cards-deck: 100-199_university::111-120_theoretic_cs::116_introduction_machine_learning

Overview of Machine learning

anchored to 116.00_anchor_machine_learning may requires prior knowledge from 105.00_math_stochastic and likely also information about 111.00_anchor

Overview

Some general ideas for the concept of machine learning compose:

[!Quote] Arthur Samuel - 1959 Field of study that gives computer the ability to learn without being explicitly programmed

[!Quote] Learning is any process by which a system improves performance from experience

Machine Learning is concerned with computer programs that automatically improve their performance through experience ~ Herber Alexander Simon - Turing Award 1975, Nobel price (business) 1978 )

Generally speaking: Machine learning is not Voodoo -> but is to use data to automatically find a suitable function or algorithm to solve a given task.

Below we may give an overview on how machine learning could be taken to use to solve tasks:

[!Example] possible usecases:

1. Weather forecast We would like to have a system that is able to predict the weather / temperature on the next day, based on data supplied –> previous information it obtained may result with: temperature tomorrow = $F(\text{ measurements of the last days})$
2. Disease diagnosis: Finding out whether a person is ill or not based on some information (images of tissues, blood data etc.) attempts to result with a function: $P(\text{ diseaseX})=F(\text{Symptoms, Measurements, Medical history})$ –> probability
3. Chat Bots predicting the outcome of for the next word(s) to answer a request $wor{d}_{next}=F(\text{ convo hist})$

[!Tip] (Oliver G. Selfridge) Find a bug in a program and fix it and the program will work today. Show the program to find and fix a bug and the program will work forever

-> at least thats expected, but not feasible xd

Historic scope:

• back in the 1945s the first artificial intelligence structures were discussed -> ranging from being a simple computer to play go against or such.
• From 1980s Machine learning was flourishing -> coming into play are the aspects of
• with advancements in the field the whole topic of deep learning was developing fast starting in the 2010s

[!Tip] The friendship algorithm (TBBT)

[!Idea] Base concept of ML regarding available data and outputs #card?

The general idea to resolve problems in ML is by doing the following: Input: Data and the desired result - I want you to find cats in images Output: An algorithm that does this -I will find cats by doing things…

$output=h(input)$ -> we reverse the usual approach to work with information/data

Supervised Learning

With Supervised learning we are partaking in one of three large categories for machine learning. Here particularly we have controlled learning of a machine because our data contains labels and information about the desired results of said data.

[!Definition] Supervised learning | concept #card

With supervised learning we are adapting/modifying the core principal of machine learning a little: input: data points $\{(x,y{)}_i\}$ where: $x$ is the input(data) and $y$ is the desired output - a label classifying what it should result with we have a space of function $\mathcal{H}$ with elements $h(\cdot )$ Objective: We would like to find a function $h(\cdot )$ such that: $$h(x)\approx y,\forall (x,y)$$ for this given task –> we want to find / discover the function that is best at approximating the desired output for a given input -> to assure this: we measure the quality of this function $h$: $dist(y,h(x))$

proceeding to minimize the loss alltogether: $$h^{\ast }=\arg \min {\mathbb{E}}_{data}[dist(y,h(x))]$$ ^1721145490688

This may be done in the following subjects:

• recognizing handwritten digits
• classifying cells -> pathology images
• recognizing faces or objects in an image / video-stream
• language modelling
• regression

Unsupervised Learning

Once again we strive for the base principle of machine learning: having data and wanting to find a function that best interprets it in a certain way.

[!Definition] Unsupervised Learning | basic idea #card

with unsupervised learning we don’t use labels on given data - as it was in supervised learning. We have the following constraints / basics: input data points $\{(x{)}_i\}$ with $x\in {\mathbb{R}}^n$ ( a vectorspace!) once again we are given a function space $\mathcal{H}$ with $h(\cdot )\in \mathcal{H}$ and we would like to find a function $h$ so that: $h(x)⟼$ where $y$ is the low-dimensional representation –> This means we are reducing dimensions from high vector input (x is a vector of ${\mathbb{R}}^n$!) to a smaller space while still maintaining / not losing information on its traits - like its similarity to the original data point in this new representation we then assign each $x$ to a given cluster $y\in {\mathbb{N}}^{+}$

Objective: The objective is not directly defined, its rather the goal to generally reduce the dimension of data somehow reliably - how exactly is not defined. –> there are many algorithms that could be found to solve this task ^1721145490694

This may be used in the following subjects:

• Genome comparisons ( this is clustering)
• Finding descriptors for face expressions ( dimension reduction –> from a high dim ( the face) to a lower -> Cartesian mapping of emotions)
• Finding disentangled generating factors ( take image of faces and find few descriptive variables – generating new maybe)

Reinforcement Learning

Is a little different in its idea compared to the previous to attempts to train machines. Here we utilize a reward system to have the system train itself and gain knowledge about performance and such.

[!Definition] Reinforcement learning in its core concept? #card we provide a system to interact with: $s_{t+1}=Sys(a_t,s_t,t)$ -> $s_t$ is the state, $a_t$ is an action ( in time $t$)

We have the Function space again $\mathcal{H}$ with $h(\cdot )$ that denotes a (policy) deciding about actions to take based on the current state: $$a=h(s)$$

to evaluate / refine later we introduce a rewarding/utility function: $$r_t=U(a_t,s_t)$$

Objective: we ought to find the function $h(\cdot )$ that maximizes the expected reward $\mathbb{E}[r]$

Generally speaking these stochastic systems are described with Markov Decision Processes.

Tasks of models in reinforcement learning are therefore:

• collect own data – to develop on / with
• simultaneously learn $h$ and potentially models of $Sys$ and $U$ (utilty funct and the system to interact with)
• the reward can be sparse – only at the end of an long action sequence or such ^1721145490696

Reinforcement learning can be deployed in various fields - obviously - like:

• robot control and movements
• Deepmind AlphaGo Robot -> game bots
• walking simulations

Usage of ML

Generally speaking we have a large objective with ML:

[!Definition] To solve problems where we do not have good algorithsm but data (a lot!)

It is involved in many many fields nowadays: daily life

• sorting pictures
• predicting what will be done (advertisment / content recommendation)
• diagnosis assistant in health
• access to knowledge -> LLM science:
• automatic processing of experimental data
• new tools for searcing in fast spaces - proetin confirmations or similar

Data privacy and security:

We ought to maintain caution of the origin of our used data: -> Machine learning systems are utilizing and working with data a lot hence its important to select and source it well

Could create:

• bias due to gender/race/ethnicity
• exploit information of people etc

Other fields:

• psychology
• computer science …
• statistics