Table 7.

Future research challenges and the association of deep learning and game theory

Sr. no.	Future research challenges	Association of deep learning and game theory
1.	Some of the areas of deep learning, like deep games [113], require training heuristics. So far, it has not been investigated how training heuristics for deep games on the recurrent model can be applied.	If the behavioral game theory is used, it may be possible to design training heuristics for deep games [83]. However, experimental results are required to support these claims.
2.	The reward design function plays a crucial role in deep learning, e.g., to improve the Monte-Carlo tree search. Combining more sophisticated deep learning architectures with reinforced learning in Learning Reward-bonus Functions (LRF) [40] remains an important future work	Game theory is widely used for negotiations, bargains, and arbitration [10]. If these concepts are utilized for LRF, a step can be taken to solve future research challenges.
3.	Future work may develop and design algorithms to leverage information-rich, better yet unstructured data in healthcare [28].	Game theory can be utilized through verbal decision analysis [70] for unstructured problem analysis. If the same approach can experiment with unstructured data, the solution may be achieved for unstructured data in healthcare. Thus, verbal decision analysis may provide a path to solve the challenges.
4.	First, in genomics, the challenge of designing and developing deep learning systems that best represent and complement human experience in making medical diagnosis and decisions (for example, genome representation) is crucial [150]. The second challenge is how to avoid biasing entities in training sets and how to interpret predictions. Robustness and interpretation are two important directions for method development. Lastly, there is a requirement for iterative/recursive experimentation, in which deep learning predictions can be checked and validated by functional laboratory tests, experiments, or formal clinical assessment [150]. The future challenges related to genome data are also described in [11] in detail.	A causal functional contribution analysis based on game theory [81] may help solve the challenges related to a medical decision and bias-less training set.
5.	In deep learning, so far, there is no significant attempt to improve support for distributed computation [102] by providing efficient primitives for data parallelism.	In [1], an attempt has been made to relate distributed computation and game theory. The details of this attempt can be crucial in providing efficient primitives for data parallelism
6.	Deep learning has been widely used for medical imaging tasks. However, major medical imaging tasks are far from solved, and the optimal deep learning method and architecture for each task and application region have not yet been established [42].	In [42], an algorithm to select feature subsets for hyperspectral image classification using the principle of coalition game theory is discussed. A similar algorithm can be a starting point for research in medical imaging.
7.	Literature suggests that compression dynamics in the information plane are not a common feature of deep learning networks but are critically influenced by the nonlinearities deployed by the network. Research in this direction may reveal new concepts related to deep networks and the information plane [110].	[75, 138] described how game theory could be associated with the network and related theory. This feature of game theory can be utilized to solve the challenges related to deep networks and the information plane
8	In the future, there is a possibility of using deep learning for robust meta-analytic estimation—in conjunction with increased transparency and collaboration at the level of the meta-analysis—revives the potential of meta-analysis to function as the platinum standard in scientific debate [13].	[7, 64] describe that game theory can be helpful in estimation and meta-analysis. It can be a good starting point towards the utilization of deep learning for robust meta-analytic estimation
9	In the future, the content generation framework requires more than one domain of computational creativity within a game-theoretic context [77].	Evolutionary game theory has the capability of content generation [115].
10	AI-based deep learning applications in COVID-19 research are currently facing several obstacles, such as, scarcity, legislation, unavailability of large-scale training data, large, noisy data and rumors, limited awareness of the intersection between medicine and computer science, security issue, data privacy, and unreliable usability of text data [3].	Aslan et al. [6] have shown how graph transduction games can be helpful to deal with limited data points. Lamba [68] has shown a game-theoretical model to enhance the awareness of the models.