Invited Talks

Deep learning has had enormous success on perceptual tasks but still struggles in providing a model for inference. To address this gap, we have been developing Memory-Attention-Composition networks (MACnets). The MACnet design provides a strong prior for explicitly iterative reasoning, enabling it to learn explainable, structured reasoning, as well as achieve good generalization from a modest amount of data. The model builds from the great success of existing recurrent cells such as LSTMs: A MacNet is a sequence of a single recurrent Memory, Attention, and Composition (MAC) cell. However, its design imposes structural constraints on the operation of each cell and the interactions between them, incorporating explicit control and soft attention mechanisms. We demonstrate the model’s strength and robustness on the challenging CLEVR dataset for visual reasoning (Johnson et al. 2016), achieving a new state-of-the-art 98.9% accuracy, halving the error rate of the previous best model. More importantly, we show that the new model is more data-efficient, achieving good results from even a modest amount of training data. Joint work with Drew Hudson.

In recent years, significant progress has been made in solving challenging problems across various domains using deep reinforcement learning. However reproducing results for state-of-the-art deep RL methods is seldom straightforward. High variance of some methods can make learning particularly difficult when environments or rewards are strongly stochastic. Furthermore, results can be brittle to even minor perturbations in the domain or experimental procedure. In this talk, I will discuss challenges that arise in experimental techniques and reporting procedures in deep RL, and will suggest methods and guidelines to make future results more reproducible, reusable and robust. I will also report on findings from the ICLR 2018 reproducibility challenge.

http://www.cs.mcgill.ca/~jpineau/

In this interactive chat, we will cover a number of topics, including Daphne's work in machine learning and probabilistic graphical models, her journey in founding Coursera, and her past and ongoing work in applying machine learning to biology and health.

http://ai.stanford.edu/~koller

Deep learning in artificial intelligence (AI) has demonstrated that learning hierarchical representations is a good approach for generating useful sensorimotor behaviors. However, the key to effective hierarchical learning is a mechanism for ""credit assignment"", i.e. a means for determining which neurons and synapses in the hierarchy are ultimately responsible for behaviors/errors. In AI, credit assignment is accomplished with the backpropagation of error algorithm, but backpropagation of error is biologically infeasible, requiring distinct feedforward and feedback phases or pathways. Moreover, backpropagation applied to standard neural network architectures does not actually deliver the flexibility that characterizes animal learning. Here, we present a computational model for hierarchical credit assignment motivated by neocortical microcircuits. Thanks to the unique physiology of apical dendrites in pyramidal neurons, bottom-up and top-down signals can be multiplexed in ensembles of pyramidal neurons using spikes versus high-frequency bursts, respectively. We demonstrate that with the help of apical dendrite targeting inhibition, akin to somatostatin positive interneurons, recursive credit assignment is possible without distinct phases or pathways for feedforward and feedback signals. Moreover, by using ensembles of pyramidal neurons to encode these signals, dynamic routing of information is possible, which could help to generate more flexible representations for continual learning. Altogether, our work provides a model of hierarchical learning that is motivated by the structure of neocortical microcircuits. It also provides specific experimental predictions about which components of the neocortical microcircuit may be involved in credit assignment calculations in the real brain.

In machine learning, we use data to automatically find dependences in the world, with the goal of predicting future observations. Most machine learning methods build on statistics, but one can also try to go beyond this, assaying causal structures underlying statistical dependences. Can such causal knowledge help prediction in machine learning tasks? We argue that this is indeed the case, due to the fact that causal models are more robust to changes that occur in real world datasets. We discuss implications of causal models for machine learning tasks, focusing on an assumption of ‘independent mechanisms’.

URL: https://ei.is.tuebingen.mpg.de/person/bs

Deep generative models of speech have recently surpassed classic algorithms and achieved state-of-the-art performance in text-to-speech (TTS). Specifically, the WaveNet model has significantly increased the quality of generated speech and it has already been transferred into a fast architecture enabling its use in real world with Google Assistant. On the other hand, new developments in Deep RL agents such as the IMPALA architecture have resulted in increased performance across 30 complex 3D environments and have demonstrated positive transfer whilst solving all tasks at the same time. In my talk, I will first explain the recent developments in the WaveNet project and its application as a production TTS system at Google. I will then explain our most recent agent framework: IMPALA and demonstrate its performance on the new DMLab30 challenge domain. Finally, I will introduce a very recent work combining these two general research directions --the SPIRAL algorithm-- an agent that can learn a generative model of images by creating a visual program over brush strokes.

The status quo in visual recognition is to learn from batches of unrelated Web photos labeled by human annotators. Yet cognitive science tells us that perception develops in the context of acting and moving in the world---and without intensive supervision. Meanwhile, many realistic vision tasks require not just categorizing a well-composed human-taken photo, but also intelligently deciding where to look in order to get a meaningful observation in the first place. In the context of these challenges, we are exploring ways to learn visual representations from unlabeled video accompanied by multi-modal sensory data like egomotion and sound. Moving from passively captured video to agents that control their own cameras, we investigate the problem of how to move to intelligently acquire visual observations. In particular, we introduce policy learning approaches for active look-around behavior---both for the sake of a specific recognition task as well as for generic exploratory behavior.

Healthcare is rapidly becoming a data-intensive discipline, driven by increasing digitization of health data, novel measurement technologies, and new policy-based incentives. Critical decisions about ​whom​ and h​ ow​ to treat can be made more precisely by layering an individual’s data over that from a population. In this talk, I will begin by introducing the types of health data currently being collected and the challenges associated with learning models from these data. Next, I will describe new techniques that leverage probabilistic methods and counterfactual reasoning for tackling the aforementioned challenges. Finally, I will introduce areas where ​statistical machine-learning techniques are leading to new classes of computational diagnostic and treatment planning tools—tools that tease out subtle information from “messy” observational datasets, and provide reliable inferences given detailed context about the individual patient.

Digital computers have transformed work in almost every sector of the economy over the past several decades. We are now at the beginning of an even larger and more rapid transformation due to recent advances in machine learning (ML), which is capable of accelerating the pace of automation itself. However, although it is clear that ML is a “general purpose technology,” like the steam engine and electricity, which spawns a plethora of additional innovations and capabilities, there is no widely shared agreement on the tasks where ML systems excel, and thus little agreement on the specific expected impacts on the workforce and on the economy more broadly. We discuss what we see to be key implications for the workforce, drawing on our rubric of what the current generation of ML systems can and cannot do. Although parts of many jobs may be “suitable for ML” (SML), other tasks within these same jobs do not fit the criteria for ML well; hence, effects on employment are more complex than the simple replacement and substitution story emphasized by some. The economic effects of ML are relatively limited today, and we are not facing the imminent “end of work” as is sometimes proclaimed. However, the implications for the economy and the workforce going forward are profound.