Model — Not All Program

Content and key concepts

School numbers and functions
First we separate the computational model itself from the search, memory, and tools around it.
The model calculates the answer using numerical settings. The program around it prepares the entrance and can trigger a search or a tool.
Introductory question
→
Numerical
→
Result
Document: External source
Instrument: Separate action
Application memory: Recording outside balance
One interface can combine different mechanisms. The answer itself does not prove that it has been calculated by a trained neurosweet; the source and operation of the program must be distinguished.
The answer does not reveal the mechanism
Let the program say, "Two notebooks cost six units."
output
, but not yet a description of the reason. The phrase could have been written in advance by the developer; price 3 could be found in the table and the multiplier function transferred; the number could be evaluated by the statistical model; finally, the application could connect the search, model and calculator. The same text is compatible with different computational graphs.
The first rule of diagnosis follows:
You can't restore the internal mechanism only by the truth of one answer. You have to look at data, operations, state and boundaries of the system.
This rule protects against two symmetric errors. The first is to call the neuronet any complex interface. The second is to believe that a system without a neuronet cannot solve an intellectually defined problem. Finding a path in a graph, planning by the rules and proof in a formal system may be related to AI tasks, although the parameters of such an algorithm have not been studied on dataset.
AI, ML and DL: multiple methods, not ratings
Artificial Intelligence, AI
- An area of methods and systems for recognizing, finding solutions, planning, managing, and processing languages. The definition is deliberately broad: it describes a class of tasks and approaches rather than a single architecture.
S01
].
Machine Learning, ML
is a family of methods in which behavior is adjusted according to the data of a given criterion. Even when the coefficients are obtained automatically, the engineer chooses the space of the hypothesis, the presentation of examples, the function of loss, the procedure of optimization and the quality test. Therefore, the phrase "model itself has learned" does not change the design.
Deep Learning, DL
is part of the ML using learning transformations and multilevel performances.
h
1
=
ϕ
1
(
x
;
θ
1
)
,
h
2
=
ϕ
2
(
h
1
;
θ
2
)
,
y
=
ϕ
3
(
h
2
;
θ
3
)
,
h_1=\phi_1(x;\theta_1),\quad
h_2=\phi_2(h_1;\theta_2),\quad
y=\phi_3(h_2;\theta_3),
h
1
​
=
ϕ
1
​
(
x
;
θ
1
​
)
,
h
2
​
=
ϕ
2
​
(
h
1
​
;
θ
2
​
)
,
y
=
ϕ
3
​
(
h
2
​
;
θ
3
​
)
,
the result of the third conversion depends on the whole chain. "The depth" refers to the composition, but the single official layer threshold, after which the real DL begins, does not.
DL ⊂ ML ⊂ AI
; it does not claim that any method of statistics must be called AI.
Generativity and Foundation Model — Other Axis
Generative model
it gives a distribution from which new samples can be obtained: current sequences, images, sound or other data. This is a feature of the task and output. A deep classification can be DL and only give grade estimates. On the contrary, a simple n-gram language model builds sequences without being a deep network.
S03
So you can't put another circle inside the DL.
Language model
Assesses the probability of linguistic sequences.
LLM
in this course means a major trained model for language problems; the only numerical boundary "large" we do not introduce.
Foundation model
Identifys a well-trained framework that is adapted to a multitude of subsequent tasks [..
S08
The term does not indicate that the browser model, the file system, or the right to execute the command.
Model and application have different boundaries
Parametric model is convenient to write as
y
=
f
θ
(
x
)
,
y=f_\theta(x),
y
=
f
θ
​
(
x
)
,
where
x
- current entrance,
θ
- Parameter state, a
y
This entry does not describe the entire application. The application still decides how to collect
x
what history to include, where to look for documents, what actions to allow and how to show
y
.
It is useful to analyse three levels:
Level
Main issue
Example of an object
Objective
What is the result required?
Find route, classify, continue text
Model
How does the entrance transform with fixed parameters?
f_θ(x)
, logits, probability of class
System
Where does the entrance come from, and who can act?
Retrieval, memory, tool runtime, interface
One model can be used in several products. One application can cause several models. A document found retrieval does not automatically become a line of the parameter matrix; the result of the calculator is not required to be generated by the model.
What can and cannot be concluded
If the app changed the answer after a new request, we know that at least the entry or external state has changed.
not
If the system showed the source, we know that the interface linked the result to some of the design.
not
We know that the source is true or relevant. If a product has a generic model, it doesn't mean that every symbol shown is produced by it.
No AI icon replaces the data flow diagram. In the lower lab, you can turn off retrieval and tools and see which component stops producing the result. This is a causal experiment: one system node is changing, not the commercial name of the product.
A deep classification compiles a composition of learning performances, and a search and calculator can work on a manual basis.
How does this question go about the result?
Academic base