Deep Learning seminars

Mental simulation—the capacity to imagine what will or what could be—is a salient feature of human cognition, playing a key role in a wide range of cognitive abilities. In artificial intelligence, the last few years have seen the development of methods which are analogous to mental models and mental simulation. In this talk, I will discuss recent methods in deep learning for constructing such models from data and learning to use them via reinforcement learning, and compare such approaches to human mental simulation. While a number of challenges remain in matching the capacity of human mental simulation, I will highlight some recent progress on developing more compositional and efficient model-based algorithms through the use of graph neural networks and tree search.

All sensory information – like what we see, hear and smell – gets encoded in spike trains by sensory neurons and gets sent to the brain. Due to the complexity of neural circuits and the difficulty of quantifying complex animal behavior, it has been exceedingly hard to resolve how the brain decodes these spike trains to drive behavior. We now measure quantal signals originating from sparse photons through the most sensitive neural circuits of the mammalian retina and correlate the retinal output spike trains with precisely quantified behavioral decisions. We utilize a combination of electrophysiological measurements on the most sensitive ON and OFF retinal ganglion cell types and a novel deep-learning based tracking technology of the head and body positions of freely-moving mice. We show that visually-guided behavior relies on information from the retinal ON pathway for the dimmest light increments and on information from the retinal OFF pathway for the dimmest light decrements (“quantal shadows”). Our results show that the distribution of labor between ON and OFF pathways starts already at starlight supporting distinct pathway-specific visual computations to drive visually-guided behavior. These results have several fundamental consequences for understanding how the brain integrates information across parallel information streams as well as for understanding the limits of sensory signal processing. In my talk, I will discuss some of the most eminent consequences including the extension of this “Quantum Behavior” paradigm from mouse vision to monkey and human visual systems.

In recent years, artificial neural networks have demonstrated human-like or super-human performance in many tasks including image or speech recognition, natural language processing (NLP), playing Go, chess, poker and video-games. One remarkable feature of the resulting models is that they can develop very intuitive latent representations of their inputs. In these latent spaces, simple linear operations tend to give meaningful results, as in the well-known analogy QUEEN-WOMAN+MAN=KING. We postulate that human brain representations share essential properties with these deep learning latent spaces. To verify this, we test whether artificial latent spaces can serve as a good model for decoding brain activity. We report improvements over state-of-the-art performance for reconstructing seen and imagined face images from fMRI brain activation patterns, using the latent space of a GAN (Generative Adversarial Network) model coupled with a Variational AutoEncoder (VAE). With another GAN model (BigBiGAN), we can decode and reconstruct natural scenes of any category from the corresponding brain activity. Our results suggest that deep learning can produce high-level representations approaching those found in the human brain. Finally, I will discuss whether these deep learning latent spaces could be relevant to the study of consciousness.

Recent advances in deep learning have produced models that far outstrip human performance in a number of domains. However, where machine learning approaches still fall far short of human-level performance is in the capacity to transfer knowledge across domains. While a human learner will happily apply knowledge acquired in one domain (e.g., mathematics) to a different domain (e.g., cooking; a vinaigrette is really just a ratio between edible fat and acid), machine learning models still struggle profoundly at such tasks. I will present a case that human intelligence might be (at least partially) usefully characterised by our ability to transfer knowledge widely, and a framework that we have developed for learning representations that support such transfer. The model is compared to current machine learning approaches.

Learning a new skill requires assimilating into our brain the regularities of the external world and how our body interacts with them as we engage in this skill. Trained Recurrent Neural Networks (TRNNs) are increasingly used as models of neural circuits of animals that were trained in laboratory setups, but the learning process itself has received less attention. Furthermore, most use of TRNNs is of a heuristic, rather than theory-based, nature, leaving many open questions: Which tasks yield to this approach and why? How do initial network architecture and learning rules bias the resultant network? In this talk, I will argue that studying the learning process of TRNNs can both advance our understanding of TRNNs and set up possible comparisons to the biological process of learning. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2021-01-06. Recording duration: 00:42:41.

Activity patterns of neural populations in natural and artificial neural networks constitute representations of data. The nature of these representations and how they are learned are key questions in neuroscience and deep learning. In his talk, I will describe my group's efforts in building a theory of representations as feature maps leading to sample efficient function approximation. Kernel methods are at the heart of these developments. I will present applications to deep learning and neuronal data.

Humans and other primates exhibit rich and versatile behaviour, switching nimbly between tasks as the environmental context requires. I will discuss the neural coding patterns that make this possible in humans and deep networks. First, using deep network simulations, I will characterise two distinct solutions to task acquisition (“lazy” and “rich” learning) which trade off learning speed for robustness, and depend on the initial weights scale and network sparsity. I will chart the predictions of these two schemes for a context-dependent decision-making task, showing that the rich solution is to project task representations onto orthogonal planes on a low-dimensional embedding space. Using behavioural testing and functional neuroimaging in humans, we observe BOLD signals in human prefrontal cortex whose dimensionality and neural geometry are consistent with the rich learning regime. Next, I will discuss the problem of continual learning, showing that behaviourally, humans (unlike vanilla neural networks) learn more effectively when conditions are blocked than interleaved. I will show how this counterintuitive pattern of behaviour can be recreated in neural networks by assuming that information is normalised and temporally clustered (via Hebbian learning) alongside supervised training. Together, this work offers a picture of how humans learn to partition knowledge in the service of structured behaviour, and offers a roadmap for building neural networks that adopt similar principles in the service of multitask learning. This is work with Andrew Saxe, Timo Flesch, David Nagy, and others.

The work to be presented in this talk proposes a novel framework seamlessly providing key properties of both neural nets (learning) and symbolic logic (knowledge and reasoning). Every neuron has a meaning as a component of a formula in a weighted real-valued logic, yielding a highly interpretable disentangled representation. Inference is omnidirectional rather than focused on predefined target variables, and corresponds to logical reasoning, including classical first-order logic theorem proving as a special case. The model is end-to-end differentiable, and learning minimizes a novel loss function capturing logical contradiction, yielding resilience to inconsistent knowledge. It also enables the open-world assumption by maintaining bounds on truth values which can have probabilistic semantics, yielding resilience to incomplete knowledge.

Remarkable advances in experimental neuroscience now enable us to simultaneously observe the activity of many neurons, thereby providing an opportunity to understand how the moment by moment collective dynamics of the brain instantiates learning and cognition. However, efficiently extracting such a conceptual understanding from large, high dimensional neural datasets requires concomitant advances in theoretically driven experimental design, data analysis, and neural circuit modeling. We will discuss how the modern frameworks of high dimensional statistics and deep learning can aid us in this process. In particular we will discuss: how unsupervised tensor component analysis and time warping can extract unbiased and interpretable descriptions of how rapid single trial circuit dynamics change slowly over many trials to mediate learning; how to tradeoff very different experimental resources, like numbers of recorded neurons and trials to accurately discover the structure of collective dynamics and information in the brain, even without spike sorting; deep learning models that accurately capture the retina’s response to natural scenes as well as its internal structure and function; algorithmic approaches for simplifying deep network models of perception; optimality approaches to explain cell-type diversity in the first steps of vision in the retina.

The possibilities of machine learning and neural networks in particular are ever expanding. With increased opportunities to do good, however there are just as many opportunities to do harm and even in the case that good intentions are at the helm, evidence suggests that opportunities for good may eventually prove to be the opposite. The greatest threat to what machine learning is able to achieve and to us as humans, is machine learning that does not reflect the diversity of the users it is meant to serve. It is important that we are not so pre-occupied with advancing technology into the future that we have not taken the time to invest the energy into engineering the security measures this future requires. It is important to investigate now, as thoroughly as we investigate differing deep neural network architectures, the complex questions regarding the fact that humans and the society in which they operate is inherently biased and loaded with prejudice and that these traits find themselves in the machines we create (and increasingly allow to run our lives).

Back-propagation is a powerful supervised learning algorithm in artificial neural networks, because it solves the credit assignment problem (essentially: what should the hidden layers do?). This algorithm has led to the deep learning revolution. But unfortunately, back-propagation cannot be used directly in spiking neural networks (SNN). Indeed, it requires differentiable activation functions, whereas spikes are all-or-none events which cause discontinuities. Here we present two strategies to overcome this problem. The first one is to use a so-called 'surrogate gradient', that is to approximate the derivative of the threshold function with the derivative of a sigmoid. We will present some applications of this method for time series processing (audio, internet traffic, EEG). The second one concerns a specific class of SNNs, which process static inputs using latency coding with at most one spike per neuron. Using approximations, we derived a latency-based back-propagation rule for this sort of networks, called S4NN, and applied it to image classification.

The potential of machine learning and deep learning to advance artificial intelligence is driving a quest to build dedicated computers, such as neuromorphic hardware that emulate the biological processes of the brain. While the hardware technologies already exist, their application to real-world tasks is hindered by the lack of suitable programming methods. Advances at the interface of neural computation and machine learning showed that key aspects of deep learning models and tools can be transferred to biologically plausible neural circuits. Building on these advances, I will show that differentiable programming can address many challenges of programming spiking neural networks for solving real-world tasks, and help devise novel continual and local learning algorithms. In turn, these new algorithms pave the road towards systematically synthesizing machine intelligence in neuromorphic hardware without detailed knowledge of the hardware circuits.

Transformative advances in deep learning, such as deep reinforcement learning, usually rely on gradient-based learning methods such as backpropagation through time (BPTT) as a core learning algorithm. However, BPTT is not argued to be biologically plausible, since it requires to a propagate gradients backwards in time and across neurons. Here, we propose e-prop, a novel gradient-based learning method with local and online weight update rules for recurrent neural networks, and in particular recurrent spiking neural networks (RSNNs). As a result, e-prop has the potential to provide a substantial fraction of the power of deep learning to RSNNs. In this presentation, we will motivate e-prop from the perspective of recent insights in neuroscience and show how these have to be combined to form an algorithm for online gradient descent. The mathematical results will be supported by empirical evidence in supervised and reinforcement learning tasks. We will also discuss how limitations that are inherited from gradient-based learning methods, such as sample-efficiency, can be addressed by considering an evolution-like optimization that enhances learning on particular task families. The emerging learning architecture can be used to learn tasks by a single demonstration, hence enabling one-shot learning.

The emergence of brain-inspired neuromorphic computing as a paradigm for edge AI is motivating the search for high-performance and efficient spiking neural networks to run on this hardware. However, compared to classical neural networks in deep learning, current spiking neural networks lack competitive performance in compelling areas. Here, for sequential and streaming tasks, we demonstrate how spiking recurrent neural networks (SRNN) using adaptive spiking neurons are able to achieve state-of-the-art performance compared to other spiking neural networks and almost reach or exceed the performance of classical recurrent neural networks (RNNs) while exhibiting sparse activity. From this, we calculate a 100x energy improvement for our SRNNs over classical RNNs on the harder tasks. We find in particular that adapting the timescales of spiking neurons is crucial for achieving such performance, and we demonstrate the performance for SRNNs for different spiking neuron models.

The affinity between statistical physics and machine learning has long history, this is reflected even in the machine learning terminology that is in part adopted from physics. I will describe the main lines of this long-lasting friendship in the context of current theoretical challenges and open questions about deep learning. Theoretical physics often proceeds in terms of solvable synthetic models, I will describe the related line of work on solvable models of simple feed-forward neural networks. I will highlight a path forward to capture the subtle interplay between the structure of the data, the architecture of the network, and the learning algorithm.

Model-based approaches to control and decision making have long held the promise of being more powerful and data efficient than model-free counterparts. However, success with model-based methods has been limited to those cases where a perfect model can be queried. The game of Go was mastered by AlphaGo using a combination of neural networks and the MCTS planning algorithm. But planning required a perfect representation of the game rules. I will describe new algorithms that instead leverage deep neural networks to learn models of the environment which are then used to plan, and update policy and value functions. These new algorithms offer hints about how brains might approach planning and acting in complex environments.

We live in the age of the network: Internet social neural ecosystems. This has become one of the main metaphors for how we think about complex systems. This view also dominates the account of brain function. The role of neuronsdescribed by Cajal as the "butterflies of the soul" has become diminished to leaky integrate-and-fire point objects in many models of neural network computation. It is perhaps not surprising that networkexplanations of neural phenomena use neurons as elementary particles andascribe all their wonderful capabilities to their interactions in a network. In the network view the Connectome defines the brain and the butterflies have no role. In this talk I'd like to reclaim some key computations from the networkand return them to their rightful place at the cellular and subcellular level. I'll start with a provocative look at potential computational capacity ofdifferent kinds of brain computation: network vs. subcellular. I'll then consider different levels of pattern and sequence computationwith a glimpse of the efficiency of the subcellular solutions. Finally I propose that there is a suggestive mapping between entire nodesof deep networks to individual neurons. This in my view is how we can walk around with 1.3 litres and 20 watts of installed computational capacity still doing far more than giant AI server farms.

At extreme scale, even conventional sampling and projection methods can become too expensive. This talk treats locality-sensitive hashing as an amortized constant-time adaptive sampler, connecting probabilistic hash tables with efficient unbiased statistical estimators. A few hash lookups can support adaptive estimation at a cost close to uniform sampling. Applications include partition-function estimation for large natural-language models such as word2vec, adaptive gradient estimation for stochastic gradient descent, and sublinear deep learning with very large parameter spaces. The talk also considers memory reduction through a hashing method related to count-min sketches. An example trains a classifier with 100,000 classes and 400,000 features on a single Titan X while using at most five percent of the memory needed to store all weights; a conventional logistic-regression model for that dataset requires 320 GB.

The success of deep neural networks raises questions about the scalability of kernel methods and the roles of large datasets, depth and training algorithms. This talk examines techniques that make kernel learning practical for large datasets in both scalar and multivariate prediction. The methods combine randomized data embeddings, Quasi-Monte Carlo acceleration, distributed convex optimization and input-output kernel learning. Experiments on speech-recognition and computer-vision datasets compare randomized kernel methods with deep neural networks and report essentially matching performance. The talk explores how randomized kernel constructions can resemble neural-network architectures, and how invariant kernel learning and matrix-valued kernels might support deeper architectures. It discusses research results and connections between these approaches.