Briefly introduce what Seq2seq is

There are several large cloze-style context-question-answer datasets which was introduced 4~5 years ago - the CNN, Daily Mail news data, and the Children's Book Test. Thanks to the introduction to these large datasets, it became easier to associate text comprehension task to deep-learning techniques that seem to outperform all alternative approaches.

Intro to Few Shot Learning

[AlphaPose](http://www.mvig.org/research/alphapose.html) is an accurate multi-person pose estimator, which is the first open-source system that achieves 70+ mAP (75 mAP) on COCO dataset and 80+ mAP (82.1 mAP) on MPII dataset.

In this post, I aim to address potential problems with RNN based _neural machine translation (NMT)_ models and introduce a solution this paper proposed.

To improve the Vector Representation Quality of Skip-gram (one of the Word2Vec method). There are 4 ways to improve representation quality and computational efficiency.

How to make text to be input in deep learning?

Google Duplex is a name for the technology supporting Google Assistant. This service was first introduced in 2018 and has been mainly used in _booking_ with human-like phone calls.

In Biomedical field, the instance segmentation are frequently used such as detecting tumors based on radiography, lesion segmentation, etc. What is important here, in biomedical data, is that the output should include localization. Let's look at what is U-Net and how it works

Today, we are going to know how to crawl iherb using Python, especially information about supplements.

See what is YOLO and how it works

Today we will look at different activation functions, especially the family of **ReLU (Rectified Linear Unit)** activation function. The role of activation functions in neural networks is taking the input and mapping it into output that goes into next layer. Deciding which activation function to use heavily depends on the target.

Fast R-CNN starts from the idea: What if using convolutional feature maps for generating region proposals?

Briefly introduce one of the most important object detection papers

In this posting, we will deal with the NLP model named Big Bird. As you can see in the title, it is the model for _longer_ sequences. Let's take a brief look at the concept of Big Bird.

In this posting, I will talk about **Multi-Task Learning (MTL)** briefly and then deal with the NLP model named MT-DNN.

I will focus on the core part of this paper (=Factorized self-attention) and briefly mention rest of the stuffs.

What is One Shot Learning? examples with face recognition

Today, we will look at the basic approaches to anomaly detection.

Take baby steps towards the Computer Vision master

Today we are going to cover some important papers about object detection using deep learning architecture.

In this posting, we will take a quick look at the NLP tasks that I picked for explanation.

Text Clustering - BERT + Dimension reduction + KMeans

From SWA to TTA

BERT is a pre-trained model released by Google in 2018, and has been used a lot so far, showing the highest performance in many NLP tasks. In this post, let's learn more about the detailed structure of BERT.

Rectified Adaptive Learning Rate (RAdam)

What is learning rate scheduler?

In this post, we will dive into negative sampling for test / train instances in recommendation system.

In this posting, we will review a paper titled: Attention is all you need, which introduces the ttention mechanism and Transformer structure that are still widely used in NLP and other fields. BERT, which was covered in the last posting, is the typical NLP model using this attention mechanism and Transformer. Although Attention and Transformer are actively used in NLP, they are also used in many areas where recurrent methods were used. From now on, let's take a closer look at what Attention and Transformer are.

Although most of the NLP models already offered a pre-trained model for multilingual data, it is still difficult to put it directly into Korean. Korean is a complex language, so there are many aspects that the Tokenizer used here does not fit well. No matter how well pre-training has done, the performance will be terrible if you don't make subword vocab well.

Can we replace maxpooling with convolutional layers?

In many cases including image segmentation, a model consists of downsampling and upsampling parts and the latter restore the feature map to the input sized image. There are two types of upsampling using torch: UpSampling, ConvTranspose2D.

Today we are going to build a simple autoencoder model using pytorch. We'll flatten CIFAR-10 dataset vectors then train the autoencoder with these flattened data.

Cross-validation is one of the most popular methods to evaluate model performance and tune model parameters. Like the bootstrap, it belongs to a family of Monte Carlo methods. Today, we will go over several types of CV methods.

There are countless ways to perform NLP, and the flow of methodologies is changing very quickly. Therefore, it is important to understand the latest trends in NLP. Since it is difficult to handle each method in a single posting, this posting will cover only the overview of NLP, and then individual models will be covered in the subsequent postings in more detail.

What is transfer learning and why?

Let's get familiar with list comprehension, instead of for loops

regularization methods: early stopping, weight decay

What is object localization, object detection, semantic segmentation and instance segmentation?

Alternatives to the Fully Connected Layer(FC layer)

Fundamental recommendation system models, content-based, collaborative-filtering and knowledge-based

Hybrid method, as the name suggests, is a mixture of methodologies such as CBF and CF.

Today, I will review a paper that first integrated the idea of latent variables with deep learning.

Let me tell you who we are