RE•WORK’s Deep Learning and Chatbot Summit: My Key Takeaways
As part of the development of BOLDR (our new professional AI powered coach in private alpha now), last week, I attended RE•WORK’s most recent event — the Deep Learning Summit. The event brought together cutting-edge science and entrepreneurship to showcase how Deep Learning Tech can positively disrupt markets and our lives. It was a fantastic event, which not only introduced me to some incredible brains (entrepreneurs and academics) but also gave plenty of tips on how deep learning can enhance our product. Here are a few of the highlights.
Raia Hadsell, a senior research scientist at Google Deepmind, spoke about how complex problem solving (involving long sequences of tasks) remains a key obstacle in achieving human-level intelligence. Previous methods for continual learning, Raia argued, are not applicable to deep neural networks; however, progressive neural networks (which are immune to forgetting and are able to leverage prior knowledge via lateral connections to previously learned features) are a step in the right direction.
Raia then illustrated how such a progressive approach to deep learning can be used to transfer knowledge from simulated tasks to a real robot domain. This was not a talk for the faint-hearted.
Murray Shanahan, professor of Cogntive Robotics at Imperial College London, talked about how, despite certain dramatic successes, contemporary deep reinforcement learning methods have particular shortcomings. One of these is the slow learning they support due to their reliance on statistics from large datasets. By introducing a symbolic component, that allows for rapid generalisation at a high level of abstraction, Murray argued that deep learning processes could become faster and more accurate.
I also attended the Chatbots breakout track. This looked at how AI and deep learning can be used to create chatbots and conversational interfaces that enable deeper, more personalised one-to-one customer experiences.
Chris Bauer, senior lecturer at Goldsmiths, spoke about how he has discovered that people trust robots more than humans because there is no fear of judgement. Which is fascinating, as it implies that users will be more open with a chatbot than they will be with their friends.
Artem Rodichev, a machine learning engineer, spoke about how the company Luka is developing chatbots capable of mimicking certain personalities — such as celebrities and fictional characters. Luka have done a good job of taking a huge amount of data and apply it to create a personality of a chatbot. Still an area with lots of room for development.
Peter Gasston, creative technologist at +rehabstudio (a creative technology company) humorously spoke about his experience of developing a chatbot, TRex, aimed specifically at children. The character is the extinct T Rex species which can answer questions about himself. This was no small feat. By ‘training’ the bot with lots of input, TRex can answer questions such as ‘When did you live?’.
Peter’s learnings were that the technology is still catching up, but new tools are being created all the time. He urged developers to create better documentation, to narrow the domain of the bot and set expectations to users upfront in an effort to help build more robust and user-friendly bots.
He also talked the perils of training the chatbot and the kind of abuse to expect from the users. Even children (who were the target audience of the chatbot) got very creative with how to abuse the poor bot. So be prepared, Peter said.
The event was a great gathering of minds, and illustrated how deep learning is able to use massive amounts of data to create new “intelligent” systems and processes. I certainly learnt a lot, and very much look forward to next year’s event — which will no doubt be as well put together as this year’s.
Did you attend the Deep Learning Summit? If so, what fascinated you? And do you think that events like this are important for furthering great technology and ideas?
Let me know your thoughts below.
Would you like to be first to try BOLDR? Sign up for an invite here.
