Scribd
Upload
3 views2 pages

Notes

Intelligence is defined as the ability to generalize knowledge from past experiences to new situations, contrasting traditional software with artificial intelligence in terms of rule discovery. The document outlines the taxonomy of AI, distinguishing between Artificial Narrow Intelligence (ANI) and theoretical constructs like AGI/ASI, emphasizing the importance of generalization in machine learning. It also discusses the two types of data—unlabeled and labeled—highlighting the advantages of using unlabeled data for pre-training models to enhance their robustness and ability to recognize patterns.

Uploaded by

indianthatha93
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
3 views2 pages

Notes

Intelligence is defined as the ability to generalize knowledge from past experiences to new situations, contrasting traditional software with artificial intelligence in terms of rule discovery. The document outlines the taxonomy of AI, distinguishing between Artificial Narrow Intelligence (ANI) and theoretical constructs like AGI/ASI, emphasizing the importance of generalization in machine learning. It also discusses the two types of data—unlabeled and labeled—highlighting the advantages of using unlabeled data for pre-training models to enhance their robustness and ability to recognize patterns.

Uploaded by

indianthatha93
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd

Intelligence

Intelligence is not about memorizing facts. Intelligence is the ability to generalize. It means taking
what you learned from past experiences and correctly applying it to a brand-new situation you have
never seen before.

Traditional Software: Data + Human Rules = Answers

Artificial Intelligence: Data + Answers = The Machine Discovers the Rules

Taxonomy of AI

Artificial Narrow Intelligence (ANI): Optimization of a specific, singular objective


function. It operates within a tightly constrained state space (e.g., predicting the next token,
classifying an image, calculating a chess move). It cannot transfer learning from one domain
to another without retraining.

AGI / ASI: Theoretical constructs where an agent possesses cross-domain transfer learning,
dynamic goal-setting, and human-level reasoning capabilities. Not relevant to current
engineering.

Generalization

When teaching a machine, our goal is for it to find the real pattern, not just memorize the
data. The primary objective of any AI model is not to memorize the training data, but to generalize to
unseen data drawn from the same distribution.

 Generalization: Looking at a green apple and knowing it is an apple because of its


shape, even if all the apples you saw before were red. The model successfully learns the
underlying decision boundaries or feature representations. It ignores minor anomalies and
applies the core mathematical pattern to new, unseen inputs.

Data
In technical terms, data refers to raw facts, figures, symbols, or observations that can be
recorded, stored, and processed by a computer. On its own, data is typically unorganized and
lacks specific meaning; it only becomes information once it has been processed, structured,
and interpreted within a relevant context. In simply Data is frozen experience.
The Two Types of Data

 Unlabeled Data: Raw information. For example, a massive folder of 10,000 random
photos with no descriptions. It is cheap and easy to find on the internet. Used in
Unsupervised Learning. To discover hidden patterns, clusters, or structures within the data.

 Why Start with Unlabeled Data?

Models can be pre-trained on massive amounts of unlabeled data to learn basic features, like
shapes in images or grammar in text, before they are ever shown the specific "answers"
(labels). Labeled data is often a "clean" but limited subset. Unlabeled data is
everywhere and represents the messy, diverse reality the model will face. By looking
at unlabeled data first, the model learns to handle variations like different lighting in
photos or various accents in speech, which makes it much more robust.

Self-Training: The model learns from the labeled data first, then "guesses" labels for the
unlabeled data (called pseudo-labeling). It then adds the most confident guesses to its training
set to get smarter.

Clustering Guidance: The model uses unlabeled data to find natural groups (clusters). If a
few items in a cluster have labels, the model can safely assume everything else in that cluster
belongs to the same category.

 Labeled Data: Information that has an "answer key" attached to it by a human. For
example, a photo of a dog where a human has explicitly typed the word "Dog" on it.
Labeled data is incredibly expensive and time-consuming to create, but it is the "gold"
of the AI industry. To train a model to predict or classify future, unseen data.

You might also like