English
stringlengths
2
2.26k
French
stringlengths
2
1.87k
Instead of pushing the pairs closer to each other, you push both the representations to be inside the same cluster.
Au lieu de rapprocher les paires l’une de l’autre, on pousse les deux représentations à se trouver dans le même cluster.
Preventing trivial solution
Empêcher les solutions triviales
In a trivial solution, all the representations will be the same and thus belong to the same centroid. However, with sinkhorn, different clusters have an equal number of samples, thus the representations can’t be put into one centroid, preventing a trivial solution.
Dans une solution triviale, toutes les représentations sont identiques et appartiennent donc au même centroïde. Cependant, avec Sinkhorn, les différents clusters ont un nombre égal d’échantillons, de ce fait les représentations ne peuvent pas être placées dans un seul centroïde. Cela empêche ainsi une solution triviale.
Other methods
« Autres méthodes »
The loss function for all the previous methods including contrasting methods needs a batch or pool of negative samples, thus creating problems with distributed training. However, the loss functions of these methods are local. These methods perform well but an understanding of why they don’t collapse is not yet available. Probably there’s some implicit regularization happening in these networks to prevent them from converging to a trivial solution.
La fonction de perte de toutes les méthodes précédentes, y compris les méthodes contrastives, nécessite un batch ou un pool d’échantillons négatifs, ce qui pose des problèmes pour l’entraînement distribué. Cependant, les fonctions de perte de ces méthodes sont locales. Ces méthodes sont performantes mais on ne comprend pas encore pourquoi elles ne s’effondrent pas. Il y a probablement une régularisation implicite dans ces réseaux qui les empêche de converger vers une solution triviale.
BYOL:
BYOL
BYOL adds a predictor, predicting hy from hx​. The energy function (D) is a cosine similarity between hy and predicted hy​. There is no term for negative samples i.e., this method only pushes positive pairs closer and has no enforcement on negative pairs. It is thought that asymmetrical architecture with extra layers makes this method work.
BYOL ajoute un prédicteur, prédisant hy à partir de hx​. La fonction d’énergie (D) est une similarité cosinus entre hy​ et le hy prédit. Il n’y a pas de terme pour les échantillons négatifs, c’est-à-dire que cette méthode ne fait que rapprocher les paires positives et n’a aucune action sur les paires négatives. On pense que l’architecture asymétrique avec des couches supplémentaires permet à cette méthode de fonctionner.
SimSiam is a followup version that uses a regular backbone instead of the momentum backbone
SimSiam est une version postérieure qui utilise un backbone normal au lieu d’un backbone avec momentum.
Dino:
Dino
The two softargmax components used have different coldness or temperature. The energy function is the cross entropy between these two, pushing them together. Even this method doesn’t enforce anything on negative samples.
Les deux composants softargmax utilisés ont une froideur/température différente. La fonction d’énergie est l’entropie croisée entre ces deux, rapprochant ces composants. Cette méthode n’impose rien sur les échantillons négatifs.
Data2Vec:
Data2Vec
Adds a layer norm at the end of the representation.
Ajoute une couche de norme à la fin de la représentation.
Initialization of the network:
Initialisation du réseau
If you initialize the network with a trivial solution, then that network will never work. This is because if the trivial solution is already achieved, the loss function will produce a zero gradient and thus, can never escape from the trivial solution. However, in other cases, the training dynamic is adjusted in a way that they never converge in these methods.
Si vous initialisez le réseau avec une solution triviale, alors ce réseau ne fonctionnera jamais. En effet, si la solution triviale est déjà atteinte, la fonction de perte produira un gradient nul et ne pourra donc jamais s’échapper de la solution triviale. Cependant, dans d’autres cas, la dynamique d’entraînement est ajustée de telle sorte qu’ils ne convergent jamais dans ces méthodes.
Improvements for JEMs
Améliorations pour les JEMs
We can further improve these models by experimenting with data augmentation and network architecture. We don’t have a good understanding of these but they are very important. In fact, finding good augmentation may boost more performance than changing the loss function.
Nous pouvons encore améliorer ces modèles en faisant des expériences sur l’augmentation des données et l’architecture du réseau. Nous n’avons pas une bonne compréhension de ces éléments, mais ils sont très importants. En fait, trouver une bonne augmentation peut améliorer les performances plus que la modification de la fonction de perte.
Data Augmentation
Augmentation de données
Most dominant augmentations were proposed by simCLR and improved a little bit by BYOL:
La plupart des augmentations dominantes ont été proposées par SimCLR et améliorées un peu par BYOL :
1. Random Crop (the most critical one)
1. Recadrage aléatoire (le plus critique)
2. Flip
2. Retournement
3. Color Jitter
3. Variation de couleur
4. Gaussian Blur
4. Flou gaussien
It has been found empirically that random crop is the most critical one. It might be because the random crop is the only we can change the spatial information about the images. Flip does the same partly but is weak. Color jitter and gaussian blur change channels.
Il a été constaté empiriquement que le recadrage aléatoire est le plus critique. Cela pourrait être dû au fait que le recadrage aléatoire est le seul moyen de modifier l’information spatiale des images. Le retournement fait la même chose en partie, mais il est faible. La variation de couleur et le flou gaussien changent les canaux.
Masking augmentation:
Augmentation masquée
Recently people are moving towards masking augmentation instead of traditional augmentation in which we mask out most ( ~75% in the below image ) of the patches. It can replace random crop since it’s another way to remove the redundancy of the spatial information
Récemment, les chercheurs se sont tournés vers l’augmentation par masquage au lieu de l’augmentation traditionnelle dans laquelle nous masquons la plupart (~75% dans l’image ci-dessous) des patchs. Elle peut remplacer le recadrage aléatoire car c’est une autre façon de supprimer la redondance de l’information spatiale.
Issues: This works well only with transformer type of architecture and not with convnet. This is because masking introduces too many random artificial edges. For any transformer, the first layer is the conv layer, with kernel size equal to the patch size and thus, this never experiences artificial edges. For convnets which have sliding windows, the artificial edges can’t be ignored and will result in noise.
Problèmes : Cela ne fonctionne bien qu’avec une architecture de type transformer et non avec de type ConvNet. En effet, le masquage introduit trop de bords artificiels aléatoires. Pour n’importe quel transformer, la première couche est la couche ConvNet, avec une taille de noyau égale à la taille du patch, ce qui fait qu’il n’y a jamais de bords artificiels. Pour les ConvNets qui ont des fenêtres glissantes, les bords artificiels ne peuvent pas être ignorés et se traduiront par du bruit.
Projector/Expander:
Projecteur/Expandeur
It is a two/three-layer feed-forward neural network and empirical results show that it is always better to add this in the network architecture.
Il s’agit d’un réseau neuronal feed-forward à deux/trois couches et les résultats empiriques montrent qu’il est toujours préférable de l’ajouter dans l’architecture du réseau.
The projector is used to project into a lower dimension and the expander is used to project into a higher dimension. A projector is used only during the pretraining and removed while performing the downstream task. This is because the projector removes a lot of information even if the output dimension of the projector and the backbone are the same.
Le projecteur est utilisé pour projeter dans une dimension inférieure et l’extenseur est utilisé pour projeter dans une dimension supérieure. Le projecteur n’est utilisé que pendant le pré-entraînement et est supprimé pendant l’exécution de la tâche en aval. En effet, le projecteur supprime beaucoup d’informations même si la dimension de sortie du projecteur et du backbone est la même.
Momentum Encoder
Encodeur avec momentum
Even without a memory bank, a momentum encoder usually helps the performance of the downstream tasks, especially with weak data augmentation.
Même sans banque mémoire, un encodeur avec momentum aide généralement à la performance des tâches en aval, surtout avec une faible augmentation de données.