Event is FINISHED
[Advanced Developers Only] Exploring Generative Modelling & Transfer Learning
Machine intelligence research is advancing so rapidly it can be hard to keep up, especially when expertise is spread across different organizations. Join Dr. Dan Kuster of Cambrio, in partnership with Cenozai and the KL AI Meetup, to discuss "Learning to generate reviews and discovering sentiment" from Alec Radford and colleagues at OpenAI. Using a recurrent neural network (multiplicative LSTM) to read bytes and predict the next character of text, they discovered representations that are disentangled and interpretable.
Why is this paper significant?
To read and understand text, the best machine learning models today attempt to encode text into some representation that is useful for the task at hand. But nobody knows if there is a general representation for natural language that works across tasks, let alone how to train models to exploit it.
This paper is exciting because an unsupervised model was shown to outperform a supervised model on well known tasks. When inspecting the model, they discovered that the learned representations were disentangled and interpretable. Unexpectedly, one unit recovered the concept of sentiment. Like other RNNs, the model can be used to generate text. Unlike other models, the "sentiment neuron" means it is possible to control the sentiment of the generated text. A promising step in the search for general representations!
What you might learn
How and why this kind of model works
Strategies to apply transfer learning to text models
Generative models using text features
Tradeoffs and data efficiency in recurrent neural networks
Implications regarding better representations for text and natural language
Dr. Kuster is an exclusive partner of Cenozai and the founder + CEO of Cambrio, a Boston-based digital R&D lab that helps businesses build capability in machine learning. He has designed and executed innovation projects for Fortune 500 corporations, government agencies, nonprofits, and startups, which have been covered by Nature, The New York Times, Wired, The Economist, and NPR.
Who should come?
Computer engineers who have started dabbling in machine learning techniques and algorithms, as well as data scientists. The material is fairly advanced, so it is recommended that you have at least a base understanding of ML techniques and algorithms to facilitate better discussion of the paper.
Limited space, you will be provided location of this event upon RSVP and confirming with KL AI organisers. Location is downtown Kuala Lumpur. Please note that time and place of this event is not the usual KL AI meetup schedule.
Want to find more groups and events like this?