O’Reilly Artificial Intelligence Conference 2019 – London, UK

About this course

so let’s look at texts these days the favorite approach is to use some massive off-the-shelf model pre-trained on lots of data so for example open a I released a version of their GPT to model with

over 800 million parameters but now researchers and practitioners are also beginning to understand that you need to infuse these models acknowledge or tune them for your specific application

domains so the objective is to build a machine reading comprehension system and just like in self-driving cars where you have predetermined levels of autonomy progress in terms of machine reading

will happen in stages so right now we’re in the simple notification and FAQ assistance but over time they’ll become more personalized and we’ll have autonomous assistance now we have

numerous sessions and texts including tasks from people in healthcare finance media and more now what about pre train models for speech there are a few reasons why one-size-fits-all speech

model hasn’t quite yet materialized you know a spoken language is just not as precise as written text and even if you have access to the best speech to text technology available you often end up

with a document the sentences that contain pauses filters restarts interjections and ungrammatical constructs that make it very difficult but even with these challenges there are

many areas where speech has grown in recent years for example comScore estimates that estimates that by the end of 2020 but half of all online searches will use voice which is pretty

remarkable and smart speakers are expected to grow by more than 82 percent from 2018 to 2019 and by the end of the year the installed base for such devices is expected to exceed 300 million

you