Learning to follow navigational instructions

Can, Ozan Arkan and Yuret, Deniz. 2017. International Symposium on Brain and Cognitive Science (ISBCS2017). Invited talk.

Full post...

The third deep learning revolution

1st presentation: April 19, 2017. Koç University Alumni Club.
2nd presentation: June 4, 2017. GİF Young Scholars Seminar.
3rd presentation: November 17, 2017. İzmir Fen Lisesi.
4th presentation: March 2, 2018. FMV Özel Ayazağa Işık Fen Lisesi.
5th presentation: March 21, 2018. İTÜ Mimarlık Bölümü.
6th presentation: April 20, 2018. Borsa İstanbul. (latest version)

The first revolution took place 1958-1969.

• We figured out how to train perceptrons.
• We proved the perceptron convergence theorem.
• Interest waned after a book (Perceptrons) written by mathematicians.

The second revolution took place 1986-1995.

• We figured out how to train multi-layer perceptrons.
• We proved the universal approximation theorem.
• Interest waned after a book (SLT) written by mathematicians.

The third revolution started in 2012.

• We figured out how to train deep nets.
• A mathematician will write a book around 2022.
• The fourth revolution will not start until 2036 :)

Full post...

Overfitting, underfitting, regularization, dropout

Here is an IJulia notebook demonstrating overfitting, underfitting, regularization and dropout in Knet for my machine learning class.
Full post...

Julia ve Knet ile Derin Öğrenmeye Giriş

Sunumlar:

• Data İstanbul, 21 Aralık 2016 Çarşamba, 19:30 (URL, Sunum,Video).
• ODTÜ Yapay Öğrenme ve Bilgi İşlemede Yeni Teknikler Yaz Okulu (OBAYO), 6-9 Eylül, 2016, ODTÜ, Ankara. (URL, Sunum,Video)
• İsmail Arı Bilgisayar Bilimleri ve Mühendisliği Bilimsel Eğitim Etkinliği (ISMAIL 2017), 31 Temmuz - 4 Ağustos 2017 Boğaziçi Üniversitesi, İstanbul. (URL).

Full post...

CharNER: Character-Level Named Entity Recognition

Onur Kuru, Ozan Arkan Can and Deniz Yuret. 2016. COLING. Osaka. (PDF,Presentation)

Abstract

We describe and evaluate a character-level tagger for language-independent Named Entity Recognition (NER). Instead of words, a sentence is represented as a sequence of characters. The model consists of stacked bidirectional LSTMs which inputs characters and outputs tag probabilities for each character. These probabilities are then converted to consistent word level named entity tags using a Viterbi decoder. We are able to achieve close to state-of-the-art NER performance in seven languages with the same basic model using only labeled NER data and no hand-engineered features or other external resources like syntactic taggers or Gazetteers.

Full post...