63

Yes, indeed, neural networks are very prone to catastrophic forgetting (or interference). Currently, this problem is often ignored because neural networks are mainly trained offline (sometimes called batch training), where this problem does not often arise, and not online or incrementally, which is fundamental to the development of artificial general ...


21

Yes, the problem of forgetting older training examples is a characteristic of Neural Networks. I wouldn't call it a "flaw" though because it helps them be more adaptive and allows for interesting applications such as transfer learning (if a network remembered old training too well, fine tuning it to new data would be meaningless). In practice what you want ...


11

I'd like to add to what's been said already that your question touches upon an important notion in machine learning called transfer learning. In practice, very few people train an entire convolutional network from scratch (with random initialization), because it is time consuming and relatively rare to have a dataset of sufficient size. Modern ConvNets ...


6

Here is one way you could do that. After training your network, you can save its weights to disk. This allows you to load this weights when new data becomes available and continue training pretty much from where your last training left off. However, since this new data might come with additional classes, you now do pre-training or fine-tuning on the network ...


6

It was essentially a lack of control over crowd-sourced training data. While Tay was initially set up with some conversational ability, it seemed to be programmed to learn from interactions with other users. Once users became aware of this, they basically gamed the bot by exposing it to inappropriate language, which Tay's algorithms then picked up and ...


6

Looking at what happened, it was something similar. Though, the case differs in my eyes from one perspective: if it could only do a few comedy jokes, that probably is not a profound starting point to excel in Twitter. Firstly, Twitter is about real life, not about comedy. Discussions are sometimes tough and you easily end up to Social Media Bubbles, where ...


5

What you are describing sounds like it could be a deliberate case of fine-tuning. There is a fundamental assumption that makes minibatch gradient descent work for learning problems: It is assumed that any batch or temporal window of consecutive batches forms a decent approximation of the true global gradient of the error function with respect to any ...


3

You are referring to catastrophic forgetting which could be an issue in any neural net. More specifically for DQN refer to this article.


3

Do you know which are the state-of-the-art approaches on this topic, and could you point me to some literature on them? This answer already mentions some of the approaches. More concretely, currently, the most common approaches to continual learning (i.e. learning with progressively more data while attempting to address the catastrophic forgetting problem) ...


2

There are lots of different approaches that try to avoid catastrophic forgetting in neural networks. It is impossible to summarize all contributions here. However, in addition to the already mentioned techniques, there are sparsity approaches that try to disentangle internal representations of the network on different tasks or learning steps. Sparsity ...


2

What I understand from your questions is that you are trying to avoid catastrophic forgetting while applying online learning. This problem should be addressed by implementing methods that reduce catastrophic forgetting for different tasks. At first glance it might seem that they don't apply because it's data that change and not a particular task but ...


1

Transfer learning is a field where you apply knowledge from a source onto a target. This is a vague notion and there is an abundance of literature pertaining to it. Given your question I will work under the assumption that you are referring weight/architecture sharing between model (in other words training a model on one dataset and using it as a featurizer ...


1

Maybe in theory, but not in practice. The thing is you seem to consider only chronological/sequential training. And there are two ways to view this issue: online learning -> then it is a feature of the method offline learning -> it does not happen thanks to several order randomizations 1. Online-Training or Online Machine Learning. Using the woppal ...


1

Successively, we get a new training picture, which we want to use to train the FFNN even further. However, wouldn't this new training picture destroy the previously learned weights, which have been calibrated to recognize the former training pictures? This can happen, and happen to various degrees depending on how the neural network is set up, but it is ...


1

There are several ways to add new classes to the trained model, which require just training for the new classes. Incremental training (GitHub) continuously learn a stream of data (GitHub) online machine learning (GitHub) Transfer Learning Twice Continual learning approaches (Regularization, Expansion, Rehearsal) (GitHub)


Only top voted, non community-wiki answers of a minimum length are eligible