Deep Learning: A Critical Appraisal

By Gary F. Marcus. 2018

1,188 citations recorded by the source corpus.

Although deep learning has historical roots going back decades, neither the term "deep learning" nor the approach was popular just over five years ago, when the field was reignited by papers such as Krizhevsky, Sutskever and Hinton's now classic (2012) deep network model of Imagenet. What has the field discovered in the five subsequent years? Against a background of considerable progress in areas such as speech recognition, image recognition, and game playing, and considerable enthusiasm in the popular press, I present ten concerns for deep learning, and suggest that deep learning must be supplemented by other techniques if we are to reach artificial general intelligence.

Subfields: Artificial Intelligence, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Safety, Robustness & Interpretability

Citations in the corpus

Shortcut learning in deep neural networks (2020, background): “We can easily see the connection to machine learning—transferring knowledge to novel problems corresponds to testing generalisation beyond the narrowly learned setting [89, 90, 91].”
The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence (2020, background): “To the contrary, deep learning techniques thus far have proven to be data hungry, shallow, brittle, and limited in their ability to generalize (Marcus, 2018).”
Local Relation Networks for Image Recognition (2019, background): “1, since filter learning could potentially face a combinatorial explosion of different valid compositional possibilities [24, 34, 23].”
Scalable agent alignment via reward modeling: a research direction (2018, background): “Yet there are still many challenges to be solved in order to make deep RL useful in the real world (Stadelmann et al., 2018; Irpan, 2018; Marcus, 2018); in particular, we need algorithms that can learn to perform complex tasks as intended in the absence of a hand-engineered reward function.”
PAL: Program-aided Language Models (2022, background): “Until as recently as three years ago, reasoning was considered to be one of the most significant challenges that large language models (LLMs) had not yet overcome (Marcus, 2018; 2020; Garcez & Lamb, 2020).”
Reconciling deep learning with symbolic artificial intelligence: representing objects and relations (2019, background): “Notwithstanding its undeniable success, critics have recently drawn attention to a number of shortcomings in contemporary deep learning [9 ,10 ,11,12].”
Explanation in Human-AI Systems: A Literature Meta-Review, Synopsis of Key Ideas and Publications, and Bibliography for Explainable AI (2019, neutral): “…2016; Champlin, Bell, and Schocken, 2017; Clancey, 1986a; Cooper, 2004; Core, et al., 2006; Harford, 2014; Hawkins, 2017; Kim, 2018; Kuang, 2017; Marcus, 2017; Monroe, 2018; Pavlus, 2017; Nott, 2017; Pinker, 2017; Schwiep, 2017; Sheh and Monteath, 2018; Voosen, 2017; Wang, et al., 2019;…”
Learning First-Order Symbolic Representations for Planning from the Structure of the State Space (2019, methodology): “Learners, and in particular deep learners, have achieved considerable success but result in black boxes that do not have the flexibility, transparency, and generality of their model-based counterparts [26, 27, 32, 12, 17].”
Invariant Risk Minimization (2019, background): “Understanding which patterns are useful has been previously studied as a correlation-versus-causation dilemma, since spurious correlations stemming from data biases are unrelated to the causal explanation of interest [31, 27, 35, 52].”
An overview of deep learning in medical imaging focusing on MRI (2018, background): “Some are general: from technical challenges related to the lack of mathematical and theoretical underpinnings of many central deep learning models and techniques, and the resulting difficulty in deciding exactly what it is that makes one model better than another, to societal challenges related to m”
Relational inductive biases, deep learning, and graph networks (2018, background): “…learning about entities and relations (Mitchell, 1980), which we, joining many others (Spelke et al., 1992; Spelke and Kinzler, 2007; Marcus, 2001; Tenenbaum et al., 2011; Lake et al., 2017; Lake and Baroni, 2018; Marcus, 2018b), suggest are an essential ingredient for human-like intelligence.”
Towards a Definition of Disentangled Representations (2018, background): “Furthermore, their performance often lacks the same level of robustness and generalisability that is characteristic of biological intelligence [52, 31, 57].”
Meta-Learning in Neural Networks: A Survey (2020, background): “However there are clear limitations [4].”
Measuring abstract reasoning in neural networks (2018, background): “Generalisation in neural networks has been subject of lots of recent debate, with some emphasising the successes (LeCun et al., 2015) and others the failures (Garnelo et al., 2016; Lake & Baroni, 2017; Marcus, 2018).”
Good-Enough Compositional Data Augmentation (2019, background): “It is not the case, as sometimes claimed (Marcus, 2018), that black-box neural models are fundamentally incapable of this kind of predictable generalization—the success of these models at capturing long-range structure in text (Radford et al.”
Spatial Broadcast Decoder: A Simple Architecture for Learning Disentangled Representations in VAEs (2019, background): “Current machine learning algorithms pale in contrast, suﬀering from overﬁtting, adversarial attacks, and domain specialization [Lake et al., 2016, Marcus, 2018].”
Unsupervised Model Selection for Variational Disentangled Representation Learning (2019, background): “…of state of the art approaches are still missing many basic yet important properties, such as fairness, data efﬁcient learning, strong generalisation beyond the training data distribution, or the ability to transfer knowledge between tasks (Lake et al., 2016; Garnelo et al., 2016; Marcus, 2018).”