
With AI platforms like Alexa and Siri improving every passing day, the next natural step for tech firms is to make them more human-friendly, at least in terms of conversation. We saw Samsung attempting this recently with its ‘Neon’ platform and now Google is trying to do the same with its new open-domain chatbot named ‘Meena’. This was announced by Daniel Adiwardana, Senior Research Engineer, and Thang Luong, Senior Research Scientist, Google Research, Brain Team in a blog post.
“Meena is an end-to-end, neural conversational model that learns to respond sensibly to a given conversational context. The training objective is to minimize perplexity, the uncertainty of predicting the next token (in this case, the next word in a conversation). At its heart lies the Evolved Transformer seq2seq architecture, a Transformer architecture discovered by evolutionary neural architecture search to improve perplexity,” explains the blog post.
‘Meena’ is not really available for open-source as of now. But one can assume this to happen in the future. Google says that the chatbot has over 2.6 billion parameters, a key reason why it can respond to queries more naturally than other AI chatbots. Also mentioned is that the team trained the chatbot with around 40 billion words and 341GB of text data which also includes social media conversations. It also uses Google’s Seq2seq model, which in simple terms, is a neural network that reads words placed next to each other in a paragraph and checks if the relation between those two makes sense.
0 Comments