A5088254508- Neuroscience and Neural Engineering
- Domain Adaptation and Few-Shot Learning
- Advanced Memory and Neural Computing
- Face and Expression Recognition
- Machine Learning and Algorithms
- Neural dynamics and brain function
- Machine Learning and Data Classification
- Face recognition and analysis
- Neural Networks and Applications
- Advanced Neural Network Applications
- EEG and Brain-Computer Interfaces
- Multimodal Machine Learning Applications
- Video Surveillance and Tracking Methods
- Advanced Vision and Imaging
- Robotics and Sensor-Based Localization
- CCD and CMOS Imaging Sensors
- Advanced Image and Video Retrieval Techniques
- Generative Adversarial Networks and Image Synthesis
- Emotion and Mood Recognition
- Gaussian Processes and Bayesian Inference
- Photoreceptor and optogenetics research
- Anomaly Detection Techniques and Applications
- Image Retrieval and Classification Techniques
- Gaze Tracking and Assistive Technology
- Robotics and Automated Systems
Computer Vision Center
2011-2025
IMEC
2014-2024
Barcelona Supercomputing Center
2007-2024
Universitat Autònoma de Barcelona
2004-2024
KU Leuven
2016-2020
Imec the Netherlands
2016
Centre de Recerca Matemàtica
2007-2012
Universitatea Națională de Știință și Tehnologie Politehnica București
2008
University of the Basque Country
2000-2005
Universidade Estadual de Campinas (UNICAMP)
2002
In vivo recording of neural action-potential and local-field-potential signals requires the use high-resolution penetrating probes. Several international initiatives to better understand brain are driving technology efforts towards maximizing number sites while minimizing probe dimensions. We designed fabricated (0.13- μm SOI Al CMOS) a 384-channel configurable for large-scale in signals. Up 966 selectable active electrodes were integrated along an implantable shank (70 wide, 10 mm long, 20...
Abstract High-density, integrated silicon electrodes have begun to transform systems neuroscience, by enabling large-scale neural population recordings with single cell resolution. Existing technologies, however, provided limited functionality in nonhuman primate species such as macaques, which offer close models of human cognition and behavior. Here, we report the design, fabrication, performance Neuropixels 1.0-NHP, a high channel count linear electrode array designed enable simultaneous...
We present a high electrode density and channel count CMOS (complementary metal-oxide-semiconductor) active neural probe containing 1344 neuron sized recording pixels (20 µm × 20 µm) 12 reference 80 µm), densely packed on 50 thick, 100 wide, 8 mm long shank. The electrodes or consist of dedicated in-situ circuits for signal source amplification, which are directly located under each electrode. supports the simultaneous all 1356 with sufficient to noise ratio typical neuroscience...
Humans are capable of learning new tasks without forgetting previous ones, while neural networks fail due to catastrophic between and previously-learned tasks. We consider a class-incremental setting which means that the task-ID is unknown at inference time. The imbalance old classes typically results in bias network towards newest ones. This problem can either be addressed by storing exemplars from tasks, or using image replay methods. However, latter only applied toy datasets since...
Although CMOS fabrication has enabled a quick evolution in the design of high-density neural probes and neural-recording chips, scaling miniaturization complete data-acquisition systems happened at slower pace. This is mainly due to complexity many requirements that change depending on specific experimental settings. In essence, fundamental challenge system getting signals describing largest possible set neurons out brain down data storage for analysis. requires optimization considers...
Active learning is a paradigm aimed at reducing the annotation effort by training model on actively selected informative and/or representative samples. Another to reduce self-training that learns from large amount of unlabeled data in an unsupervised way and fine-tunes few labeled Recent developments have achieved very impressive results rivaling supervised some datasets. The current work focuses whether two paradigms can benefit each other. We studied object recognition datasets including...
In vivo recording of neural action-potential (AP) and local-field-potential (LFP) signals requires the use high-resolution penetrating probes. Driven by need for large-scale minimal tissue damage, a technology roadmap has been defined next-generation probes aiming to maximize number sites while minimizing probe dimensions [1]. this paper we present 384-channel configurable active high-density which implements in situ buffering under each electrode minimize crosstalk between adjacent metal...
Computing advances and increased smartphone use gives technology system designers greater flexibility in exploiting computer vision to support visually impaired users. Understanding these users' needs will certainly provide insight for the development of improved usability computing devices.
To understand the neural basis of behavior, it is essential to sensitively and accurately measure activity at single neuron spike resolution. Extracellular electrophysiology delivers this, but has biases in neurons detects imperfectly resolves their action potentials. minimize these limitations, we developed a silicon probe with much smaller denser recording sites than previous designs, called Neuropixels Ultra (
We present a high density CMOS neural probe with active electrodes (pixels), consisting of dedicated in-situ circuits for signal source amplification. The complete contains 1356 neuron sized (20×20 μm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ) pixels densely packed on 50 thick, 100 wide and 8 mm long shank. It allows simultaneous high-performance recording from 678 possibility to simultaneously observe all the increased noise....
Active learning aims to reduce the labeling effort that is required train algorithms by an acquisition function selecting most relevant data for which a label should be requested from large unlabeled pool. generally studied on balanced datasets where equal amount of images per class available. However, real-world suffer severe imbalanced classes, so called long-tail distribution. We argue this further complicates active process, since pool can result in suboptimal classifiers. To address...
Autonomous driving systems require huge amounts of data to train. Manual annotation this is time-consuming and prohibitively expensive since it involves human resources. Therefore, active learning emerged as an alternative ease effort make more manageable. In paper, we introduce a novel approach for object detection in videos by exploiting temporal coherence. Our criterion based on the estimated number errors terms false positives negatives. The detections obtained detector are used define...
We review recent progress in neural probes for brain recording, with a focus on the Neuropixels platform. Historically number of neurons' recorded simultaneously, follows Moore's law like behavior, numbers doubling every 6.7 years. Using traditional techniques probe fabrication, continuing to scale up electrode densities is very challenging. describe custom CMOS process technology that enables counts well beyond 1000 electrodes; aim characterize large populations single neuron spatial...
In this paper, we investigate the continual learning of Vision Transformers (ViT) for challenging exemplar-free scenario, with special focus on how to efficiently distill knowledge its crucial self-attention mechanism (SAM). Our work takes an initial step towards a surgical investigation SAM designing coherent methods in ViTs. We first carry out evaluation established regularization techniques. then examine effect when applied two key enablers SAM: (a) contextualized embedding layers, their...
Abstract It is an uninformative truism to state that the brain operates at multiple spatial and temporal scales, each with own set of emergent phenomena. More worthy attention point our current understanding it cannot clearly indicate which these phenomenological scales are significant contributors brain’s function primary output (i.e. behaviour). Apart from sheer complexity problem, a major contributing factor this affairs lack instrumentation can simultaneously address without causing...
Transferring the knowledge of pretrained networks to new domains by means finetuning is a widely used practice for applications based on discriminative models. To best our this has not been studied within context generative deep networks. Therefore, we study domain adaptation applied image generation with adversarial We evaluate several aspects adaptation, including impact target size, relative distance between source and domain, initialization conditional GANs. Our results show that using...