Perhaps the most ambitious scientific quest in human history is the creation of general artificial intelligence, which roughly means AI that is as smart or smarter than humans. The dominant approach in the machine learning community is to attempt to discover each of the pieces required for intelligence, with the implicit assumption that some future group will complete the Herculean task of figuring out how to combine all of those pieces into a complex thinking machine. I call this the ``manual AI approach.'' This paper describes another exciting path that ultimately may be more successful at producing general AI. It is based on the clear trend in machine learning that hand-designed solutions eventually are replaced by more effective, learned solutions. The idea is to create an AI-generating algorithm (AI-GA), which automatically learns how to produce general AI. Three Pillars are essential for the approach: (1) meta-learning architectures, (2) meta-learning the learning algorithms themselves, and (3) generating effective learning environments. I argue that either approach could produce general AI first, and both are scientifically worthwhile irrespective of which is the fastest path. Because both are promising, yet the ML community is currently committed to the manual approach, I argue that our community should increase its research investment in the AI-GA approach. To encourage such research, I describe promising work in each of the Three Pillars. I also discuss AI-GA-specific safety and ethical considerations. Because it it may be the fastest path to general AI and because it is inherently scientifically interesting to understand the conditions in which a simple algorithm can produce general AI (as happened on Earth where Darwinian evolution produced human intelligence), I argue that the pursuit of AI-GAs should be considered a new grand challenge of computer science research.

Complex structures are typical in machine learning. Tailoring learning algorithms for every structure requires an effort that may be saved by defining a generic learning procedure adaptive to any complex structure. In this paper, we propose to map any complex structure onto a generic form, called serialization, over which we can apply any sequence-based density estimator. We then show how to transfer the learned density back onto the space of original structures. To expose the learning procedure to the structural particularities of the original structures, we take care that the serializations reflect accurately the structures' properties. Enumerating all serializations is infeasible. We propose an effective way to sample representative serializations from the complete set of serializations which preserves the statistics of the complete set. Our method is competitive or better than state of the art learning algorithms that have been specifically designed for given structures. In addition, since the serialization involves sampling from a combinatorial process it provides considerable protection from overfitting, which we clearly demonstrate on a number of experiments.

Instance segmentation is an important problem in computer vision, with applications in autonomous driving, drone navigation and robotic manipulation. However, most existing methods are not real-time, complicating their deployment in time-sensitive contexts. In this work, we extend an existing approach to real-time instance segmentation, called `Straight to Shapes' (STS), which makes use of low-dimensional shape embedding spaces to directly regress to object shape masks. The STS model can run at 35 FPS on a high-end desktop, but its accuracy is significantly worse than that of offline state-of-the-art methods. We leverage recent advances in the design and training of deep instance segmentation models to improve the performance accuracy of the STS model whilst keeping its real-time capabilities intact. In particular, we find that parameter sharing, more aggressive data augmentation and the use of structured loss for shape mask prediction all provide a useful boost to the network performance. Our proposed approach, `Straight to Shapes++', achieves a remarkable 19.7 point improvement in mAP (at IOU of 0.5) over the original method as evaluated on the PASCAL VOC dataset, thus redefining the accuracy frontier at real-time speeds. Since the accuracy of instance segmentation is closely tied to that of object bounding box prediction, we also study the error profile of the latter and examine the failure modes of our method for future improvements.

To help understand the underlying mechanisms of neural networks (NNs), several groups have, in recent years, studied the number of linear regions $\ell$ of piecewise linear functions generated by deep neural networks (DNN). In particular, they showed that $\ell$ can grow exponentially with the number of network parameters $p$, a property often used to explain the advantages of DNNs over shallow NNs in approximating complicated functions. Nonetheless, a simple dimension argument shows that DNNs cannot generate all piecewise linear functions with $\ell$ linear regions as soon as $\ell > p$. It is thus natural to seek to characterize specific families of functions with $\ell$ linear regions that can be constructed by DNNs. Iterated Function Systems (IFS) generate sequences of piecewise linear functions $F_k$ with a number of linear regions exponential in $k$. We show that, under mild assumptions, $F_k$ can be generated by a NN using only $\mathcal{O}(k)$ parameters. IFS are used extensively to generate, at low computational cost, natural-looking landscape textures in artificial images. They have also been proposed for compression of natural images, albeit with less commercial success. The surprisingly good performance of this fractal-based compression suggests that our visual system may lock in, to some extent, on self-similarities in images. The combination of this phenomenon with the capacity, demonstrated here, of DNNs to efficiently approximate IFS may contribute to the success of DNNs, particularly striking for image processing tasks, as well as suggest new algorithms for representing self similarities in images based on the DNN mechanism.

It is fundamental and challenging to train robust and accurate Deep Neural Networks (DNNs) when noisy labels exist. Although great progress has been made, there is still one crucial research question which is not thoroughly explored yet: What training examples should be focused and how much more should they be emphasised when training DNNs under label noise? In this work, we study this question and propose gradient rescaling (GR) to solve it. GR modifies the magnitude of logit vector's gradient to emphasise on relatively easier training data points when severe noise exists, which functions as explicit emphasis regularisation to improve the generalisation performance of DNNs. Apart from regularisation, we also interpret GR from the perspectives of sample reweighting and designing robust loss functions. Therefore, our proposed GR helps connect these three approaches in the literature. We empirically demonstrate that GR is highly noise-robust and outperforms the state-of-the-art noise-tolerant algorithms by a large margin, e.g., increasing 7% on CIFAR-100 with 40% noisy labels. It is also significantly superior to standard regularisors. Furthermore, we present comprehensive ablation studies to explore the behaviours of GR under different cases, which is informative for applying GR in real-world scenarios.

Value functions are crucial for model-free Reinforcement Learning (RL) to obtain a policy implicitly or guide the policy updates. Value estimation heavily depends on the stochasticity of environmental dynamics and the quality of reward signals. In this paper, we propose a two-step understanding of value estimation from the perspective of future prediction, through decomposing the value function into a reward-independent future dynamics part and a policy-independent trajectory return part. We then derive a practical deep RL algorithm from the above decomposition, consisting of a convolutional trajectory representation model, a conditional variational dynamics model to predict the expected representation of future trajectory and a convex trajectory return model that maps a trajectory representation to its return. Our algorithm is evaluated in MuJoCo continuous control tasks and shows superior results under both common settings and delayed reward settings.

Spoken language understanding (SLU) acts as a critical component in goal-oriented dialog systems. It typically involves identifying the speakers intent and extracting semantic slots from user utterances, which are known as intent detection (ID) and slot filling (SF). SLU problem has been intensively investigated in recent years. However, these methods just constrain SF results grammatically, solve ID and SF independently, or do not fully utilize the mutual impact of the two tasks. This paper proposes a multi-head self-attention joint model with a conditional random field (CRF) layer and a prior mask. The experiments show the effectiveness of our model, as compared with state-of-the-art models. Meanwhile, online education in China has made great progress in the last few years. But there are few intelligent educational dialog applications for students to learn foreign languages. Hence, we design an intelligent dialog robot equipped with different scenario settings to help students learn communication skills.

Using different sources of information to support automated extracting of relations between biomedical concepts contributes to the development of our understanding of biological systems. The primary comprehensive source of these relations is biomedical literature. Several relation extraction approaches have been proposed to identify relations between concepts in biomedical literature, namely using neural networks algorithms. The use of multichannel architectures composed of multiple data representations, as in deep neural networks, is leading to state-of-the-art results. The right combination of data representations can eventually lead us to even higher evaluation scores in relation extraction tasks. Thus, biomedical ontologies play a fundamental role by providing semantic and ancestry information about an entity. The incorporation of biomedical ontologies has already been proved to enhance previous state-of-the-art results.

A limitation of the Graph Convolutional Networks (GCNs) is that it assumes at a particular $l^{th}$ layer of the neural network model only the $l^{th}$ order neighbourhood nodes of a social network are influential. Furthermore, the GCN has been evaluated on citation and knowledge graphs, but not extensively on friendship-based social graphs. The drawback associated with the dependencies between layers and the order of node neighbourhood for the GCN can be more prevalent for friendship-based graphs. The evaluation of the full potential of the GCN on friendship-based social network requires openly available datasets in larger quantities. However, most available social network datasets are not complete. Also, the majority of the available social network datasets do not contain both the features and ground truth labels. In this work, firstly, we provide a guideline on simulating dynamic social networks, with ground truth labels and features, both coupled with the topology. Secondly, we introduce an open-source Python-based simulation library. We argue that the topology of the network is driven by a set of latent variables, termed as the social DNA (sDNA). We consider the sDNA as labels for the nodes. Finally, by evaluating on our simulated datasets, we propose four new variants of the GCN, mainly to overcome the limitation of dependency between the order of node-neighbourhood and a particular layer of the model. We then evaluate the performance of all the models and our results show that on 27 out of the 30 simulated datasets our proposed GCN variants outperform the original model.

This paper presents a technique called evolving self-supervised neural networks - neural networks that can teach themselves, intrinsically motivated, without external supervision or reward. The proposed method presents some sort-of paradigm shift, and differs greatly from both traditional gradient-based learning and evolutionary algorithms in that it combines the metaphor of evolution and learning, more specifically self-learning, together, rather than treating these phenomena alternatively. I simulate a multi-agent system in which neural networks are used to control autonomous foraging agents with little domain knowledge. Experimental results show that only evolved self-supervised agents can demonstrate some sort of intelligent behaviour, but not evolution or self-learning alone. Indications for future work on evolving intelligence are also presented.