taken a statistics you have heard of the "wisdom of the crowd" : we can ask a bunch of people to estimate and then we take the average of their guesses.
Could we teach a machine to learn how much to
trust each person’s guess — and adjust based
on feedback?
ROUND 1. Present the jar, record the guesses,
compute the average, reveal the actual number of
jelly beans.
Remind ourselves that computing the average is
the same as computing a weighted sum where the weights are all 1/n where n is the number of people guessing. We count the number of folks in the room
and compute 1/n. "This is your initial weight."
ROUND 2. We ask for show of hands: how many had a guess that was high? Low? If your guess last round was higher than the actual number, subtract 0.05 from your weight. If it was lower, add 0.05.
If you were spot-on, leave it as is.
Now we introduce a new jar and ask folks to write down their guess.
We have folks do a little math - multiply their
guess times their weight - and we add up all the results.
Then we reveal the actual number of jelly beans.
ROUND 3. We repeat the process.
STOP+THINK. What's going to happen to folks who consistently guess too low?
We can introduce some limits like "your weight can never go above 1 or below 0. Or, after we increment or decrement everyone's weights we could adjust them all so they still add up to 1. This would preserve the math of the weighted average.
STOP+THINK. Would it be better to add/subtract more weight from people who were really far off, and less from those who were just a little off?
HMIA 2025 AI Bootcamp
CLASS
HMIA 2025 AI Bootcamp
\frac {1}{5} \times 200 + \frac {1}{5} \times 150 +
\frac {1}{5} \times 250 +
\frac {1}{5} \times 250 + \frac {1}{5} \times 150 = 200
\text {Actual} = 220
\begin{aligned}
& \text{New Weights: } 0.25, 0.25, 0.15, 0.15, 0.25
\end{aligned}
\text {Actual} = 400
\begin{aligned}
& \text{New Guesses: } 300, 350, 450, 500, 500
\end{aligned}
\begin{aligned}
& \text{New Weighted Average } \\ &= 300 \times 0.25 + 350 \times 0.25 + 450 \times 0.15 + 500 \times 0.15 + 500 \times 0.25 \\
&= 75 + 187.5 + 67.5 + 75 + 125 = 455
\end{aligned}
\begin{aligned}
& \text{New Weights: } 0.3, 0.3, 0.1, 0.1, 0.2
\end{aligned}
\text {Actual} = 150
\begin{aligned}
& \text{New Guesses: } 100, 150, 250, 200, 300
\end{aligned}
\begin{aligned}
& \text{New Weighted Average } \\ &= 100 \times 0.3 + 150 \times 0.3 + 250 \times 0.1 + 200 \times 0.1 + 300 \times 0.2 \\
&= 30 + 45 + 25 + 20 + 60 = 180
\end{aligned}
\begin{aligned}
& \text{Guesses: } 200, 150, 250, 250, 150
\end{aligned}
low
high
high
low
low
CLASS
HMIA 2025 AI Bootcamp
\begin{aligned}
& \text{Guess: } 100
\end{aligned}
\begin{aligned}
& \text{Guess: } 300
\end{aligned}
\begin{aligned}
& \text{Guess: } 200
\end{aligned}
\begin{aligned}
& \text{Guess: } 250
\end{aligned}
\begin{aligned}
& \text{Guess: } 150
\end{aligned}
\begin{aligned}
& \text{Weight: } 0.3
\end{aligned}
\begin{aligned}
& \text{Weight: } 0.2
\end{aligned}
\begin{aligned}
& \text{Weight: } 0.1
\end{aligned}
\begin{aligned}
& \text{Weight: } 0.1
\end{aligned}
\begin{aligned}
& \text{Weight: } 0.3
\end{aligned}
\begin{aligned}
& \text{New Weighted Average } \\ &= 100 \times 0.3 + 150 \times 0.3 + 250 \times 0.1 + 200 \times 0.1 + 300 \times 0.2 \\
&= 30 + 45 + 25 + 20 + 60 = 180
\end{aligned}
CLASS
Models as Functions
HMIA 2025 AI Bootcamp
CLASS
Input
Hidden Layers
Output
X_1
X_2
X_3
X_4
X_5
X_6
\hat{Y} = f(x_1,x_2,x_3,x_4,x_5,x_6)
Models
HMIA 2025 AI Bootcamp
CLASS
HMIA 2025 AI Bootcamp
CLASS
HMIA 2025
HMIA 2025 AI Bootcamp
By Dan Ryan
