You are here:  Home  >  Data Education  >  BI / Data Science News, Articles, & Education  >  BI / Data Science Articles  >  Current Article

Deep Learning Updates: Machine Learning, Deep Reinforcement Learning, and Limitations

By   /  February 1, 2018  /  No Comments

deep learningIn recent years, some astonishing technological breakthroughs in the field Artificial Intelligence (AI) and its sub-field Deep Learning have begun to train machines to behave like humans.

As machines are increasingly emulating complex cognitive functions such as deductive reasoning, inferences, and informed decision-making, robots functioning as humans are a reality in many industry practices today.

However, machines are still behind in articulating the reasons behind their choices or actions. In other words, a machine witness still cannot be used in a court of law to solve a case as it cannot “justify” past actions. The noteworthy achievements in AI applications include the inclusion of neural networks and Deep Learning (DL), which combine unique training opportunities for machines to learn from layers of knowledge, and then to apply that knowledge to achieve particular goals.

Neural networks and Deep Learning penetrate the intensely complex realms of physics, mathematics, statistics, signal processing, Machine Learning, neuroscience, and many others.

Machine Learning (ML),neural networks, and Deep Learning together represent rapidly advancing core group of technologies within the bigger field of Artificial Intelligence.  Nowadays, machines are capable of instantly solving many complex problems using these techniques, which normally would take capable human brains to figure out over a much longer length of time.

Human society is increasingly relying on smart machines to make decisions and solve day to day problems. This has been made possible due to the presence of neural networks and Deep Learning in AI applications.

Triumph of Deep Reinforcement Learning: Deep Q-Network (DQN)

Deep Q-Network or DQN is DeepMind’s brainchild in the area of Deep Reinforcement Learning. The initial paper on DQN was published in Nature Magazine in 2015, after which many reputed research organizations entered this field of study. The entire research community believes that Deep Neural Networks (DNN) can enhance Reinforcement Learning (RL) for interacting with HD images and others because of the presence of DQN techniques. From Google to Facebook, many other industry leaders are patiently awaiting the results of advanced research with DQN.

Market Success of Evolution Strategies in Reinforcement Learning

Evolution Strategies (ES) seem to have made a comeback in Reinforcement Learning. The reason for its apparent success are the explorative algorithms that do not rely on gradients, scalability of algorithms, and cheap hardware requirements – no expensive GPU is necessary for fast parallel processing.

Deep Learning Frameworks

2017 was certainly the Year of Deep Learning Frameworks. Although both Google’s Tensorflow and Facebook’s PyTorch gained much acclaim among the Natural Language Processing (NLP) research communities, Tensorflow is more suited for static graphs. On the other hand, PyTorch is ideal for dynamic structures. Both Tensorflow and PyTorch have received positive user feedback in their respective arenas in 2017 and are making bigger plans moving into 2018.

Reinforcement Learning Agent Beats Human AlphaGo Players

This Nature paper reminiscences a glorious moment when a Reinforcement Learning agent beats the world’s best human Go players. In the first version of AlphaGo, training data collected from human players was used, and further tuned through the combined use of self play and Monte Carlo Tree Search. Thereafter in AlphaGo Zero, the machine learned to play on its own without any human intervention. The science and technology used in this version of the game is aptly described in the paper titled Thinking Fast and Slow with Deep Learning and Tree Search . These games motivated so many human players to perfect their techniques that DeepMind released an AlphaGo Teach to train human players.

The Next Hurdle for DeepMind

Witnessing the recent success and popularity of AlphaGo series of games, DeepMind started contemplating multi-player Poker games using RL techniques. DeepMind is also working on Starcraft 2 – an RL research environment.

Role of Deep Learning in AI: Misinformation is Possible

According to industry watchdog KDNugget, Deep Learning is not the AI future. As both Google and Facebook have marketed their DL offerings globally, the users have mistakenly received the wrong messages about the importance of DL in Artificial Intelligence applications. KDNuggets thinks that the worldwide publicity of DL techniques is more hype than substance. In its defense, KDNuggets further states that techniques like Decision Trees as used in XGBoost “are not making headlines,” but are as important as Deep Learning if not more.

Even in the case of AlphaGo, the media hyped about DL, whereas in truth, the Monte Carlo Tree Search method contributed to the success of the game as much as DL. In RL, many learning tasks are achieved via Neuroevolution’s NEAT and not due to back-propagation, as claimed by media.

The Biggest Limitation of Deep Learning: Machines Cannot Provide Legal Explanations

DL has two major drawbacks right now: the first being the tendency to abruptly forget previous learning, and the second being the inability to question or rationalize the information provided. In other world, in the DL world, the machine believes what it is fed in the learning process, and it does not have the capability to question the acquired knowledge. The last trend is dangerous as the machine can be fed nonsense in the name of “truisms.”

The NLP comments provided in DL applications cannot be accepted by a judge or entered in a court case as evidence or argument. In other words, even the smartest AI systems cannot be held “accountable” for their actions for reasons just described. In future, many DL-driven AI systems will be deemed illegal or non-complaint because of this huge limitation.

DL-Enabled Data Collection Systems Do Not Meet GDPR Requirements

As Data Security will be of cardinal importance in the Data Science world of 2018, many DL- enabled AI systems will fail the security regulations set forth by various regulatory bodies. A good example of this security non-compliance is General Data Protection Regulation (GDPR) which requires that all automated decision-making is commented with highly detailed, logical explanations to prevent discrimination based on race, health conditions, personal opinions, and so on. GDPR will go into effect by May 25, 2018. Thus, all data collection agencies active in the 28 European nations will be forced to abandon their DL-enabled apps or platforms long before that or risk severe penal actions.



Photo Credit: Zapp2Photo/Shutterstock.com

About the author

Paramita Ghosh has over two and a half decades of business writing experience, much of which has been writing for technology and business domains. She has written extensively for a broad range of industries, including but not limited to data management and data technologies. Paramita has also contributed to blended learning projects. She received her M.A. degree in English Literature in 1984 from Jadavpur University in India, and embarked on her career in the United States in 1989 after completing professional coursework. Having ghostwritten and authored hundreds of articles, blog posts, white papers, case studies, marketing content, and learning modules, Paramita has included authorship of one or two books on the business of business writing as part of her post-retirement projects. She thinks her professional strength is “lifelong learning.”

You might also like...

RWDG Slides: The Non-Invasive Data Governance Framework

Read More →
We use technologies such as cookies to understand how you use our site and to provide a better user experience. This includes personalizing content, using analytics and improving site operations. We may share your information about your use of our site with third parties in accordance with our Privacy Policy. You can change your cookie settings as described here at any time, but parts of our site may not function correctly without them. By continuing to use our site, you agree that we can save cookies on your device, unless you have disabled cookies.
I Accept