Category Archives: A.I.

popular technologies in deep learning

For more info : NN for Language

Presenter : Juno

This seminar introduces several popular technologies in deep learning or neural network for natural language processing such as distributed representation of word, neural network language model, recurrent neural network, and LSTM (Long short term memory). In particular, recent advances and a couple of application examples will also be discussed in the last part.

- Reference: Using Neural Network for Modeling  and Representing Natural Languages by Tomas Mikolov (Facebook)
- Reference: Recurrent Nets and LSTM by Nando de Freitas (Oxford Univ.)

Deep Learning

For more info : DEEP LEARNING.compressed

Reference :  (

Presenter : Juno

This seminar introduces Deep learning, which has been a buzz word in machine learning area for recent years. Since it is based on neural network, the seminar first introduces the basic theory and history of Neural network and explains why Neural network failed in 1990s and early 2000s. Unsupervised way of feature learning, is then followed to describe the core of Deep learning. The latter half of the seminar focuses on the algorithms that learn features from data in an unsupervised way. As a result, main algorithms such as Hopfield Net, Boltzmann machine, and Restricted Boltzmann machine will be discussed.


Brief Introduction to H/W Design

For more info : Brief Introduction to H_W Design_FORAMY

Presenter : Neil

I first introduced the most essential circuit elements for developing the hardware device and then explained how this element works.

Then I explained the two ways in which the signal runs through the circuit, how the aforementioned elements control the two signals, and how it applies practically.

And finally, with the introduction of the theories and additional information for the planning and development of the hardware, I have explained how to develop applications for the Wand product which I have created.

I have introduced all of the development sources and programs to our colleagues.

However, I have been deleted from our company blog to prevent improper use.


A Basic Tutorial to Logistic Regression

For more info  : log

presenter : Jay

The seminar covered the basics of logistic regression, which is the simplest method for classification. After going through an numeric example and the principle of maximum likelihood, an overview of logistic regression was given in relationship with other algorithms.The following was shown:
1. The equivalence between the Maximum Entropy algorithm and logistic regression
2. The equivalence between Naive bayes and logistic regression
3. How Conditional Random Fields are basically logistic regression with selective mapping of feature functions
Conclusively, it was shown that logistic regression is a simple algorithm which works well when classifying linearly separable data with a reasonable running time.


On Air

date : (Korean time): AM 9 :00 Fri 28th Nov 

Where: AKA Seoul + Santa Monica R&D Center

About : A brief Introduction to the Hardware Design

presenter: Neil Park

Seminar on Hologram

For more info : A_intro_to_the_Holographic_Projection_David_Kim_AKA_20141106.pdf


This is an introduction to Holographic projection by David Kim.

This is about principle and applications of analog and digital holograms.

There are lot of fun facts about hologram such as hologram concerts including 2pac, Vocaloid Hatsune Miku, Michael Jackson, etc. and hologram transmission, companies and displays used for smartphones.


Check more about hologram in the attachment!


Seminar on Speaker verification & GMM

For more info :  Seminar on Speaker verification and GMM_141024

This seminar is about Speaker verification, which tells whether the speaker’s voice is same as the registered voice, and GMM (Gaussian Mixture Model) which is one of the theories used in speaker modeling.

Basically, Speech recognition includes processes of Background modeling, Feature Extraction and Speaker modeling, which check on whether the input voice is same as the existing one.

In each stage, it proceeds the modeling with the characteristics, such as frequency, extracted from the vocal data. There are two ways to do this:

1. Generative Model – measures the similarity between the samples from each model.

2. Discriminative Model – discriminates each model with their characteristics and defines in which part they each belong to.

and GMM (Gaussian Mixture Model) is one of the Generative Model.



입력받은 음성이 이미 등록되어 있는 어떤 한명의 화자와 동일한 음성인지 확인하는 Speaker verification의 과정의 소개와 speaker의 모델링에서 사용되는 이론 중 Gaussian mixture model (GMM)에 대한 설명을 담았다.

일반적으로 화자 인식은 백그라운드 모델(비교모델), 특정 화자 모델을 순차적으로 생성하는 등록단계와 등록된 모델과 입력된 음성이 동일한 화자인지 확인하는 확인단계로 구성된다.

각 단계에서 모델링을 하기 위해 음성 데이터로부터 주파수등의 특징을 추출하는 과정을 거치고, 특징들을 가지고 모델링을 하게된다.
모델링의 방법은 다음의 두가지가 있다.
1. 각각의 모델에 입력받은 샘플을 1:1로 비교하여 비슷한 정도를 측정하는 방식
2. 각 모델간의 특징이 되는 영역을 구분하고 입력받은 샘플이 어떤 영역으로 속하는지 확인하는 방식
GMM은 전자의 방법 중 하나이다.



Seminar on System Architecture

For more info :  System_Arhchitecture


***Peeking into the Seminar***

This seminar is a short explanation on key features and details to design an architecture for stable services. Key features and basics in making an architecture are guaranteed mass storage transaction, guaranteed scalability, guaranteed high-availability of the service, efficiency in operation management, reinforcing system security, etc.

안정적인 서비스를 위해 Archiecture를 설계하기 위한 주요 항목과 세부 중점 고려사항에 대해 간략히 설명하였다. 아키텍쳐 설계를 위한 주용 항목으로는 대용량 트랜젝선 성능 보장, 아키텍쳐 확장성 보장, 서비스 고 가용성 보장, 운영관리 효율성, 시스템 보안 강화 등이 있다.