15
votes
Why LLMs and RNNs learn so fast during inference but, ironically, are so slow during training?
There is huge difference between what is happening with the information during training and during inference and one can not be used for the other.
Let me start with an analogy to the human brain (...
- 551
11
votes
Why LLMs and RNNs learn so fast during inference but, ironically, are so slow during training?
They are not "learning" during inference at all.
Learning is the process of updating the weights of the model (to lower loss). This does not happen during inference. The model weights stay ...
- 144
6
votes
Accepted
What are the differences between transfer learning and meta learning?
First of all, I would like to say that it is possible that these terms are used inconsistently, given that at least transfer learning, AFAIK, is a relatively new expression, so, the general trick is ...
- 39.7k
6
votes
What are the differences between transfer learning and meta learning?
Meta-learning is more about speeding up and optimizing hyperparameters for networks that are not trained at all, whereas transfer learning uses a net that has already been trained for some task and ...
- 327
5
votes
Accepted
Can we optimize an optimization algorithm?
First, you need to consider what are the "parameters" of this "optimization algorithm" that you want to "optimize". Let's take the most simple case, a SGD without momentum. The update rule for this ...
- 3,173
4
votes
How does rotating an image and adding new 'rotated classes' prevent overfitting?
How can data augmentation reduce overfitting?
You write that you can already maybe see how data augmentation can help prevent overfitting in general, but it sounds a bit uncertain and it's still ...
- 10.1k
3
votes
What are the differences between transfer learning and meta learning?
The difference really comes down to the fact that in meta-learning, there is a population of tasks $\tau$ which have distribution $p(\tau)$. The goal is to perform well on a task drawn from $p(\tau)$. ...
- 276
3
votes
Accepted
Why LLMs and RNNs learn so fast during inference but, ironically, are so slow during training?
As pointed out by others, what you call "learning" at inference, is nothing more than providing more context. The model can indeed memorize in its short-term, but it is only working for the ...
- 466
3
votes
Why LLMs and RNNs learn so fast during inference but, ironically, are so slow during training?
It's kind of like short-term memory versus long-term memory. Giving a language model a small amount of information at inference time allows it to use that information, and so you might say that the ...
- 151
3
votes
Accepted
What are the state-of-the-art meta-reinforcement learning methods?
One of the most recent papers on meta-RL is meta-Q-learning. This paper introduces Meta-Q-Learning (MQL), a new off-policy algorithm for meta-reinforcement learning (meta-RL). MQL builds upon three ...
- 381
3
votes
How important is learning to learn for the development of AGI?
Learning to learn (also known as meta-learning) is very important for the development of artificial general intelligence (AGI), given that one of the desirable and fundamental properties of an AGI is ...
- 39.7k
3
votes
Do genetic algorithms also evolve?
In principle, yes, you can also evolve the genetic algorithm (or, in general, evolutionary algorithm), i.e. you can evolve its operations (such as the mutation and cross-over) and hyper-parameters (...
- 39.7k
2
votes
How does rotating an image and adding new 'rotated classes' prevent overfitting?
Over-fitting in the context of convergence in a neural network can have many causes. When the model implied in the design of the network is not well fitted for the task, the network may still ...
- 7,443
2
votes
Accepted
What are the features get from a feature extraction using a CNN?
You get what we call high-level features, which are basically abstract representations of the parts that carry information in the image you want to classify.
Imagine you want to classify a car. The ...
- 3,173
2
votes
Can we optimize an optimization algorithm?
We usually optimize with respect to something. For example, you can train a neural network to locate cats in an image. This operation of locating cats in an image can be thought of as a function: ...
- 39.7k
2
votes
Accepted
Is it possible to learn the number of layers?
I like the idea, but I fear this approach may be a dead end. I see a few problems:
Layers in front of (closer to the output than) the currently selected layer(s) don't affect the output, so they won't ...
- 491
2
votes
How to split data for meta-learning?
I assume in your case what you need to be doing is to collate your 3 datasets together - these would form the training dataset, and then leave the testing dataset aside.
During Meta-Training, the code ...
- 131
2
votes
Understanding the derivation of the first-order model-agnostic meta-learning
$\nabla_{\theta_{i-1}} \theta_{i-1} = \mathbf{I}$ in a similar way that $\frac{d f}{dx} = 1$ for $f(x) = x$. Strictly speaking, $\mathbf{I}$ should be a vector of $1s$ with the same dimensionality as $...
- 39.7k
2
votes
Accepted
What is the difference between meta-learning and zero-shot learning?
First see the definition of meta-learning:
Meta-learning is a subfield of machine learning where automatic learning algorithms are applied to metadata about machine learning experiments. As of 2017 ...
- 1,751
1
vote
Is it possible to learn the number of layers?
The work had been done before, take a look at this paper. The author not only search for the number of layers but also the whole model architecture.
By using reinforcement learning, the system makes a ...
- 902
1
vote
Is it possible to learn the number of layers?
It is always possible to use a Dense layer to allow the network to built its own menu of layers
...
- 51
1
vote
What are recent AI software systems and research papers close to J. Pitrat's ideas?
Today one of the challenges is learning representations/concepts that are causally invariant. Once we have good representations then we can work on the reasoning aspect. There are 2 camps of people ...
- 119
1
vote
What AI conferences in Europe should I consider submitting papers to explaining the ongoing work on RefPerSys?
Probably to as many as possible. Average accept rate of papers is around 20%. You can find the best conferences on AI & ML Event.
- 111
1
vote
What are the differences between transfer learning and meta learning?
The way nbro describes meta-learning, it sounds identical to hyper-parameter optimization. Here I would like to clarify possible differences:
According to Dataset2Vec: learning dataset meta-features
...
- 163
1
vote
What are the state-of-the-art meta-reinforcement learning methods?
Meta-Reinforcement Learning can refer to a broad range of ideas. Also, different algorithms are SOTA under different evaluation metrics (sample efficiency, agent performance, adaptation speed on a new ...
1
vote
What does "episodic training" mean?
It consists of organizing training in a series of learning problems, each relying on small "support" and "query" sets to mimic the few-shot circumstances encountered during ...
- 111
1
vote
Why not go another layer deeper with Auto-AutoML?
Logically it is possible, but you will just end up complicating the entire task.
The aim of AutoML is to provide a drop in solution to the customers. To do this, a trained network decides and ...
- 393
1
vote
What is the internal state of a Simple Neural Attentive Meta-Learner(SNAIL)?
Here's what I understand, welcome to point out any mistakes.
When starting a new episode(but still in the same task), SNAIL does not clear its batches. Instead, it makes decisions based on the ...
- 451
1
vote
Do genetic algorithms also evolve?
A genetic algorithm is a class of evolutionary algorithms.
They do get better at searching through the solution possibilities for each trial (generation) over time because evolution usually starts ...
- 136
Only top scored, non community-wiki answers of a minimum length are eligible