Self-supervised visual learning from interactions with objects

Self-supervised learning (SSL) has revolutionized visual representation learning, but has not achieved the robustness of human vision. A reason for this could be that SSL does not leverage all the data available to humans during learning. When learning about an object, humans often purposefully turn or move around objects and research suggests that these interactions can substantially enhance their learning. Here we explore whether such object-related actions can boost SSL. For this, we extract the actions performed to change from one ego-centric view of an object to another in four video datasets. We then introduce a new loss function to learn visual and action embeddings by aligning the performed action with the representations of two images extracted from the same clip. This permits the performed actions to structure the latent visual representation. Our experiments show that our method consistently outperforms previous methods on downstream category recognition. In our analysis, we find that the observed improvement is associated with a better viewpoint-wise alignment of different objects from the same category. Overall, our work demonstrates that embodied interactions with objects can improve SSL of object categories.

Self-Supervised Learning of Color Constancy

Color constancy (CC) describes the ability of the visual system to perceive an object as having a relatively constant color despite changes in lighting conditions. While CC and its limitations have been carefully characterized in humans, it is still unclear how the visual system acquires this ability during development. Here, we present a first study showing that CC develops in a neural network trained in a self-supervised manner through an invariance learning objective. During learning, objects are presented under changing illuminations, while the network aims to map subsequent views of the same object onto close-by latent representations. This gives rise to representations that are largely invariant to the illumination conditions, offering a plausible example of how CC could emerge during human cognitive development via a form of self-supervised learning.

Caregiver Talk Shapes Toddler Vision: A Computational Study of Dyadic Play

Infants' ability to recognize and categorize objects develops gradually. The second year of life is marked by both the emergence of more semantic visual representations and a better understanding of word meaning. This suggests that language input may play an important role in shaping visual representations. However, even in suitable contexts for word learning like dyadic play sessions, caregivers utterances are sparse and ambiguous, often referring to objects that are different from the one to which the child attends. Here, we systematically investigate to what extent caregivers' utterances can nevertheless enhance visual representations. For this we propose a computational model of visual representation learning during dyadic play. We introduce a synthetic dataset of ego-centric images perceived by a toddler-agent that moves and rotates toy objects in different parts of its home environment while hearing caregivers' utterances, modeled as captions. We propose to model toddlers' learning as simultaneously aligning representations for 1) close-in-time images and 2) co-occurring images and utterances. We show that utterances with statistics matching those of real caregivers give rise to representations supporting improved category recognition. Our analysis reveals that a small decrease/increase in object-relevant naming frequencies can drastically impact the learned representations. This affects the attention on object names within an utterance, which is required for efficient visuo-linguistic alignment. Overall, our results support the hypothesis that caregivers' naming utterances can improve toddlers' visual representations.

MIMo: A Multi-Modal Infant Model for Studying Cognitive Development

Human intelligence and human consciousness emerge gradually during the process of cognitive development. Understanding this development is an essential aspect of understanding the human mind and may facilitate the construction of artificial minds with similar properties. Importantly, human cognitive development relies on embodied interactions with the physical and social environment, which is perceived via complementary sensory modalities. These interactions allow the developing mind to probe the causal structure of the world. This is in stark contrast to common machine learning approaches, e.g., for large language models, which are merely passively ``digesting'' large amounts of training data, but are not in control of their sensory inputs. However, computational modeling of the kind of self-determined embodied interactions that lead to human intelligence and consciousness is a formidable challenge. Here we present MIMo, an open-source multi-modal infant model for studying early cognitive development through computer simulations. MIMo's body is modeled after an 18-month-old child with detailed five-fingered hands. MIMo perceives its surroundings via binocular vision, a vestibular system, proprioception, and touch perception through a full-body virtual skin, while two different actuation models allow control of his body. We describe the design and interfaces of MIMo and provide examples illustrating its use. All code is available at https://github.com/trieschlab/MIMo .

An information-theoretic perspective on intrinsic motivation in reinforcement learning: a survey

The reinforcement learning (RL) research area is very active, with an important number of new contributions; especially considering the emergent field of deep RL (DRL). However a number of scientific and technical challenges still need to be resolved, amongst which we can mention the ability to abstract actions or the difficulty to explore the environment in sparse-reward settings which can be addressed by intrinsic motivation (IM). We propose to survey these research works through a new taxonomy based on information theory: we computationally revisit the notions of surprise, novelty and skill learning. This allows us to identify advantages and disadvantages of methods and exhibit current outlooks of research. Our analysis suggests that novelty and surprise can assist the building of a hierarchy of transferable skills that further abstracts the environment and makes the exploration process more robust.

Time to augment contrastive learning

Biological vision systems are unparalleled in their ability to learn visual representations without supervision. In machine learning, contrastive learning (CL) has led to major advances in forming object representations in an unsupervised fashion. These systems learn representations invariant to augmentation operations over images, like cropping or flipping. In contrast, biological vision systems exploit the temporal structure of the visual experience. This gives access to augmentations not commonly used in CL, like watching the same object from multiple viewpoints or against different backgrounds. Here, we systematically investigate and compare the potential benefits of such time-based augmentations for learning object categories. Our results show that time-based augmentations achieve large performance gains over state-of-the-art image augmentations. Specifically, our analyses reveal that: 1) 3-D object rotations drastically improve the learning of object categories; 2) viewing objects against changing backgrounds is vital for learning to discard background-related information. Overall, we conclude that time-based augmentations can greatly improve contrastive learning, narrowing the gap between artificial and biological vision systems.

Embodied vision for learning object representations

Recent time-contrastive learning approaches manage to learn invariant object representations without supervision. This is achieved by mapping successive views of an object onto close-by internal representations. When considering this learning approach as a model of the development of human object recognition, it is important to consider what visual input a toddler would typically observe while interacting with objects. First, human vision is highly foveated, with high resolution only available in the central region of the field of view. Second, objects may be seen against a blurry background due to infants' limited depth of field. Third, during object manipulation a toddler mostly observes close objects filling a large part of the field of view due to their rather short arms. Here, we study how these effects impact the quality of visual representations learnt through time-contrastive learning. To this end, we let a visually embodied agent "play" with objects in different locations of a near photo-realistic flat. During each play session the agent views an object in multiple orientations before turning its body to view another object. The resulting sequence of views feeds a time-contrastive learning algorithm. Our results show that visual statistics mimicking those of a toddler improve object recognition accuracy in both familiar and novel environments. We argue that this effect is caused by the reduction of features extracted in the background, a neural network bias for large features in the image and a greater similarity between novel and familiar background regions. We conclude that the embodied nature of visual learning may be crucial for understanding the development of human object perception.

DisTop: Discovering a Topological representation to learn diverse and rewarding skills

The optimal way for a deep reinforcement learning (DRL) agent to explore is to learn a set of skills that achieves a uniform distribution of states. Following this,we introduce DisTop, a new model that simultaneously learns diverse skills and focuses on improving rewarding skills. DisTop progressively builds a discrete topology of the environment using an unsupervised contrastive loss, a growing network and a goal-conditioned policy. Using this topology, a state-independent hierarchical policy can select where the agent has to keep discovering skills in the state space. In turn, the newly visited states allows an improved learnt representation and the learning loop continues. Our experiments emphasize that DisTop is agnostic to the ground state representation and that the agent can discover the topology of its environment whether the states are high-dimensional binary data, images, or proprioceptive inputs. We demonstrate that this paradigm is competitiveon MuJoCo benchmarks with state-of-the-art algorithms on both single-task dense rewards and diverse skill discovery. By combining these two aspects, we showthat DisTop achieves state-of-the-art performance in comparison with hierarchical reinforcement learning (HRL) when rewards are sparse. We believe DisTop opens new perspectives by showing that bottom-up skill discovery combined with representation learning can unlock the exploration challenge in DRL.

ELSIM: End-to-end learning of reusable skills through intrinsic motivation

Taking inspiration from developmental learning, we present a novel reinforcement learning architecture which hierarchically learns and represents self-generated skills in an end-to-end way. With this architecture, an agent focuses only on task-rewarded skills while keeping the learning process of skills bottom-up. This bottom-up approach allows to learn skills that 1- are transferable across tasks, 2- improves exploration when rewards are sparse. To do so, we combine a previously defined mutual information objective with a novel curriculum learning algorithm, creating an unlimited and explorable tree of skills. We test our agent on simple gridworld environments to understand and visualize how the agent distinguishes between its skills. Then we show that our approach can scale on more difficult MuJoCo environments in which our agent is able to build a representation of skills which improve over a baseline both transfer learning and exploration when rewards are sparse.

A survey on intrinsic motivation in reinforcement learning

Despite numerous research work in reinforcement learning (RL) and the recent successes obtained by combining it with deep learning, deep reinforcement learning (DRL) is still facing many challenges. Some of them, like the ability to abstract actions or the difficulty to explore the environment with sparse rewards, can be addressed by the use of intrinsic motivation. In this article, we provide a survey on the role of intrinsic motivation in DRL. We categorize the different kinds of intrinsic motivations and detail their interests and limitations. Our investigation shows that the combination of DRL and intrinsic motivation enables to learn more complicated and more generalisable behaviours than standard DRL. We provide an in-depth analysis describing learning modules through an unifying scheme composed of information theory, compression theory and reinforcement learning. We then explain how these modules could serve as building blocks over a complete developmental architecture, highlighting the numerous outlooks of the domain.