Deep Learning seminars

How do humans perceive daily objects of various features and categorize these seemingly intuitive and effortless mental representations? Prior literature focusing on the role of the inferotemporal region (IT) has revealed object category clustering that is consistent with the semantic predefined structure (superordinate, ordinate, subordinate). It has however been debated whether the neural signals in the IT regions are a reflection of such categorical hierarchy [Wen et al.,2018; Bracci et al., 2017]. Visual attributes of images that correlated with semantic and category dimensions may have confounded these prior results. Our study aimed to address this debate by building and comparing models using the DNN AlexNet, to explain the variance in representational dissimilarity matrix (RDM) of neural signals in the IT region. We found that mid and high level perceptual attributes of the DNN model contribute the most to neural RDMs in the IT region. Semantic categories, as in predefined structure, were moderately correlated with mid to high DNN layers (r = [0.24 - 0.36]). Variance partitioning analysis also showed that the IT neural representations were mostly explained by DNN layers, while semantic categorical RDMs brought little additional information. In light of these results, we propose future works should focus more on the specific role IT plays in facilitating the extraction and coding of visual features that lead to the emergence of categorical conceptualizations.

A recent approach to understanding the mammalian visual system is to show correspondence between the sequential stages of processing in the ventral stream with layers in a deep convolutional neural network (DCNN), providing evidence that visual information is processed hierarchically, with successive stages containing ever higher-level information. However, correspondence is usually defined as shared variance between brain region and model layer. We propose that task-relevant variance is a stricter test: If a DCNN layer corresponds to a brain region, then substituting the model’s activity with brain activity should successfully drive the model’s object recognition decision. Using this approach on three datasets (human fMRI and macaque neuron firing rates) we found that in contrast to the hierarchical view, all ventral stream regions corresponded best to later model layers. That is, all regions contain high-level information about object category. We hypothesised that this is due to recurrent connections propagating high-level visual information from later regions back to early regions, in contrast to the exclusively feed-forward connectivity of DCNNs. Using task-relevant correspondence with a late DCNN layer akin to a tracer, we used Granger causal modelling to show late-DCNN correspondence in IT drives correspondence in V4. Our analysis suggests, effectively, that no ventral stream region can be appropriately characterised as ‘early’ beyond 70ms after stimulus presentation, challenging hierarchical models. More broadly, we ask what it means for a model component and brain region to correspond: beyond quantifying shared variance, we must consider the functional role in the computation. We also demonstrate that using a DCNN to decode high-level conceptual information from ventral stream produces a general mapping from brain to model activation space, which generalises to novel classes held-out from training data. This suggests future possibilities for brain-machine interface with high-level conceptual information, beyond current designs that interface with the sensorimotor periphery.

Deep neural networks (DNNs) with the flexibility to learn good top-layer representations have eclipsed shallow kernel methods without that flexibility. Here, we take inspiration from deep neural networks to develop a new family of deep kernel method. In a deep kernel method, there is a kernel at every layer, and the kernels are jointly optimized to improve performance (with strong regularisation). We establish the representational power of deep kernel methods, by showing that they perform exact inference in an infinitely wide Bayesian neural network or deep Gaussian process. Next, we conjecture that the deep kernel machine objective is unimodal, and give a proof of unimodality for linear kernels. Finally, we exploit the simplicity of the deep kernel machine loss to develop a new family of optimizers, based on a matrix equation from control theory, that converges in around 10 steps.

It is an open question whether preferences for visual art can be lawfully predicted from the basic constituent elements of a visual image. Here, we developed and tested a computational framework to investigate how aesthetic values are formed. We show that it is possible to explain human preferences for a visual art piece based on a mixture of low- and high-level features of the image. Subjective value ratings could be predicted not only within but also across individuals, using a regression model with a common set of interpretable features. We also show that the features predicting aesthetic preference can emerge hierarchically within a deep convolutional neural network trained only for object recognition. Our findings suggest that human preferences for art can be explained at least in part as a systematic integration over the underlying visual features of an image.

Deep learning has made tremendous progress in the last year but it's high computational and memory requirements impose challenges in using deep learning on edge devices. There has been some progress in lowering memory requirements of deep neural networks (for instance, use of half-precision) but there has been minimal effort in developing alternative efficient computational paradigms. Inspired by the brain, Spiking Neural Networks (SNN) provide an energy-efficient alternative to conventional rate-based neural networks. However, SNN architectures that employ the traditional feedforward and feedback pass do not fully exploit the asynchronous event-based processing paradigm of SNNs. In the first part of my talk, I will present my work on predictive coding which offers a fundamentally different approach to developing neural networks that are particularly suitable for event-based processing. In the second part of my talk, I will present our work on development of approaches for SNNs that target specific problems like low response latency and continual learning. References Dora, S., Bohte, S. M., & Pennartz, C. (2021). Deep Gated Hebbian Predictive Coding Accounts for Emergence of Complex Neural Response Properties Along the Visual Cortical Hierarchy. Frontiers in Computational Neuroscience, 65. Saranirad, V., McGinnity, T. M., Dora, S., & Coyle, D. (2021, July). DoB-SNN: A New Neuron Assembly-Inspired Spiking Neural Network for Pattern Classification. In 2021 International Joint Conference on Neural Networks (IJCNN) (pp. 1-6). IEEE. Machingal, P., Thousif, M., Dora, S., Sundaram, S., Meng, Q. (2021). A Cross Entropy Loss for Spiking Neural Networks. Expert Systems with Applications (under review).

Last year’s SNUFA workshop report concluded “Moving toward neuron numbers comparable with biology and applying these networks to real-world data-sets will require the development of novel algorithms, software libraries, and dedicated hardware accelerators that perform well with the specifics of spiking neural networks” [1]. Taking inspiration from machine learning libraries — where techniques such as parallel batch training minimise latency and maximise GPU occupancy — as well as our previous research on efficiently simulating SNNs on GPUs for computational neuroscience [2,3], we are extending our GeNN SNN simulator to pursue this vision. To explore GeNN’s potential, we use the eProp learning rule [4] — which approximates RTRL — to train SNN classifiers on the Spiking Heidelberg Digits and the Spiking Sequential MNIST datasets. We find that the performance of these classifiers is comparable to those trained using BPTT [5] and verify that the theoretical advantages of neuron models with adaptation dynamics [5] translate to improved classification performance. We then measured execution times and found that training an SNN classifier using GeNN and eProp becomes faster than SpyTorch and BPTT after less than 685 timesteps and much larger models can be trained on the same GPU when using GeNN. Furthermore, we demonstrate that our implementation of parallel batch training improves training performance by over 4⨉ and enables near-perfect scaling across multiple GPUs. Finally, we show that performing inference using a recurrent SNN using GeNN uses less energy and has lower latency than a comparable LSTM simulated with TensorFlow [6].

Norse aims to exploit the advantages of bio-inspired neural components, which are sparse and event-driven - a fundamental difference from artificial neural networks. Norse expands PyTorch with primitives for bio-inspired neural components, bringing you two advantages: a modern and proven infrastructure based on PyTorch and deep learning-compatible spiking neural network components.

Gradient-based optimization powered by the backpropagation algorithm proved to be the pivotal method in the training of non-spiking artificial neural networks. At the same time, spiking neural networks hold the promise for efficient processing of real-world sensory data by communicating using discrete events in continuous time. We derive the backpropagation algorithm for a recurrent network of spiking (leaky integrate-and-fire) neurons with hard thresholds and show that the backward dynamics amount to an event-based backpropagation of errors through time. Our derivation uses the jump conditions for partial derivatives at state discontinuities found by applying the implicit function theorem, allowing us to avoid approximations or substitutions. We find that the gradient exists and is finite almost everywhere in weight space, up to the null set where a membrane potential is precisely tangent to the threshold. Our presented algorithm, EventProp, computes the exact gradient with respect to a general loss function based on spike times and membrane potentials. Crucially, the algorithm allows for an event-based communication scheme in the backward phase, retaining the potential advantages of temporal sparsity afforded by spiking neural networks. We demonstrate the optimization of spiking networks using gradients computed via EventProp and the Yin-Yang and MNIST datasets with either a spike time-based or voltage-based loss function and report competitive performance. Our work supports the rigorous study of gradient-based optimization in spiking neural networks as well as the development of event-based neuromorphic architectures for the efficient training of spiking neural networks. While we consider the leaky integrate-and-fire model in this work, our methodology generalises to any neuron model defined as a hybrid dynamical system.

Recent advances in neuromorphic hardware and Surrogate Gradient (SG) learning highlight the potential of Spiking Neural Networks (SNNs) for energy-efficient signal processing and learning. Like in Artificial Neural Networks (ANNs), training performance in SNNs strongly depends on the initialization of synaptic and neuronal parameters. While there are established methods of initializing deep ANNs for high performance, effective strategies for optimal SNN initialization are lacking. Here, we address this gap and propose flexible data-dependent initialization strategies for SNNs.

Depth estimation is an important computer vision task, useful in particular for navigation in autonomous vehicles, or for object manipulation in robotics. Here we solved it using an end-to-end neuromorphic approach, combining two event-based cameras and a Spiking Neural Network (SNN) with a slightly modified U-Net-like encoder-decoder architecture, that we named StereoSpike. More specifically, we used the Multi Vehicle Stereo Event Camera Dataset (MVSEC). It provides a depth ground-truth, which was used to train StereoSpike in a supervised manner, using surrogate gradient descent. We propose a novel readout paradigm to obtain a dense analog prediction –the depth of each pixel– from the spikes of the decoder. We demonstrate that this architecture generalizes very well, even better than its non-spiking counterparts, leading to state-of-the-art test accuracy. To the best of our knowledge, it is the first time that such a large-scale regression problem is solved by a fully spiking network. Finally, we show that low firing rates (<10%) can be obtained via regularization, with a minimal cost in accuracy. This means that StereoSpike could be implemented efficiently on neuromorphic chips, opening the door for low power real time embedded systems.

Tali's work emphasized the tradeoff between compression and information preservation. In this talk I will explore this theme in the context of deep learning. Artificial neural networks have recently revolutionized the field of machine learning. However, we still do not have sufficient theoretical understanding of how such models can be successfully learned. Two specific questions in this context are: how can neural nets be learned despite the non-convexity of the learning problem, and how can they generalize well despite often having more parameters than training data. I will describe our recent work showing that gradient-descent optimization indeed leads to 'simpler' models, where simplicity is captured by lower weight norm and in some cases clustering of weight vectors. We demonstrate this for several teacher and student architectures, including learning linear teachers with ReLU networks, learning boolean functions and learning convolutional pattern detection architectures.

The efficacy of images to create quantitative measures of urban perception has been explored in psychology, social science, urban planning and architecture over the last 50 years. The ability to scale these measurements has become possible only in the last decade, due to increased urban surveillance in the form of street view and satellite imagery, and the accessibility of such data. This talk will present a series of projects which make use of imagery and CNNs to predict, measure and interpret the social and physical environments of our cities.

In this workshop you'll learn the fundamentals of PyTorch using an incremental, from-first-principles approach. We'll start with tensors, autograd, and the dynamic computation graph, and then move on to developing and training a simple model using PyTorch's model classes, datasets, data loaders, optimizers, and more. You should be comfortable using Python, Jupyter notebooks, Google Colab, Numpy and, preferably, object oriented programming.

The success of deep learning sparked interest in whether the brain learns by using similar techniques for assigning credit to each synaptic weight for its contribution to the network output. However, the majority of current attempts at biologically-plausible learning methods are either non-local in time, require highly specific connectivity motives, or have no clear link to any known mathematical optimization method. Here, we introduce Deep Feedback Control (DFC), a new learning method that uses a feedback controller to drive a deep neural network to match a desired output target and whose control signal can be used for credit assignment. The resulting learning rule is fully local in space and time and approximates Gauss-Newton optimization for a wide range of feedback connectivity patterns. To further underline its biological plausibility, we relate DFC to a multi-compartment model of cortical pyramidal neurons with a local voltage-dependent synaptic plasticity rule, consistent with recent theories of dendritic processing. By combining dynamical system theory with mathematical optimization theory, we provide a strong theoretical foundation for DFC that we corroborate with detailed results on toy experiments and standard computer-vision benchmarks.

Robust detection of biological structures such as neuronal dendrites in brightfield micrographs, tumor tissue in histological slides, or pathological brain regions in MRI scans is a fundamental task in bio-image analysis. Detection of those structures requests complex decision making which is often impossible with current image analysis software, and therefore typically executed by humans in a tedious and time-consuming manual procedure. Supervised pixel classification based on Deep Convolutional Neural Networks (DNNs) is currently emerging as the most promising technique to solve such complex region detection tasks. Here, a self-learning artificial neural network is trained with a small set of manually annotated images to eventually identify the trained structures from large image data sets in a fully automated way. While supervised pixel classification based on faster machine learning algorithms like Random Forests are nowadays part of the standard toolbox of bio-image analysts (e.g. Ilastik), the currently emerging tools based on deep learning are still rarely used. There is also not much experience in the community how much training data has to be collected, to obtain a reasonable prediction result with deep learning based approaches. Our software YAPiC (Yet Another Pixel Classifier) provides an easy-to-use Python- and command line interface and is purely designed for intuitive pixel classification of multidimensional images with DNNs. With the aim to integrate well in the current open source ecosystem, YAPiC utilizes the Ilastik user interface in combination with a high performance GPU server for model training and prediction. Numerous research groups at our institute have already successfully applied YAPiC for a variety of tasks. From our experience, a surprisingly low amount of sparse label data is needed to train a sufficiently working classifier for typical bioimaging applications. Not least because of this, YAPiC has become the "standard weapon” for our core facility to detect objects in hard-to-segement images. We would like to present some use cases like cell classification in high content screening, tissue detection in histological slides, quantification of neural outgrowth in phase contrast time series, or actin filament detection in transmission electron microscopy.

There has been a recent surge of interest in the question of whether and how deep learning algorithms might be capable of abstract reasoning, much of which has centered around datasets based on Raven’s Progressive Matrices (RPM), a visual analogy problem set commonly employed to assess fluid intelligence. This has led to the development of algorithms that are capable of solving RPM-like problems directly from pixel-level inputs. However, these algorithms require extensive direct training on analogy problems, and typically generalize poorly to novel problem types. This is in stark contrast to human reasoners, who are capable of solving RPM and other analogy problems zero-shot — that is, with no direct training on those problems. Indeed, it’s this capacity for zero-shot reasoning about novel problem types, i.e. fluid intelligence, that RPM was originally designed to measure. I will present some results from our recent efforts to model this capacity for zero-shot reasoning, based on an extension of a recently proposed approach to analogical mapping we refer to as Probabilistic Analogical Mapping (PAM). Our RPM model uses deep learning to extract attributed graph representations from pixel-level inputs, and then performs alignment of objects between source and target analogs using gradient descent to optimize a graph-matching objective. This extended version of PAM features a number of new capabilities that underscore the flexibility of the overall approach, including 1) the capacity to discover solutions that emphasize either object similarity or relation similarity, based on the demands of a given problem, 2) the ability to extract a schema representing the overall abstract pattern that characterizes a problem, and 3) the ability to directly infer the answer to a problem, rather than relying on a set of possible answer choices. This work suggests that PAM is a promising framework for modeling human zero-shot reasoning.

(1) We (Grace Lindsay) used convolutional neural nets to model attention, by scaling the input/output function of neurons in an imagenet-trained network according to their selectivity for the feature or object category being attended. While this was effective in improving performance on difficult tasks, it was far less effective in earlier than in later layers. This indicated that neurons selective for a feature in earlier layers did not necessarily drive neurons selective for that feature in later layers. In contrast, applying attention according to the gradient for improving task performance worked well in early as well as late layers. This raises the question whether biological attentional modulation might reflect task requirements and not only the features of the stimuli to be attended. We suggest a simple experiment to answer this question, which we hope to convince an appropriate lab to carry out. (2) In E/I networks, a "paradoxical" response to stimulation has been shown: If the excitatory neurons would be unstable by themselves, but are stabilized by feedback inhibition (an "inhibition-stabilized network", or ISN), then, in response to addition of excitatory input to inhibitory neurons, their steady-state firing rates paradoxically decrease. In circuits with multiple inhibitory cell types, this has been generalized: in an ISN, if there is an added stimulus only to inhibitory cells, there will be a paradoxical change in the net inhibition received by excitatory cells -- e.g., if excitatory firing rates increase, so too will the net inhibition they receive. This does not imply that the firing rates of any particular inhibitory cell type will change paradoxically. Here we (Agostina Palmigiano along with Francesco Fumarola, and experimental work of Dan Mossing in the Adesnik lab) generalize the conditions for a paradoxical firing rate response, including in responses to partial as well as full perturbation of the neurons of a given cell type. We work in the context of the circuit with three inhibitory cell types (PV, SOM, VIP) in mouse V1. We show that, if a given cell type shows a paradoxical response to its own full stimulation, then the circuit without that cell type is unstable. This and experimental results to date, as well as our models fitted to data, suggest that PV but not SOM interneurons stabilize the circuit of layer 2/3 of mouse V1, at least for smaller visual stimulus sizes. For partial perturbations of a fraction f of a cell type that responds paradoxically to a full perturbation, there is a "fractional paradoxical effect": the proportion of all the cells of that type, stimulated and unstimulated, that respond opposite to the stimulation (i.e. negative response to excitation), changes non-monotonically, approaching 1 for f→0, decreasing with increasing f, and then increasing again to again approach 1 as f→1. I'll explain the origins of this behavior.3) We (Mario Dipoppa, in collaboration with the experimental work of Andy Keller and Morgane Roth from the Scanziani lab) have studied the E-PV-SOM-VIP circuit underlying contextual modulation in layer 2/3 of mouse V1. Experiments showed that E, PV, and VIP are suppressed by a surround stimuus that has the same orientation as, but not by one orthogonal to, the center stimulus. SOM neurons show the opposite behavior, being suppressed by an orthogonal but much less by a parallel surround. A combination of theory and optogenetic experiments show that the disinhibitory circuit -- VIP inhibits SOM, which inhibits E -- modulate responses between the two conditions. However, it does so, as part of the recurrent circuit, primarily by changing the recurrent excitation E cells receive, rather than by directly changing the inhibition received, in a manner reminiscent of the paradoxical response. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2021-06-09. Recording duration: 00:51:59.

Humans can adapt to a novel task on our first try. By contrast, artificial intelligence systems often require immense amounts of data to adapt. In this talk, I will discuss my recent work (https://www.pnas.org/content/117/52/32970) on creating deep learning systems that can adapt on their first try by exploiting relationships between tasks. Specifically, the approach is based on transforming a representation for a known task to produce a representation for the novel task, by inferring and then using a higher order function that captures a relationship between the tasks. This approach can be interpreted as a type of analogical reasoning. I will show that task transformation can allow systems to adapt to novel tasks on their first try in domains ranging from card games, to mathematical objects, to image classification and reinforcement learning. I will discuss the analogical interpretation of this approach, an analogy between levels of abstraction within the model architecture that I refer to as homoiconicity, and what this work might suggest about using deep-learning models to infer analogies more generally.

DeepLabStream is a python based multi-purpose tool that enables the realtime tracking and manipulation of animals during ongoing experiments. Our toolbox was orginally adapted from the previously published DeepLabCut (Mathis et al., 2018) and expanded on its core capabilities, but is now able to utilize a variety of different network architectures for online pose estimation (SLEAP, DLC-Live, DeepPosekit's StackedDenseNet, StackedHourGlass and LEAP). Our aim is to provide an open-source tool that allows researchers to design custom experiments based on real-time behavior-dependent feedback. My personal ideal goal would be a swiss-army knife like solution where we could integrate the many brilliant python interfaces. We are constantly upgrading DLStream with new features and integrate other open-source solutions.

One may think that Deep Learning technology works in ways that are similar to the human brain. This is not really true. Our best AI technology still does not mimic the brain sufficiently well to be a match in intelligence. I will describe seven differences on how our minds work in ways diametrically opposite to those of Deep Learning technology.