Deep Learning: Architectures & Methods an der TU Darmstadt | Karteikarten & Zusammenfassungen

Lernmaterialien für Deep Learning: Architectures & Methods an der TU Darmstadt

Greife auf kostenlose Karteikarten, Zusammenfassungen, Übungsaufgaben und Altklausuren für deinen Deep Learning: Architectures & Methods Kurs an der TU Darmstadt zu.

TESTE DEIN WISSEN

What is the goal of Deep Architectures?

Lösung anzeigen
TESTE DEIN WISSEN

Deep learning methods aim at

• learning feature hierarchies
• where features from higher levels of the hierarchy are formed by lower level features
Lösung ausblenden
TESTE DEIN WISSEN

Are multiple hidden layers better than one hidden layer?

Lösung anzeigen
TESTE DEIN WISSEN

In practice, NN with multiple hidden layers work better than with a single hidden layer

Lösung ausblenden
TESTE DEIN WISSEN

What is hype?

Lösung anzeigen
TESTE DEIN WISSEN

Hype: "extravagant or intensive publicity or promotion"

Lösung ausblenden
TESTE DEIN WISSEN

What is hope?

Lösung anzeigen
TESTE DEIN WISSEN

Hope: "expectation of fulfillment or success"

Lösung ausblenden
TESTE DEIN WISSEN

What are the milestones of DL?

Lösung anzeigen
TESTE DEIN WISSEN
• Digit Recognition
• Image Classification
• Speech Recognition
• Language Translation
• Deep Reinforcement Learning
Lösung ausblenden
TESTE DEIN WISSEN

Why is the success of DNNs surprising?

Lösung anzeigen
TESTE DEIN WISSEN
• From both complexity and learning theory perspectives, simple networks are very limited
• The most successful DNN training algorithm is a version of gradient descent which will only find local optima. In other words, it' a greedy algorithm.
• Greedy algorithms are even more limited in what they can represent and how well they learn
Lösung ausblenden
TESTE DEIN WISSEN

What is the task of Classification + Localization?

Lösung anzeigen
TESTE DEIN WISSEN
• Classification: C classes
• Input: Image
• Output: Class label
• Evaluation metric: Accuracy
• Localization:
• Input: Image
• Output: Box in the image (x, y, w, h)
• Evaluation metric: Intersection over Union
Lösung ausblenden
TESTE DEIN WISSEN

Detection as Classification?

Lösung anzeigen
TESTE DEIN WISSEN

Problem: Need to test many positions and scales, and use a computationally demanding classifier (CNN)

Solution: If your classifier is fast enough, just do it. Only look at a tiny subset of possible positions

Lösung ausblenden
TESTE DEIN WISSEN

Object Detection: Evaluation

Lösung anzeigen
TESTE DEIN WISSEN
• We use a metric called "mean average precision" (mAP)
• Compute average precision (AP) separately for each class, then average over classes
• A detection is a true positive if it has IoU with a ground-truth box greater than some threshold (usually 0.5) (mAP@0.5)
• Combine all detections from all test images to draw a precision/recall curve for each class; AP is area under the curve
• TL;DR MAP is a number from 0 to 100; high is good
Lösung ausblenden
TESTE DEIN WISSEN

What are problems with R-CNN?

Lösung anzeigen
TESTE DEIN WISSEN
1. Slow at test-time: need to run full forward pass of CNN for each region proposal
2. SVMs and regressors are post-hoc: CNN features not updated in response to SVMs and regressors
3. Complex multistage training pipeline
Lösung ausblenden
TESTE DEIN WISSEN

What have we learnt about BN?

Lösung anzeigen
TESTE DEIN WISSEN
• BNs encode joint distributions
• They are DAGs (nodes=RVs, edges=dependencies)
• Inference is NP-hard
• Variable Elimination is one of the most basic inference approaches; there are many other inference approaches
• We have skipped learning BNs
Lösung ausblenden
TESTE DEIN WISSEN

What have we learnt about SPNs?

Lösung anzeigen
TESTE DEIN WISSEN

Sum-product networks (SPNs)

• DAG of sums and products
• They are instances of Arithmetic Circuits (ACs)
• Compactly represent partition function
• Learn many layers of hidden variables

Efficient marginal inference

Easy learning

Can outperform well-known alternatives

Lösung ausblenden
• 86515 Karteikarten
• 2067 Studierende
• 117 Lernmaterialien

Beispielhafte Karteikarten für deinen Deep Learning: Architectures & Methods Kurs an der TU Darmstadt - von Kommilitonen auf StudySmarter erstellt!

Q:

What is the goal of Deep Architectures?

A:

Deep learning methods aim at

• learning feature hierarchies
• where features from higher levels of the hierarchy are formed by lower level features
Q:

Are multiple hidden layers better than one hidden layer?

A:

In practice, NN with multiple hidden layers work better than with a single hidden layer

Q:

What is hype?

A:

Hype: "extravagant or intensive publicity or promotion"

Q:

What is hope?

A:

Hope: "expectation of fulfillment or success"

Q:

What are the milestones of DL?

A:
• Digit Recognition
• Image Classification
• Speech Recognition
• Language Translation
• Deep Reinforcement Learning
Q:

Why is the success of DNNs surprising?

A:
• From both complexity and learning theory perspectives, simple networks are very limited
• The most successful DNN training algorithm is a version of gradient descent which will only find local optima. In other words, it' a greedy algorithm.
• Greedy algorithms are even more limited in what they can represent and how well they learn
Q:

What is the task of Classification + Localization?

A:
• Classification: C classes
• Input: Image
• Output: Class label
• Evaluation metric: Accuracy
• Localization:
• Input: Image
• Output: Box in the image (x, y, w, h)
• Evaluation metric: Intersection over Union
Q:

Detection as Classification?

A:

Problem: Need to test many positions and scales, and use a computationally demanding classifier (CNN)

Solution: If your classifier is fast enough, just do it. Only look at a tiny subset of possible positions

Q:

Object Detection: Evaluation

A:
• We use a metric called "mean average precision" (mAP)
• Compute average precision (AP) separately for each class, then average over classes
• A detection is a true positive if it has IoU with a ground-truth box greater than some threshold (usually 0.5) (mAP@0.5)
• Combine all detections from all test images to draw a precision/recall curve for each class; AP is area under the curve
• TL;DR MAP is a number from 0 to 100; high is good
Q:

What are problems with R-CNN?

A:
1. Slow at test-time: need to run full forward pass of CNN for each region proposal
2. SVMs and regressors are post-hoc: CNN features not updated in response to SVMs and regressors
3. Complex multistage training pipeline
Q:

What have we learnt about BN?

A:
• BNs encode joint distributions
• They are DAGs (nodes=RVs, edges=dependencies)
• Inference is NP-hard
• Variable Elimination is one of the most basic inference approaches; there are many other inference approaches
• We have skipped learning BNs
Q:

What have we learnt about SPNs?

A:

Sum-product networks (SPNs)

• DAG of sums and products
• They are instances of Arithmetic Circuits (ACs)
• Compactly represent partition function
• Learn many layers of hidden variables

Efficient marginal inference

Easy learning

Can outperform well-known alternatives

Erstelle und finde Lernmaterialien auf StudySmarter.

Greife kostenlos auf tausende geteilte Karteikarten, Zusammenfassungen, Altklausuren und mehr zu.

Das sind die beliebtesten StudySmarter Kurse für deinen Studiengang Deep Learning: Architectures & Methods an der TU Darmstadt

Für deinen Studiengang Deep Learning: Architectures & Methods an der TU Darmstadt gibt es bereits viele Kurse, die von deinen Kommilitonen auf StudySmarter erstellt wurden. Karteikarten, Zusammenfassungen, Altklausuren, Übungsaufgaben und mehr warten auf dich!

Das sind die beliebtesten Deep Learning: Architectures & Methods Kurse im gesamten StudySmarter Universum

Deep Learning

Hochschule Konstanz

Deep Learning

Universität Hamburg