Questions tagged [meta-learning]

For questions related to the concept of meta-learning (or learning-to-learn).

Filter by
Sorted by
Tagged with
0 votes
0 answers
24 views

Why is the embedding of a task using Task2Vec not depend on the model?

I saw this in the Task2Vec paper: TASK2VEC depends solely on the task, and ignores interactions with the model which may however play an important role. To address this, we learn a joint task and ...
2 votes
3 answers
179 views

Is it possible to learn the number of layers?

Is it possible, in a transformer or other deep architecture, to include the number of layers as a parameter of the model so it could be learned? In fact, I have a keras layer that I use to change the ...
  • 51
1 vote
1 answer
130 views

How does MAML inner loop optimization works?

I started to learn meta-learning, reading the MAML paper https://arxiv.org/pdf/1703.03400.pdf In the inner loop, I am calculating adapted parameters for each task, I will be doing multiple steps of ...
0 votes
0 answers
27 views

Resolving Derivation Discrepancies for Differentiating through Optimization Paths

I'm reading the paper "Optimizing Millions of Hyperparameters by Implicit Differentiation". The key contribution of the paper is to show that you can replace optimizing through the ...
  • 1
0 votes
0 answers
37 views

Are tasks created at random for training, validation, and testing of Meta Learning algorithms?

When we feed the data to a Meta Learning algorithm, e.g., Prototypical Network, do we create the dataloader in a way to see each training instance only once a epoch, or is it just random for ease of ...
  • 101
0 votes
0 answers
26 views

What are practical methods to acquire a large number of tasks for Meta-learning?

It appears that it may be necessary to acquire a very large number of tasks for meta-learning , because MAML for example says that each task is analogous to a single training example in regular ...
  • 296
3 votes
0 answers
90 views

How many tasks are needed for meta-learning?

This is an empirical question, essentially how many tasks do you need data for, to make a useful meta learning model (e.g. using MAML)? I'm looking for ranges based on personal experience or if anyone ...
  • 296
1 vote
0 answers
54 views

Why doesn't anyone use reinforcement learning to find the best possible alternative to backpropagation?

To be clear, I'm very uninformed on the topic of alternative learning algorithms to backprop, all my knowledge comes from articles like these: lets-not-stop-at-backprop backprop-alternatives we-need-a-...
  • 103
1 vote
0 answers
34 views

What's mutual exclusivity in meta-learning?

What do we mean by mutual exclusivity of tasks? This work (E Pan, 21) and this one (M Yin, 20) state that most classification meta-learning algorithms fail for non-mutually exclusive tasks as the ...
  • 517
1 vote
1 answer
202 views

How to split data for meta-learning?

I've been trying to understand the meta-learning paradigm, more precisely, the optimization-based models, such as MAML, but I have a hard time understanding how I should correctly split my data to ...
  • 1,098
0 votes
1 answer
37 views

Which meta-learning approach selection methodology should I use for similarity learning of an image?

Meta-learning has 3 broad approaches: model, metric and optimization-based approach. Each of them has its own sub-approach, like matching network, meta-agonistic and Siamese-based network, and so on. ...
2 votes
1 answer
86 views

What is $ \nabla_{\theta_{k-1}} \theta_{k}$ in the context of MAML?

I am attempting to fully understand the explicit derivation and computation of the Hessian and how it is used in MAML. I came across this blog: https://lilianweng.github.io/lil-log/2018/11/30/meta-...
2 votes
0 answers
174 views

In few-shot classification, should I use my custom dataset as the validation dataset and mini-ImageNet as the training dataset?

I am new to few-shot learning, and I wanted to get a hands-on understanding of it, using Reptile algorithm, applied to my custom dataset. My custom dataset has 30 categories, with 5 images per ...
3 votes
1 answer
75 views

What exactly does meta-learning in reinforcement learning setting mean?

We can use DDPG to train agents to stack objects. And stacking objects can be viewed as first grasping followed by pick and place. In this context, how does meta-reinforcement learning fit? Does it ...
  • 31
1 vote
1 answer
67 views

Finding the optimal policy from a set of fixed policies in reinforcement learning

This is an open-ended question.Suppose I have a reinforcement learning task that is being solved using many different fixed policies, one of which is optimal. The goal of the agent is not to figure ...
1 vote
1 answer
175 views

How are mujoco environments used for meta-rl?

Afaik, investigating meta reinforcement learning algorithms requires a collection of two or more environments which have similar structure but are still different enough. When I read this paper it was ...
1 vote
0 answers
24 views

How do I format task features with a one-hot task identification vector to ensure separate weight matrices for each task in multi-task RL?

I am on Lecture 2 of Stanford CS330 Multi-Task and Meta-learning, and on slide 10, the professor describes using a one-hot input vector to represent the task, and she also explained that there would ...
  • 116
3 votes
0 answers
41 views

What is the difference between "out-of-distribution (generalisation)" and "(meta)-transfer learning"?

I'm trying to develop a better understanding of the concept of "out-of-distribution" (generalization) in the context of Bengio's "Moving from System 1 DL to System 2 DL" and the concept of "(meta)-...
1 vote
1 answer
729 views

Understanding the derivation of the first-order model-agnostic meta-learning

According to the authors of this paper, to improve the performance, they decided to drop backward pass and using a first-order approximation I found a blog which discussed how to derive the math ...
  • 133
2 votes
1 answer
173 views

What are recent AI software systems and research papers close to J. Pitrat's ideas?

J. Pitrat (born in 1934) was a French leading artificial intelligence scientist (the first to get a Ph.D. in France mentioning "artificial intelligence"). His blog is still online and of ...
1 vote
1 answer
62 views

What AI conferences in Europe should I consider submitting papers to explaining the ongoing work on RefPerSys?

https://afia.asso.fr/journee-hommage-j-pitrat/ is a seminar on March 6th, 2020, in Paris (France, European Union), in honor of the late Jacques Pitrat, who advocated during all his professional life a ...
7 votes
4 answers
8k views

What are the differences between transfer learning and meta learning?

What are the differences between meta-learning and transfer learning? I have read 2 articles on Quora and TowardDataScience. Meta learning is a part of machine learning theory in which some ...
  • 133
6 votes
2 answers
1k views

What are the state-of-the-art meta-reinforcement learning methods?

This question can seem a little bit too broad, but I am wondering what are the current state-of-the-art works on meta reinforcement learning. Can you provide me with the current state-of-the-art in ...
  • 63
3 votes
1 answer
1k views

What does "episodic training" mean?

I'm reading the book Hands-On Meta Learning with Python, and in Prototypical networks said: So, we use episodic training—for each episode, we randomly sample a few data points from each class in our ...
4 votes
1 answer
122 views

How important is learning to learn for the development of AGI?

Some people say that abstract thinking, intuition, common sense, and understanding cause and effect are important to make AGI. How important is learning to learn for the development of AGI?
2 votes
1 answer
625 views

What are the features get from a feature extraction using a CNN?

I've just started to learn CNN and somewhere I have read if I remove the last FCL I will get the features extracted from the input image but... what are those features? Are they numbers? Labels? An ...
4 votes
3 answers
164 views

Can we optimize an optimization algorithm?

In this answer to the question Is an optimization algorithm equivalent to a neural network?, the author stated that, in theory, there is some recurrent neural network that implements a given ...
  • 331
4 votes
1 answer
53 views

Why not go another layer deeper with Auto-AutoML?

So I'm finding AutoML to be pretty interesting but I'm still learning how it all works. I've played with the incredibly broken AutoKeras and got some decent results. The question is, if you are using ...
2 votes
1 answer
246 views

What is the internal state of a Simple Neural Attentive Meta-Learner(SNAIL)?

In the paper A Simple Neural Attentive Meta-Learner, the authors mentioned right before Section 3.1: we preserve the internal state of a SNAIL across episode boundaries, which allows it to have ...
  • 431
2 votes
1 answer
647 views

What is the difference between meta-learning and zero-shot learning?

What is the difference between meta-learning and zero-shot learning? Are they synonymous? I have seen articles where they seem to imply that they are at least very similar concepts.
  • 173
6 votes
2 answers
299 views

How does rotating an image and adding new 'rotated classes' prevent overfitting?

From Meta-Learning with Memory-Augmented Neural Networks in section 4.1: To reduce the risk of overfitting, we performed data augmentation by randomly translating and rotating character images. We ...
  • 161
5 votes
2 answers
354 views

Do genetic algorithms also evolve?

After witnessing the rise of deep learning as automatic feature/pattern recognition over classic machine learning techniques, I had an insight that the more you automate at each level, the better the ...