Before's choice of model describes the task: what receives, what returns and how you will know if it works. That habit prevents using AI where a rule and helps to compare different solutions.
Objectives: To distinguish AI, ML and DL; To recognize the tasks of language, audio and vision. separate generation classification and evaluate an exit without treating his confidence as certainty.
Key concepts
Artificial Intelligence (AI) It's the field that builds systems capable of performing tasks associated with perception, reasoning, language or decision. Machine Learning (ML) Learn patterns from data. Deep Learning (DL) is a family of ML that uses multi-layered neural networks to learn complex representations. Thus, DL's inside ML and ML's inside AI.
The Generative AI production of text, images, audio or code. Many today's generating systems have been using deep learning. It doesn't mean that they generate without predicting: a LLM can build text by predicting successive tokens. A profound network that classifies photographs remains DL even if it doesn't generate images.
A solution specialized in transcribing calls doesn't prove general intelligence. AGI describes an overall and comprehensive capacity To recognize objects, by itself, doesn't credit them. It's also appropriate to separate AI from data scienceThat includes exploring data, designing experiments and communicating results with or without ML models.
Tasks and data types
| Input | Task | Departure that you have to recognize |
|---|---|---|
| The text of a claim | Classify theme or feeling | Label or score |
| The text of a policy | Generate a Summary | New text using the source |
| Audio of an call | Voice recognition, ASR | Transcription |
| Photography of equipment | Object detection | Labels and delimiting boxes |
| Invoice scanned | Extract fields and tables | Data structured |
| Reparations History | To estimate duration | Value numeric |
Data structured have an outline, as an order table. The Unstructuredas photos or free text, they also have an internal structure but they are not organized into business columns. JSON is a regular example of semistructured data.
The text becomes tokens and numeric representations. The audio is represented with time samples; Sample frequency and bit depth describe different aspects. A image contains pixels and channels: an isolated pixel rarely identifies the object. The context and relations between elements matter.
To choose between rules and learning
A rule as "code should have eight characters" does not require training. ML is useful when recognizing the pattern with manual rules would be difficult and appropriate data are available. Learning supervised there are examples with answers. in unsupervised an exploration of a structure without these labels; in strengthening decisions are learned from interplay rewards.
The name of the application does not determine a single method. A standalone vehicle can combine monitored perception, planning and other components. An anomaly doesn't prove fraud either: it only points to something that deserves review.
Step-by-step example
A receives company "The computer goes back out" with a photo and an audio. He wants to classify the problem and write an initial response.
- Separate inputs: message, image and recording. Don't treat all the files as text.
- He transforms his audio so that he can look and analyze it. Keep the original to check errors.
- Classifies text as a category If you need to figure out a piece in the photo, use an item detection.
- Query's authorized guarantee policy and generates an explanation supported by it.
- Check dubious data and decide whether human intervention is necessary before an action.
The system generates an answer, but that doesn't mean that you have passed a guarantee. The decision and the permissions belong to the business process.
Common errors
- To merge a label, as an "engine," with a new image of an engine.
- To assume that all AI uses profound networks or that every profound network is generative.
- Read a high confidence score as a correction guarantee in any situation.
- To think that sending a message to a chatbot automatically modifies model weights.
- To say that restoring images is never AI: there are models that do that task. That's the same as categorizing or detecting.
Selection exercise
A workshop wants to check a serial format, detect a piece in a photo and write an understandable notice. Do you have the same model for anything?
Solution
It uses determinist validation for the serial format, vision to locate the piece and text generation to write the notice from verified data. Keep the three outputs separate: valid format does not prove that the piece exists, and convincing text does not confirm a repair.
Sources and review
Review AI Foundations at official route. It contrasts the tasks of images with OCI Vision. Before to advance, gives an example of its own classification, regression and generation.