BA Internet · From data to model
A data set does not convey the world as it is; A model does not produce the truth itself.
Data is selected, named, subtracted and affected by its context. The model learns patterns from this man-made representation and produces new outputs. The BAI approach does not view the data as orphaned raw material and the model as a ruling mind; It keeps the source, right, privacy, deficiency and human responsibility visible at every stage.
Source · Selection · Permission · Context · Underrepresentation · Redaction
Basic distinction
Data is a trace; the data set is a selected representation; The model produces probability and output based on this representation.
When we reduce a person to a few behavioral records, a city to sensor numbers, a language to texts found on the internet, or the land to measured values, we may lose the essence that remains. Therefore, knowing that data is not the same as existence and model is not the same as truth; It is not a technical detail, but the beginning of the right decision.
What we can measure is a trace of reality; It is not the whole truth.
Truth and representation
The real thing and the representation we record and calculate are not of the same rank.
Human, living being, society and field
Contains substance and context
- It has immeasurable meanings and relationships
- Varies with time, place and condition
- It carries rights, dignity and privacy
- She has her own say and the right to object
Dataset and model
Produces elected representation
- What is recorded is human choice
- May be missing, outdated or unstable
- Works according to certain purposes and criteria
- Requires re-examination in a new context
Four record types
Read the technical words along with the human choice they carry.
This distinction makes record keeping easier; no record in itself confers credibility or publication authorization.
Data
A recorded trace of a person, event, text, device, or site. It is important to whom it belongs, in what context and with what permission it was created.
Dataset
A set of data selected, cleaned, labeled and organized for a specific purpose. It shows what is left out as well as what is included.
Algorithm
Steps and rules that determine how input will be processed. It is a human decision which criterion is highlighted and which error is accepted.
Model
A representation that learns patterns from data and method and produces prediction, classification or new output. Appearing fluent and successful does not give authority to judge.
From life to model
At every stage there is a choice of people; These choices must be visible and correctable.
Before using the dataset or model
Answer the same six questions clearly.
Where did this recording come from?Are the source, date, method of collection, creator and modification history known? Who has the right and permission?Individual, child, family, community, work and institution rights; Are the terms of use and withdrawal clear? Who is invisible or misrepresented?How are language, geography, age, gender, occupation, access and other field differences missing from the data? Where does the model go wrong?It's not just the success rate; Are examples of false confidence, spurious results, discriminatory outcomes, and high-risk errors recorded? Which decision remains with the person?Are the place where the model provides support and the competent person's place of verification, acceptance, rejection, publication and responsibility separated? Is there any way to appeal and rectify?Can people see, correct or delete their own data? Can a model error be reported and use stopped?
Sharing and access decision
Not every data set or model has to be in open circulation.
SEDD and İHSAN are not a later publication stamp. When purpose, rights, damage and care are evaluated together, four different decisions can be made.
Never collect or use
Generating data and building the model if there is no real need, privacy or risk of harm cannot be justified.
Keep it private and limited
Establish a narrow use within the household, school, profession or institution, to which only qualified and authorized persons have access.
Share conditionally
Grant controlled access with specific purpose, age, duration, license, attribution, human consent and misuse limits.
Hungry for common good
Make a WAQF candidate with a dated and correctable record of rights, permissions, privacy, resources, security and maintenance responsibilities, if appropriate.
Joint representation
Turning a community into data without its words, language and competence is not being a NATION.
Joint data and model production is important not only for the technical team; It requires the people represented, native speakers, professionals, young people and caregivers to work together.
Main language and concept
A language is not just about words to be translated. The concept, context, history and living usage are recorded with native speakers.
Professional and field license
From medicine to agriculture, evaluate what the data shows and what it hides with the experts in the field.
The word of the person represented
People are not the only source of data; They can see the classes established for them, object and participate in the correction process.
Intergenerational care
Consider who the data and model established today will impact tomorrow; Do not leave the responsibility for updating, deleting and maintenance unattended.
Data and model review
Do you want to examine a dataset or model idea together in terms of source, rights, representation and human decision?
Within the scope of BAI, joint data and model review rings are still in the preparation phase. You can share the real need, the data source, the person represented, and the benefit or risk you see. This call is not an active model catalogue, data service or program promise.