What are Attention Mechanisms in AI?
AI is a volatile field that is transforming everyday and it’s getting hard to keep up with it. Deep learning is one of the core components of AI and the recent trend in Deep Learning is Attention Mechanisms. Attention Mechanisms in Neural Network is one of the latest advancements in the field of AI. They are based on the real human visual attention mechanism. Being able to focus on a specific and certain point is what human attention mechanism is based on. Exciting, right? Focusing while a certain point can be seen in high resolution and everything else gets blurred and after that certain point is processed, results are efficiently perfect.
Attention Mechanism is not an old study, it has its roots in image recognitions. All the experiments that incorporated Attention Mechanisms went successful. Including the Boltzmann machine experiment with third-order connections learned how to process a small part of an image with the retina of its lens. The data was evaluated on a synthetic dataset along with two image classification datasets accentuating that if the model is well-trained, it can process whole images.
Attention Mechanisms in Neural Networks are usually for NLP (Natural Language Processing). But the question is, what problem Attention Mechanism will solve for chatbots? To understand this, we have an example of NMT (Neural Machine Translation).
NMT relies on feature engineering that is solely based on statistical properties of a text. It’s like feeding utterances to the MS LUIS bot framework. A particular task that is to be performed by the bot is fed into the system and it can only be activated if the user sends a message that contains particular “keywords”. In NMT, the meaning of a sentence is mapped into a vector representation that works as a translation to the back-end engine that comes back with a response based on that vector.
Attention Mechanism blesses the backend with memory. The network is allowed to refer back with the input sequence and the chatbot is able to memorize the conversation and will get your references if you have made any. Another win of using Attention Mechanisms is that we will no longer need to add all the utterances and intents in a chatbot, we will only integrate a small fragment of it just to give the bot a taste of it and it will come up with generic responses. In simple words, we are granting our chatbot the access to its internal memory that contains the whole conversation just so it can make the conversation more human-like and doesn’t make the user go through the hassle of saying everything again.