Free trial
A full month, every feature.
$0
for 30 days
- Every feature
- 50 credits to try it
- Ask with your files
- Approve what it drafts
- Students receive it
Ask.Create.Approve.Send.Feedback.
Week 5 slides.pptx
Section A roster.xlsx
Create an exam on self-attention from @Week 5 slides.pptx for @Section A roster.xlsx
Working from 2 course files
DraftApproved12 questionsWeek 5 slides, p. 4EditApproveSent to Section A
Create an exam on self-attention from @Week 5 slides.pptx for @Section A roster.xlsx. 12 questions, cited to Week 5 slides, p. 4. Waiting for your approval.Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
1 · Ask with your files
Ask · Think · Create
Create.Ask.Think.Create.
Create an exam on “Intro to AI” from Week 1–14 slides.
Ask Fazah or create something from your course…Create an exam on “Intro to AI” from Week 1–14 slides.0:00
Options
Speak it in one sentence. Finding the right materials is Fazah's job.
Create an exam on “Intro to AI” from Week 1–14 slides.
Syllabus.pdf
Week 1 slides.pdf
Week 2 slides.pdf
⋯ 11 more decks
Week 14 slides.pdf
Midterm rubric.pdf
Searching your course
Choosing what matters“Week 1–14” → Lectures · 14 decks
Reading the 14 decks
Writing the exam questions
Checking level and answers
FazahD
✦
Courses
Files
Students
Library
Deep Learning
Mon 9:00 AM
Create an exam on “Intro to AI” from Week 1–14 slides.
✦FAZAH · Course-grounded
✓Used 14 course files · Prepared exam setup · 26s
From your course materials
TypeQuizExamAssignment⋯
TopicIntro to AI✦
Total questionsQuickStandardDetailedComprehensiveCustom
Sources selected14 sources· of 14 filesChange
Exam of 24 questions worth 60 marks, from 14 sources.Advanced optionsGenerate exam
Using sources:Lectures · 14 files
Ask Fazah or create something from your course…
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
2 · Approve what it drafts
Grade · Feedback
Feedback.Grade.Feedback.
Create a quiz, slides, notes and an assignment from this week's materials.Working from Week 5 materials.pdf
Used 1 course file · Prepared 4 cards · 12s
QuizNeeds review
Self-attention and scaling
10 questions · 13 marks
Review quizEdit
SlidesNeeds review
Transformer architecture
12 slides
Review deckEdit
Student NotesNeeds review
Attention, in plain terms
6 sections
Review notesEdit
AssignmentNeeds review
Positional encoding
3 tasks
Review assignmentEdit
Working from Week 5 materials.pdf
Ask Fazah or create something from your course…
Options
0:00
Question 1 of 10
Why is the dot product scaled before the softmax?
A
B
C
D
Approve
Approved
Ready to publish
Live with students
Publish to students
View live assignment
Publish to students
CS 471 · Deep Learning · 42 students
Deadline
Tomorrow
In 3 days
In 1 weekIn 1 week
Pick a time in the future
Students
AHLMSKCourse roster+39 more
Accept late submissions
Send to 42 students
Published to students
Students in CS 471 have been emailed their private submission link.
View live assignment
FFazah for Dr. Al-Rashidnow
Self-attention and scaling
10 questions · 13 marks · Due Thu 3 Sep, 23:59
Open quizPrivate link, yours only
Self-attention and scaling42Submitted39Graded3Pending2LateGrade nowRelease 3 grades
StudentStatusScoreTimeliness
AHAmal H.ReleasedNot gradedQueuedGrading…Ready to release11 / 1384.6%On time
LMLina M.ReleasedNot gradedQueuedGrading…Ready to release10 / 1376.9%Submitted late
SKSami K.ReleasedNot gradedQueuedGrading…Ready to release13 / 13100%On time
NANoor A.ReleasedNot gradedQueuedGrading…Ready to release12 / 1392.3%On time
OROmar R.ReleasedNot gradedQueuedGrading…Ready to release9 / 1369.2%On time
HSHana S.ReleasedNot gradedQueuedGrading…Ready to release8 / 1361.5%On time
Class results
82%Class average93%Pass rate
Well understoodWhy the dot product is scaled before the softmax.
Common slipPositional encodings are added to the embeddings, not concatenated.
TryReteach positional encoding from the Week 5 notes, §3; 14 students slipped there.
Self-attention and scaling11 / 13 · 4 Sep
Overall feedback
Amal: a solid grasp of scaled attention. Positional encoding needs one more pass.
Strengths
Explained why √dₖ keeps the softmax gradients healthy.
Areas for improvement
Positional encodings are added to the embeddings, not concatenated.
Question-by-question feedback
Q1: 1/1
Correct: scaling stops the dot products from saturating the softmax.
Q7: 0/2
Misconception:Concatenating the encoding onto the embedding.
Next step:Week 5 notes, §3: Positional encoding.
Amal H. has been emailed that their grade is released
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
3 · Students receive it
Review · Share
Review.Review.Share.
Create a quiz, slides, notes and an assignment from this week's materials.Working from Week 5 materials.pdf
Used 1 course file · Prepared 4 cards · 12s
QuizNeeds review
Self-attention and scaling
10 questions · 13 marks
Review quizEdit
SlidesNeeds review
Transformer architecture
12 slides
Review deckEdit
Student NotesNeeds review
Attention, in plain terms
6 sections
Review notesEdit
AssignmentNeeds review
Positional encoding
3 tasks
Review assignmentEdit
Working from Week 5 materials.pdf
Ask Fazah or create something from your course…
Options
0:00
Question 1 of 10
Why is the dot product scaled before the softmax?
A
B
C
D
Approve
Approved
Ready to publish
Live with students
Publish to students
View live assignment
Publish to students
CS 471 · Deep Learning · 42 students
Deadline
Tomorrow
In 3 days
In 1 weekIn 1 week
Pick a time in the future
Students
AHLMSKCourse roster+39 more
Accept late submissions
Send to 42 students
Published to students
Students in CS 471 have been emailed their private submission link.
View live assignment
FFazah for Dr. Al-Rashidnow
Self-attention and scaling
10 questions · 13 marks · Due Thu 3 Sep, 23:59
Open quizPrivate link, yours only
Self-attention and scaling42Submitted39Graded3Pending2LateGrade nowRelease 3 grades
StudentStatusScoreTimeliness
AHAmal H.ReleasedNot gradedQueuedGrading…Ready to release11 / 1384.6%On time
LMLina M.ReleasedNot gradedQueuedGrading…Ready to release10 / 1376.9%Submitted late
SKSami K.ReleasedNot gradedQueuedGrading…Ready to release13 / 13100%On time
NANoor A.ReleasedNot gradedQueuedGrading…Ready to release12 / 1392.3%On time
OROmar R.ReleasedNot gradedQueuedGrading…Ready to release9 / 1369.2%On time
HSHana S.ReleasedNot gradedQueuedGrading…Ready to release8 / 1361.5%On time
Class results
82%Class average93%Pass rate
Well understoodWhy the dot product is scaled before the softmax.
Common slipPositional encodings are added to the embeddings, not concatenated.
TryReteach positional encoding from the Week 5 notes, §3; 14 students slipped there.
Self-attention and scaling11 / 13 · 4 Sep
Overall feedback
Amal: a solid grasp of scaled attention. Positional encoding needs one more pass.
Strengths
Explained why √dₖ keeps the softmax gradients healthy.
Areas for improvement
Positional encodings are added to the embeddings, not concatenated.
Question-by-question feedback
Q1: 1/1
Correct: scaling stops the dot products from saturating the softmax.
Q7: 0/2
Misconception:Concatenating the encoding onto the embedding.
Next step:Week 5 notes, §3: Positional encoding.
Amal H. has been emailed that their grade is released
The shelf
Everything a course needs.
Drafted from your materials. Released by you.
✓ ApprovedFrom Week 5 materials.pdf · p. 4
Self-attention and scaling10 questions · 13 marks
Why is the dot product scaled before the softmax?
To make training faster
✓To keep the softmax stable
To reduce the parameter count
✓ Approved
Midterm examWeeks 1–5
Variant 01Variant 02Variant 03Variant 04Variant 05Variant 06Variant 07Variant 08Variant 09Variant 10
30 safe variants · same difficulty✓ ApprovedDue: In 1 week
Positional encoding3 tasks
✓Explain why the transformer needs position information
✓Compare sinusoidal and learned encodings
Apply the encoding to a short sequence
✓ Approved
Transformer architecture12 slides
Working from Week 5 materials.pdf
✓ Approved
Attention, in plain terms6 sections
Key concepts highlighted
✓ ApprovedEvery claim cited to its page
Where does attention go wrong?12 sources
✓Vaswani et al., 2017 · the original paper
✓Your reading list · Week 5
✓Your notes · Week 5, p. 4
✓ ApprovedAgainst your rubric. You release them
Week 5 Quiz42 of 42 marked
Amal H.11/13
Sami K.13/13
Lina M.7/13
✓ Approved
Personalized feedback42 of 42
Amal11 / 13
What went wellNext step
One approval gate for all of it
Ten grounded questions, each cited to the page it came from.
One exam, thirty safe variants: one per student, same difficulty.
Tasks with rubrics and a deadline, ready to send.
A teaching deck from this week's chapter, in your voice.
The key concepts, written for the student who missed class.
Beyond your slides: sources found, read, and cited to the page.
Every submission marked against your rubric. You release the marks.
A personal note for every student: what went well, what's next.
In every courseFrom your course
From your course
to every student.
Fazah drafts. You decide.
01 / 05
01Teacher
You approve.
✓Approved
02Student
Each student gets feedback.
Amal0 / 13
What went wellNext step
03Back to you
The class comes back to you.
76%Class average88%Pass rate
↩ Reteach
Week
Class
Approved
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Assignment
Positional encoding
3 tasks
Slides
Transformer architecture
12 slides
Explained why √dₖ keeps the softmax gradients healthy.
Positional encodings are added to the embeddings, not concatenated.
×0:03
Thought for 4sCorrect: scaling stops the dot products from saturating the softmax.
Working from Week 5 materials.pdf
Create a quiz, slides, notes and an assignment from this week's materials.
Student Notes
Attention, in plain terms
6 sections
Quiz
Self-attention and scaling
10 questions · 13 marks
82%Class average93%Pass rate
Positional encodings are added to the embeddings, not concatenated.
Live with studentsCS 471 · Deep Learning · 42 students
Sources usedWeek 3 lecture notes, pages 4 to 9
Pricing
Pick the size of your teaching week.
Faculty Starter
Fazah preps. You teach.
$8/month
$96 billed yearly
- 1 course
- 200 credits a month
Faculty Standard
RecommendedFazah preps and grades.
$16/month
$192 billed yearly
- 3 courses
- 800 credits a month
Faculty Plus
Fazah runs your whole term.
$40/month
$480 billed yearly
- Unlimited courses
- 2,400 credits a month
University
Fazah runs the department.
Custom
Priced to your faculty size. Pooled credits, agreed with you.
- Unlimited teachers
- Credits: Custom
Free 30-day trial · No card required · Change or cancel any time
USD prices exclude sales tax.
Why Fazah
General-purpose AI writes for everyone, about everything. University teaching is not a general purpose.
You decide
Nothing reaches a student until you release it.
Week 3 quizDraft, only youReleaseYour course at the center
Every draft starts from your materials, cited to the page.
Lecture 4.pdfp. 12Your data stays yours
Never used to train AI models.
Your course filesNot used for training
Governance or procurement questions? Contact us
Before you decide
Questions faculty ask.
How is Fazah different from ChatGPT?
ChatGPT answers from everything it has read. Fazah works from the files you upload to your course and shows which file each draft came from. Nothing reaches a student until you release it.
What if Fazah marks an answer wrong?
Every mark arrives as a draft beside the student's answer and your rubric. Change any mark or comment first. Students see nothing until you release it.
What happens to my files and my students' work?
They are never used to train any AI model, ours or a provider's. Students submit through your share link, without an account. If you delete your account, your content is purged after a 30-day grace period.
Does it work with Blackboard, Moodle or Canvas?
It works alongside them. Upload the files you already teach from, share work with students by link, and download quizzes and exams as Word or PDF. Fazah does not connect to your LMS.
Can students copy answers during a quiz?
Turn on Prevent copying and pasting. It blocks copy and paste, warns the student, and records each attempt and each time they leave the page.
Bring your course. Keep your evenings.
Free for 30 days once you’re in. No card.