Special | A | B | C | D | E | F | G | H | I | J | K | L | M | N | O | P | Q | R | S | T | U | V | W | X | Y | Z | ALL
A |
---|
Activation function: contemporary options | ||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
This knowledge base entry follows discussion of artificial neural networks and backpropagation. Contemporary options for are the non-saturating activation functions [Mur22, Sec. 13.4.3], although the term is not accurate. Below, ( should be understood as the output of the summing junction.
References
| ||||||||||||||||||||
Active learning | ||
---|---|---|
Adversarial machine learning | ||||||
---|---|---|---|---|---|---|
Adversarial machine learning (AML) as a field can be traced back to [HJN+11]. The impact of adversarial examples on deep learning is well known within the computer vision community, and documented in a body of literature that has been growing exponentially since Szegedy et al.’s discovery [SZS+14]. The field is moving so fast that the taxonomy, terminology and threat models are still being standardised. See MITRE ATLAS. References
| ||||||
Artificial neural networks and backpropagation | |||
---|---|---|---|
See 👇 attachment or the latest source on Overleaf.
| |||
B |
---|
Batch normalisation (BatchNorm) | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|
Watch a high-level explanation of BatchNorm: Watch more detailed explanation of BatchNorm by Prof Ng: Watch coverage of BatchNorm in Stanford 2016 course CS231n Lecture 5 Part 2: References
| ||||||||||
C |
---|
Cross-entropy loss | ||||
---|---|---|---|---|
[Cha19, pp. 11-14] References
| ||||
D |
---|
Domain adaptation | |||
---|---|---|---|
Domain adaptation is learning a discriminative classifier or other predictor in the presence of a shift of data distribution between the source/training domain and the target/test domain [GUA+16]. References | |||
Dropout | ||||
---|---|---|---|---|
Deep neural networks (DNNs) employ a large number of parameters to learn complex dependencies of outputs on inputs, but overfitting often occurs as a result. Large DNNs are also slow to converge. The dropout method implements the intuitive idea of randomly dropping units (along with their connections) from a network during training [SHK+14]. References
| ||||
F |
---|
Few-shot learning | ||||||
---|---|---|---|---|---|---|
References
| ||||||
P |
---|
Problems of vanishing gradients and exploding gradients | ||||||||
---|---|---|---|---|---|---|---|---|
This knowledge base entry follows discussion of artificial neural networks and backpropagation. The backpropagation (“backprop” for short) algorithm calculates gradients to update each weight. Unfortunately, gradients often shrink as the algorithm progresses down to the lower layers, with the result that the lower layers’ weights remain virtually unchanged, and training fails to converge to a good solution — this is called the vanishing gradients problem [G22, Ch. 11]. The opposite can also happen: the gradients can keep growing until the layers get excessively large weight updates and the algorithm diverges — this is the exploding gradients problem [G22, Ch. 11]. Both problems plague deep neural networks (DNNs) and recurrent neural networks (RNNs) over very long sequences [Mur22, Sec. 13.4.2]. More generally, deep neural networks suffer from unstable gradients, and different layers may learn at widely different speeds. Watch Prof Ng’s explanation of the problems: The problems were observed decades ago and were the reasons why DNNs were mostly abandoned in the early 2000s [G22, Ch. 11].
Watch Prof Ng’s explanation of weight initialisation: References
| ||||||||