Yeah, good morning, everybody.

Welcome back to our seminar, meta learning.

And today we have a presentation entitled meta networks and the presentation will be

given by a Lakha Dutta.

So this stage is yours.

Thanks, Professor.

Hi, good morning, everyone.

Today I'm going to present the paper named meta network, written by Saint Surin Mukthilai

and Hung Phu.

Here is the contents of the presentation.

We will look at the motivation of the paper as mentioned in the introduction.

Then we will see the works related to this paper after which we will dive into the main

topics, meta networks and layer augmentation.

After that, we'll see some experimental results.

And lastly, we will talk about the future scope of the paper.

So first, what is the purpose of this paper?

It is inherent of human to learn quickly and continuously from few examples.

And it is very desirable property in AI models also.

Conventional AI models needs a large data set and lack of ability to learn continuously.

For this ability, our meta network or meta learning introduced here.

So the concept of meta learning is that it is a learning subsystem which adapts with

experience.

There are two levels of meta learning.

One is slow learning and other is fast learning.

First learning is the base level performed with each task and slow learning is performed

across tasks.

Now we'll see the motivation behind the concept.

The goal of the neural net should learn and generalize a new task or concept from a single

example.

A learning model introduced as meta net consists of these three parts.

One is base learner, next meta learner and external memory.

There are two types of training weights.

One is slow weights and other is fast weight.

Slow weights are updated through a learning algorithm and fast weights are updated within

the scope of each task.

There are two types of loss function.

A main task loss used for the input task objective and a representation loss is for the learner

criteria.

Next we will see some related works.

There are a lot of work related to this concept.

Given the time constraint, I'll go through a few.

One of them is Vinyl's et al. in 2016, the NEMD combined the training and testing procedure

one short learner and developed an end-to-end differentiable nearest neighbor method for

one-shot learning.

This also followed in this paper.

Another one is Revi and Larochelle.

In 2017, they proposed the LSTM-based one-shot optimizer.

After that, we see that a memory augmented neural network man has been introduced and

in our seminar, we already discussed this network.

Next we will see the meta network here.