It also tends towards high capacity storage of all pieces of information which may be relevant for future reasoning [42, 3, 4]. By contrast, longterm memory changes slowly [45, 41], is highly selective and involves repeated consolidation. It contains a set of memories that summarize the entire past, only storing details about observations whicharemostrelevant[28,6]. Deep Learning has seen a variety of architectures for processing sequential data [36, 57, 18].

Recurrent neural networks have a strong inductive bias towards learning temporally compressed representations, as the entire history of a sequence is represented by a single vector. By contrast, Transformers have little inductive bias towards learning temporally compressed representations, as they allow for attention over all previously computed elements in a sequence. Having a more compressed representation of a sequence may be beneficial for generalization, as a high-level representation may be more easily re-used and re-purposed and will contain fewer irrelevant details. At the same time, excessive compression of representations comes at the cost of expressiveness. We propose a solution which divides computation into two streams. A slow stream that is recurrent in nature aims to learn a specialized and compressed representation, by forcing chunks of K time steps into a single representation which is divided into multiple vectors. At the same time, a fast stream is parameterized as a Transformer to process chunks consisting of K time-steps conditioned on the information in the slow-stream. In the proposed approach we hope to gain the expressiveness of the Transformer, while encouraging better compression and structuring of representations in the slow stream. We show the benefits of the proposed method in terms of improved sample efficiency and generalization performance as compared to various competitive baselines for visual perception and sequential decision making tasks.

In this paper I want to propose an argument to support Jerry Fodor's thesis (Fodor 1983) that input systems are modular and thus informationally encapsulated. The argument starts with the suggestion that there is a "grounding problem" in perception, i. e. that there is a problem in explaining how perception that can yield a visual experience is possible, how sensation can become meaningful perception of something for the subject. Given that visual experience is actually possible, this invites a transcendental argument that explains the conditions of its possibility. I propose that one of these conditions is the existence of a visual module in Fodor's sense that allows the step from sensation to object-identifying perception, thus enabling visual experience. It seems to follow that there is informationally encapsulated nonconceptual content in visual perception.

The ability to model the underlying dynamics of visual scenes and reason about the future is central to human intelligence. Many attempts have been made to empower intelligent systems with such physical understanding and prediction abilities. However, most existing methods focus on pixel-to-pixel prediction, which suffers from heavy computational costs while lacking a deep understanding of the physical dynamics behind videos. Recently, object-centric prediction methods have emerged and attracted increasing interest. Inspired by it, this paper proposes an unsupervised object-centric prediction model that makes future predictions by learning visual dynamics between objects. Our model consists of two modules, perceptual, and dynamic module. The perceptual module is utilized to decompose images into several objects and synthesize images with a set of object-centric representations. The dynamic module fuses contextual information, takes environment-object and object-object interaction into account, and predicts the future trajectory of objects. Extensive experiments are conducted to validate the effectiveness of the proposed method. Both quantitative and qualitative experimental results demonstrate that our model generates higher visual quality and more physically reliable predictions compared to the state-of-the-art methods.

Recurrent neural networks have a strong inductive bias towards learning temporally compressed representations, as the entire history of a sequence is represented by a single vector. By contrast, Transformers have little inductive bias towards learning temporally compressed representations, as they allow for attention over all previously computed elements in a sequence. Having a more compressed representation of a sequence may be beneficial for generalization, as a high-level representation may be more easily re-used and re-purposed and will contain fewer irrelevant details. At the same time, excessive compression of representations comes at the cost of expressiveness. We propose a solution which divides computation into two streams. A slow stream that is recurrent in nature aims to learn a specialized and compressed representation, by forcing chunks of $K$ time steps into a single representation which is divided into multiple vectors. At the same time, a fast stream is parameterized as a Transformer to process chunks consisting of $K$ time-steps conditioned on the information in the slow-stream. In the proposed approach we hope to gain the expressiveness of the Transformer, while encouraging better compression and structuring of representations in the slow stream. We show the benefits of the proposed method in terms of improved sample efficiency and generalization performance as compared to various competitive baselines for visual perception and sequential decision making tasks.