Critiquing Deep Learning: limits of deep learning by Gary Marcus

In his paper titled “Deep Learning: A Critical Appraisal,” Gary Marcus offers an overview of the limitations associated with deep learning and cautions against the risk of excessive hype surrounding this technology. Marcus also introduces alternative approaches to deep learning. The paper begins by discussing what deep learning excels at, for example, face and speech recognition. Subsequently, it delves into the limitations of deep learning, and in the conclusion explores possible ways for achieving human-level intelligence in computers, which include unsupervised learning, symbolic AI, and drawing insights from human cognition. In this essay, I will examine Marcus’ ideas on deep learning and provide a summary, complemented by my commentary.

What is deep learning and what does it do well?
Deep learning is a branch of machine learning, a subset of artificial intelligence. It instructs computers to execute tasks by exposing them to examples. Despite its existence for some time, deep learning has gained popularity in the past five years. It can be regarded as a statistical technique for pattern classification, employing sample data and multi-layered neural networks. These networks consist of input units, multiple hidden layers, and output units connected by nodes, with the weights of the connections adjustable through backpropagation to minimize a loss function. The organization of input, hidden, and output nodes closely mimics the connections found in biological neurons, hence it is known as an artificial neural network.

Limitations of deep learning:
Deep learning has made a profound impact on various domains, including speech recognition, image recognition, and language translation, causing companies to invest heavily in the technology. However, as Gary Marcus argues, deep learning may be approaching its limitations, and it might be wise to explore alternative solutions if our objective is to surpass human intelligence with computers. In the following lines, I will delve into Marcus’ critique of deep learning and examine the limitations he identifies.

Based on Marcus’ paper, the limitations of deep learning begin with its inability to make generalizations based on a few trials; instead, it requires thousands or millions of training examples. In contrast, humans can learn abstract relationships with only a few examples due to their ability to grasp abstractions through verbal definitions. Deep learning, on the other hand, heavily relies on vast amounts of data to train the model and make accurate predictions. The reason for this lies in the necessity for deep learning models to have a large sample size to extract meaningful insights. The more data the model is exposed to, the better it can understand underlying patterns and generalize to new, unseen data. However, a risk of overfitting arises when trained with a limited amount of data, where the model becomes too closely tied to the known training data, resulting in poor performance and inaccurate predictions. Despite this, using large amounts of data has its downsides, as it can be a laborious and time-consuming task. To summarize, deep learning heavily relies on using vast amounts of data, and in cases of limited data, alternative solutions should be considered.

The second limitation highlighted by Marcus is the lack of comprehension of abstract concepts, indicating that deep learning is shallower than initially thought. An example can be found in DeepMind’s Atari game, where deep learning combined with reinforcement learning can beat human experts at games. However, when the system encounters scenarios different from those it was trained with, it performs poorly, showcasing the inability of deep learning to understand abstract concepts. Overall, deep learning seems to extract patterns that are only on a surface level.

Deep learning seems to extract patterns that are only on a surface level

The third limitation presented by Marcus involves the inability to deal with hierarchical structures and struggles with open-ended inference. He suggests that deep learning models cannot understand the hierarchy of language and its complex structure, treating sentences as if they are merely lists of words. This poses challenges for deep learning models when attempting to comprehend language. The underlying problem is that deep learning learns correlations in a simple, unstructured list. While hierarchical structures are not directly represented, deep learning systems are forced to use inadequate proxies such as the sequential position of a word in a sequence. As a result, deep learning models have less success in tasks where inference goes beyond what is explicitly stated in a text. Comparing deep learning models with humans reveals a significant difference. People frequently draw new and meaningful conclusions from texts by implying them, such as discerning a character’s motivations through indirect language. Currently, no deep learning system, using knowledge from the real world, can make open-ended conclusions as accurately as humans.

The transparency issue arises because it’s not always clear to most people how the system is making decisions

Another limitation argued by Marcus is the insufficient transparency of deep learning. Deep learning systems have a vast number of parameters, making it challenging even for developers to wrap their minds around it. Despite some progress in visualizing individual nodes, the majority of individuals remain confused by deep learning. The transparency issue arises because it’s not always clear to most people how the system is making decisions. This might pose a future problem, especially as more technological devices incorporate deep learning. In cases where individuals want to understand how the system works and comes up with decisions, such as in medical diagnoses or financial trades, the lack of transparency becomes significant.

The fifth limitation is the lack of integration with prior knowledge. This means that while deep learning is effective for problems with many labeled examples, it has limited usefulness for open-ended problems that require commonsense reasoning.

Another limitation Marcus raises is the difficulty in distinguishing causation from correlation. Deep learning excels at learning complex correlations between input and output features but lacks the ability to understand causality. For instance, a deep learning system can comprehend the correlation between vocabulary learning and height but cannot grasp how this correlation is caused by growth and development.

The sixth limitation is the inability to use deep learning in unstable situations. In other words, for a deep learning model to make accurate predictions, there must be a set of stable rules it follows. Applying deep learning to tasks with predictable rules, such as board games, would likely yield better results compared to using it for unpredictable situations, as seen in 2013 when Google’s flu trends system missed the peak of flu season.

Another limitation concerns the trustworthiness of deep learning. Marcus argues that we cannot fully trust these models as they only work well as approximations. He further states that while deep learning systems can be useful in a specific area, they can also be easily tricked. Numerous studies and real-world examples have demonstrated this, such as deep learning systems in the vision domain mistaking stripes for school buses and 3D-printed turtles for weapons. To sum things up, although deep learning performs effectively in the majority of situations, its reliability cannot be guaranteed in all circumstances.

The last significant challenge mentioned in Gary Marcus’s paper is the difficulty of robust engineering with the deep learning method. This implies that it’s easier to create systems that work in limited circumstances but much harder to guarantee their performance in new data or when used as part of a larger system.

Discussion
As mentioned earlier, Marcus views deep learning as a powerful tool but acknowledges its limitations. When faced with a limited amount of data or when the test set differs from the training set, deep learning may not perform optimally. The author’s experiments in 1997 on neural networks attempting to generalize language patterns revealed that the networks could not extrapolate beyond the training space. This challenge of not being able to generalize beyond the training space continues to persist.

Furthermore, Marcus argues that deep learning is inherently shallow, unable to comprehend abstract concepts, hierarchical structures, and open-ended inference. Additionally, deep learning lacks transparency, proving difficult to understand even for developers. Despite these limitations, Marcus argues that deep learning has a significant impact on various domains. He suggests that more research is required to explore alternative solutions that could enable computers to surpass human intelligence.

Despite the challenges he outlines, Marcus does not advocate for abandoning deep learning. Instead, he proposes the idea of reconceptualization, advocating for using deep learning as a tool. He also discusses the risks associated with the hype towards deep learning, warning of the potential for another “AI winter” similar to the one that occurred in the 70s when the field became too narrow for practical use. Marcus expresses concern that AI could become fixated on limited models like deep learning, neglecting more promising approaches. To reach its full potential, Marcus believes the field must avoid stagnation.

Furthermore, according to Marcus, alternatives to deep learning exist, such as unsupervised learning, symbolic AI, and drawing insights from human cognition.

Conclusion
In conclusion, deep learning stands out as a powerful tool applicable across various fields. Nevertheless, it is crucial to recognize its limitations, as articulated by Gary Marcus in his paper. The paper begins by acknowledging the strengths of deep learning and subsequently outlines its limitations, including its reliance on vast amounts of data, shallow understanding, challenges with hierarchical structures and open-ended inference, and a lack of transparency.

I believe that even despite the remarkable advancements deep learning has demonstrated in specific areas, by exploring and addressing these limitations, we can gain a better appreciation of its potential and work towards enhancements. Furthermore, it is prudent to explore alternative solutions, particularly in areas where the objective is to propel computers beyond human intelligence.