It was only a matter of time until Google’s DeepMind would learn a new trick. Developing artificial intelligence and improving upon existing breakthroughs remains challenging, yet this latest update is quite significant. DeepMind now has a memory capacity, which can be used to tackle multiple Atari games. This also removes the need to retrain DeepMind for every game it plays.
DeepMind Is Taking Another Step Forward
One of the most frustrating aspects about training an AI is how it forgets previous skills obtained during the training process. In the case of DeepMind, developers could teach it to beat one specific game. Once that goal was achieved, they had to wipe its memory to train it for a new game. A very frustrating feature for the engineers, as there was no viable way to have the AI solution store this information for later use. That problem has now been solved, thanks to the Elastic Weight Consolidation.
To be more specific this new algorithm allows DeepMind to retain knowledge of all of its previously taught skills. In this particular case, it remains the memory and skills to beat the Atari games in question. Moreover, this new algorithm allows the AI solution to learn multiple games in succession without any delays or problems. A fantastic development on the one hand, yet a disconcerting fact at the same time.
It is of the utmost importance artificial intelligence solutions adopt the Elastic Weight Consolidation algorithm moving forward. As brilliant as AI projects can be, they remain prone to conditional forgetting. Not being able to retain knowledge about more than one particular industry could hinder growth in the AI industry moving forward. EWC Solves these problems, although it remains to be seen how much information DeepMind can retain outside of Atari games.
One of the main reasons why companies and entrepreneurs are so excited about AI is how it can be taught virtually anything. However, if it needs to be retrained every single time and forgets its previous knowledge, the novelty can wear off rather quickly. The deep neural networks used by machine learning can only learn multiple tasks when data is presented in one block of information. This poses significant problems that needed to be solved sooner rather than later.
This is only the first step towards achieving adaptive real-time learning capabilities for artificial intelligence solutions. Making the learning process more flexible and efficient remains the top priority, and EWC is a great way to establish that goal in the long run. It is definitely not a perfect solution, though, despite it bringing some interesting possibilities to the table. Moreover, it also provides valuable insights as to how the human brain operates when it encounters new information.
DeepMind has always been a force to be reckoned with. The AI solution has made media headlines several times now, and the EWC algorithm will only make this offering more robust. In fact, it brings our society yet another step closer to developing an advanced artificial intelligence, even though there are major technological hurdles that need to be overcome first and foremost.
If you liked this article, follow us on Twitter @themerklenews and make sure to subscribe to our newsletter to receive the latest bitcoin, cryptocurrency, and technology news.